pulse-monitor-v2/main.go

package main

import (
	"fmt"
	"image"
	"image/color"
	"os"
	"os/signal"
	"strconv"
	"strings"
	"syscall"
	"time"

	"gocv.io/x/gocv"
	"gopkg.in/yaml.v3"
)

// Config holds application configuration
type Config struct {
	Camera struct {
		RTSPURL string `yaml:"rtsp_url"`
	} `yaml:"camera"`
	HomeAssistant struct {
		URL   string `yaml:"url"`
		Token string `yaml:"token"`
	} `yaml:"home_assistant"`
	OCR struct {
		APIKey string `yaml:"api_key"`
	} `yaml:"ocr"`
	Processing struct {
		IntervalSeconds int `yaml:"interval_seconds"`
	} `yaml:"processing"`
}

func init() {
	// Suppress GStreamer and OpenCV warnings
	os.Setenv("OPENCV_LOG_LEVEL", "ERROR")
	os.Setenv("GST_DEBUG", "0")
}

func loadConfig(path string) (*Config, error) {
	data, err := os.ReadFile(path)
	if err != nil {
		return nil, err
	}

	var cfg Config
	if err := yaml.Unmarshal(data, &cfg); err != nil {
		return nil, err
	}

	// Defaults
	if cfg.Processing.IntervalSeconds == 0 {
		cfg.Processing.IntervalSeconds = 2
	}

	return &cfg, nil
}

func main() {
	// Load config
	cfg, err := loadConfig("config.yaml")
	if err != nil {
		fmt.Printf("Failed to load config: %v\n", err)
		os.Exit(1)
	}

	// Set up components
	capture, err := NewCapture(cfg.Camera.RTSPURL)
	if err != nil {
		fmt.Printf("Failed to connect to camera: %v\n", err)
		os.Exit(1)
	}
	defer capture.Close()

	// Load pattern store
	patterns, err := NewPatternStore("patterns.csv")
	if err != nil {
		fmt.Printf("Failed to load patterns: %v\n", err)
		os.Exit(1)
	}
	fmt.Printf("Loaded %d patterns\n", patterns.Count())

	// Initialize Home Assistant client
	var hass *HASS
	if cfg.HomeAssistant.URL != "" && cfg.HomeAssistant.Token != "" {
		hass = NewHASS(cfg.HomeAssistant.URL, cfg.HomeAssistant.Token)
		fmt.Println("Home Assistant integration enabled")
	}

	// Create training directory
	os.MkdirAll("training", 0755)

	// Start web server for training
	StartWebServer(8090)

	// Handle graceful shutdown
	sigChan := make(chan os.Signal, 1)
	signal.Notify(sigChan, syscall.SIGINT, syscall.SIGTERM)

	fmt.Println("Starting pulse monitor...")

	// Find bar region (retry until valid)
	var barRegion Region
	for {
		frame := capture.NextFrame()
		if frame.Empty() {
			frame.Close()
			time.Sleep(time.Second)
			continue
		}
		prep := preprocessFrame(frame)
		frame.Close()
		gocv.IMWrite("bw_image.png", prep)
		barRegion = findBar(prep)
		prep.Close()
		fmt.Println("Saved bw_image.png and bars_visualization.png")
		fmt.Printf("Bar region: (%d,%d) %dx%d\n", barRegion.X, barRegion.Y, barRegion.W, barRegion.H)
		if barRegion.W > 0 && barRegion.H > 0 {
			break
		}
		fmt.Println("Invalid bar region, retrying in 2s...")
		time.Sleep(2 * time.Second)
	}

	// Track last known values (for display and HASS)
	var lastLeft, lastRight string
	var lastPostedLeft, lastPostedRight int
	var prevDisplayLeft, prevDisplayRight string // for CR/LF logic

	// Rate limit: 2 frames per second
	lastProcess := time.Now()
	processInterval := 500 * time.Millisecond

	// Main loop - detect changes in left/right
	for {
		select {
		case <-sigChan:
			fmt.Printf("\nShutting down... Patterns: %d\n", patterns.Count())
			GenerateHTML("training")
			return
		default:
		}

		frame := capture.NextFrame()
		if frame.Empty() {
			frame.Close()
			time.Sleep(100 * time.Millisecond)
			continue
		}

		// Skip processing if not enough time has passed
		if time.Since(lastProcess) < processInterval {
			frame.Close()
			continue
		}
		lastProcess = time.Now()

		prep := preprocessFrame(frame)
		frame.Close()

		// Crop bar and find digit regions
		barCrop := CropRegion(prep, barRegion)
		rows := barCrop.Rows()

		// Find digit boundaries
		spo2Start, spo2End, hrStart, hrEnd := visualizeColumns(barCrop)

		// Crop tightly to each digit region
		leftRegion := barCrop.Region(image.Rect(spo2Start, 0, spo2End, rows))
		leftCrop := gocv.NewMat()
		leftRegion.CopyTo(&leftCrop)
		leftRegion.Close()

		rightRegion := barCrop.Region(image.Rect(hrStart, 0, hrEnd, rows))
		rightCrop := gocv.NewMat()
		rightRegion.CopyTo(&rightCrop)
		rightRegion.Close()

		leftDigitCrop := leftCrop
		rightDigitCrop := rightCrop

		// Add 1px black border so contours don't touch edges (FindContours ignores border-touching shapes)
		leftPadded := gocv.NewMatWithSize(leftCrop.Rows()+2, leftCrop.Cols()+2, leftCrop.Type())
		rightPadded := gocv.NewMatWithSize(rightCrop.Rows()+2, rightCrop.Cols()+2, rightCrop.Type())
		leftRegionPad := leftPadded.Region(image.Rect(1, 1, leftCrop.Cols()+1, leftCrop.Rows()+1))
		rightRegionPad := rightPadded.Region(image.Rect(1, 1, rightCrop.Cols()+1, rightCrop.Rows()+1))
		leftCrop.CopyTo(&leftRegionPad)
		rightCrop.CopyTo(&rightRegionPad)
		leftRegionPad.Close()
		rightRegionPad.Close()

		// Vectorize contours for shape comparison (RetrievalList = all contours including holes)
		leftContours := gocv.FindContours(leftPadded, gocv.RetrievalList, gocv.ChainApproxSimple)
		rightContours := gocv.FindContours(rightPadded, gocv.RetrievalList, gocv.ChainApproxSimple)
		leftPadded.Close()
		rightPadded.Close()

		// Recognize numbers using hole/stroke analysis
		leftResult := recognizeNumber(leftDigitCrop)
		rightResult := recognizeNumber(rightDigitCrop)

		// Update values if recognized
		if leftResult != "" {
			lastLeft = leftResult
			if hass != nil {
				if val, err := strconv.Atoi(leftResult); err == nil && val != lastPostedLeft {
					hass.PostSpO2(val)
					lastPostedLeft = val
				}
			}
		}

		if rightResult != "" {
			lastRight = rightResult
			if hass != nil {
				if val, err := strconv.Atoi(rightResult); err == nil && val != lastPostedRight {
					hass.PostHR(val)
					lastPostedRight = val
				}
			}
		}
		if lastLeft != "" || lastRight != "" {
			timeStr := time.Now().Format("15:04:05")
			valuesChanged := lastLeft != prevDisplayLeft || lastRight != prevDisplayRight
			if valuesChanged {
				fmt.Printf("\n%s  SpO2: %-3s  HR: %-3s", timeStr, lastLeft, lastRight)
				prevDisplayLeft, prevDisplayRight = lastLeft, lastRight
			} else {
				fmt.Printf("\r%s  SpO2: %-3s  HR: %-3s", timeStr, lastLeft, lastRight)
			}
		}

		// Cleanup
		leftContours.Close()
		rightContours.Close()
		leftCrop.Close()
		rightCrop.Close()
		barCrop.Close()
		prep.Close()
	}
}

// preprocessFrame crops timestamp and rotates the frame
func preprocessFrame(frame gocv.Mat) gocv.Mat {
	// Crop timestamp (top 68 pixels)
	cropped := frame.Region(image.Rect(0, 68, frame.Cols(), frame.Rows()))
	rotated := gocv.NewMat()
	gocv.Rotate(cropped, &rotated, gocv.Rotate90Clockwise)
	cropped.Close()

	// Convert to grayscale then threshold to B&W
	gray := gocv.NewMat()
	gocv.CvtColor(rotated, &gray, gocv.ColorBGRToGray)
	rotated.Close()

	bw := gocv.NewMat()
	gocv.Threshold(gray, &bw, 210, 255, gocv.ThresholdBinary)
	gray.Close()
	return bw
}

// parseNumber extracts a number from string
func parseNumber(s string) int {
	var n int
	for _, c := range s {
		if c >= '0' && c <= '9' {
			n = n*10 + int(c-'0')
		}
	}
	return n
}

// abs returns the absolute value of an int
func abs(x int) int {
	if x < 0 {
		return -x
	}
	return x
}

// findBar detects the digit bar region using B&W image analysis
func findBar(frame gocv.Mat) Region {
	// Input is B&W (0 or 255) from preprocessFrame
	rows := frame.Rows()
	cols := frame.Cols()

	// Count white pixels per row
	whiteCount := make([]int, rows)
	for y := 0; y < rows; y++ {
		for x := 0; x < cols; x++ {
			if frame.GetUCharAt(y, x) > 0 {
				whiteCount[y]++
			}
		}
	}

	// Find ALL bars - transitions between rows with/without white pixels
	type Bar struct {
		Start, End  int
		HasWhite    bool
		WhitePixels int
	}

	var bars []Bar
	var transitions []int
	hasWhite := whiteCount[0] > 0
	barStart := 0
	var pixelSum int

	transitions = append(transitions, 0)

	for y := 0; y < rows; y++ {
		rowHasWhite := whiteCount[y] > 0
		pixelSum += whiteCount[y]

		if rowHasWhite != hasWhite || y == rows-1 {
			height := y - barStart
			if height >= 5 {
				bars = append(bars, Bar{
					Start:       barStart,
					End:         y - 1,
					HasWhite:    hasWhite,
					WhitePixels: pixelSum,
				})
				transitions = append(transitions, y)
			}
			barStart = y
			pixelSum = whiteCount[y]
			hasWhite = rowHasWhite
		}
	}

	// Save visualization
	vis := gocv.NewMat()
	gocv.CvtColor(frame, &vis, gocv.ColorGrayToBGR)

	for _, y := range transitions {
		gocv.Line(&vis, image.Point{0, y}, image.Point{cols, y}, color.RGBA{255, 0, 0, 255}, 2)
		gocv.PutText(&vis, fmt.Sprintf("%d", y), image.Point{10, y - 5}, gocv.FontHersheyPlain, 1.5, color.RGBA{0, 255, 0, 255}, 2)
	}
	gocv.IMWrite("bars_visualization.png", vis)
	vis.Close()

	// Find first and last white bar to define content bounds
	firstWhite, lastWhite := -1, -1
	for i, b := range bars {
		if b.HasWhite {
			if firstWhite == -1 {
				firstWhite = i
			}
			lastWhite = i
		}
	}

	// Calculate vertical span of white content
	var contentStart, contentEnd int
	if firstWhite >= 0 && lastWhite >= 0 {
		contentStart = bars[firstWhite].Start
		contentEnd = bars[lastWhite].End
	}
	contentSpan := contentEnd - contentStart

	// Print bars with position as percentage of white content span
	fmt.Println("Detected bars (% = position within white content):")
	for _, b := range bars {
		t := "BLACK"
		if b.HasWhite {
			t = "WHITE"
		}
		height := b.End - b.Start + 1

		// Calculate position as percentage of content span
		startPct := float64(b.Start-contentStart) / float64(contentSpan) * 100
		endPct := float64(b.End-contentStart) / float64(contentSpan) * 100

		fmt.Printf("  Y=%d-%d (%d px) %s  %.0f%%-%.0f%%\n", b.Start, b.End, height, t, startPct, endPct)
	}

	// Find the digit bar by position within content span
	// It sits around 40-60% and has a significant black gap above it (>5%)
	var digitBar Bar
	found := false
	for i, bar := range bars {
		if !bar.HasWhite {
			continue
		}

		// Calculate position as percentage of content span
		startPct := float64(bar.Start-contentStart) / float64(contentSpan) * 100
		endPct := float64(bar.End-contentStart) / float64(contentSpan) * 100

		// Check if bar is in the 30%-70% range (middle of content)
		if startPct < 30 || startPct > 70 {
			continue
		}

		// Check black gap above (previous bar should be black with significant height)
		if i > 0 {
			prevBar := bars[i-1]
			if !prevBar.HasWhite {
				gapPct := float64(prevBar.End-prevBar.Start+1) / float64(contentSpan) * 100
				// Need at least 5% black gap above
				if gapPct >= 5 {
					digitBar = bar
					found = true
					fmt.Printf("Selected digit bar: %.0f%%-%.0f%% (gap above: %.1f%%)\n", startPct, endPct, gapPct)
					break
				}
			}
		}
	}

	if !found {
		return Region{}
	}

	// Find leftmost and rightmost white pixels in this bar
	minX := cols
	maxX := 0
	for y := digitBar.Start; y <= digitBar.End; y++ {
		for x := 0; x < cols; x++ {
			if frame.GetUCharAt(y, x) > 0 {
				if x < minX {
					minX = x
				}
				if x > maxX {
					maxX = x
				}
			}
		}
	}

	// Save cropped bar
	region := Region{
		X: minX,
		Y: digitBar.Start,
		W: maxX - minX + 1,
		H: digitBar.End - digitBar.Start + 1,
	}
	barCrop := CropRegion(frame, region)
	gocv.IMWrite("bar_cropped.png", barCrop)
	barCrop.Close()
	fmt.Printf("Bar cropped: (%d,%d) %dx%d\n", region.X, region.Y, region.W, region.H)

	return region
}

// parseValues extracts SpO2 and HR from API response like "SPO2:91 HR:117"
func parseValues(s string) (spo2, hr int) {
	s = strings.ToUpper(s)
	var nums []int
	current := ""
	for _, c := range s {
		if c >= '0' && c <= '9' {
			current += string(c)
		} else if current != "" {
			if n, err := strconv.Atoi(current); err == nil {
				nums = append(nums, n)
			}
			current = ""
		}
	}
	if current != "" {
		if n, err := strconv.Atoi(current); err == nil {
			nums = append(nums, n)
		}
	}

	// Assign by range: SpO2 is 70-100, HR is 40-220
	for _, n := range nums {
		if n >= 70 && n <= 100 && spo2 == 0 {
			spo2 = n
		} else if n >= 40 && n <= 220 && hr == 0 {
			hr = n
		}
	}
	return
}

// Region defines a rectangular area
type Region struct {
	X, Y, W, H int
}

// ColumnType classifies a column's content
type ColumnType int

const (
	ColEmpty    ColumnType = iota // no white pixels
	ColFullHeight                  // white pixels span most of height (digits)
	ColBottomOnly                  // white pixels only in bottom portion (labels)
)

// classifyColumn determines if a column is empty, full-height (digit), or bottom-only (label)
func classifyColumn(img gocv.Mat, x int) ColumnType {
	rows := img.Rows()
	topThird := rows / 3

	topPixels := 0
	bottomPixels := 0

	for y := 0; y < rows; y++ {
		if img.GetUCharAt(y, x) > 0 {
			if y < topThird {
				topPixels++
			} else {
				bottomPixels++
			}
		}
	}

	totalPixels := topPixels + bottomPixels
	if totalPixels < 2 {
		return ColEmpty
	}

	// If has ANY content in top third, it's a full-height digit
	if topPixels >= 1 {
		return ColFullHeight
	}

	// Content only in bottom = small label text
	return ColBottomOnly
}

// visualizeColumns draws column classification and finds digit regions
// Returns: spo2Start, spo2End, hrStart, hrEnd (x coordinates)
func visualizeColumns(img gocv.Mat) (int, int, int, int) {
	rows, cols := img.Rows(), img.Cols()

	// Classify all columns
	colTypes := make([]ColumnType, cols)
	for x := 0; x < cols; x++ {
		colTypes[x] = classifyColumn(img, x)
	}

	// Find digit regions: contiguous runs of full-height columns
	// with gaps of at least minGap EMPTY columns between them
	// Labels (BOTTOM columns) don't count as gaps - they're adjacent to digits
	// But we crop to the FULL columns only (excluding labels)
	minGap := 30
	type DigitRun struct{ Start, End int }
	var runs []DigitRun

	inRun := false
	runStart := 0
	lastFull := 0 // track last full-height column for tight cropping
	emptyCount := 0

	for x := 0; x < cols; x++ {
		isDigit := colTypes[x] == ColFullHeight
		isEmpty := colTypes[x] == ColEmpty

		if isDigit {
			if !inRun {
				inRun = true
				runStart = x
			}
			lastFull = x
			emptyCount = 0
		} else if inRun {
			if isEmpty {
				emptyCount++
			}
			// Only end run if we've seen enough truly empty columns
			if emptyCount >= minGap {
				// End current run at last FULL column (tight crop, excluding labels)
				runs = append(runs, DigitRun{runStart, lastFull + 1})
				inRun = false
			}
		}
	}
	if inRun {
		runs = append(runs, DigitRun{runStart, lastFull + 1})
	}

	// Create visualization
	vis := gocv.NewMat()
	gocv.CvtColor(img, &vis, gocv.ColorGrayToBGR)
	defer vis.Close()

	// Draw column type markers
	for x := 0; x < cols; x++ {
		var c color.RGBA
		switch colTypes[x] {
		case ColEmpty:
			c = color.RGBA{0, 0, 255, 255} // red
		case ColFullHeight:
			c = color.RGBA{0, 255, 0, 255} // green
		case ColBottomOnly:
			c = color.RGBA{255, 0, 0, 255} // blue
		}
		gocv.Line(&vis, image.Point{x, 0}, image.Point{x, 3}, c, 1)
	}

	// Draw vertical lines at run boundaries (white)
	for _, run := range runs {
		gocv.Line(&vis, image.Point{run.Start, 0}, image.Point{run.Start, rows}, color.RGBA{255, 255, 255, 255}, 1)
		gocv.Line(&vis, image.Point{run.End, 0}, image.Point{run.End, rows}, color.RGBA{255, 255, 255, 255}, 1)
	}

	// Draw mid-line (cyan) - shows upper/lower half boundary for digit pattern analysis
	gocv.Line(&vis, image.Point{0, rows/2}, image.Point{cols, rows/2}, color.RGBA{0, 255, 255, 255}, 1)

	// Save debug image once, but always print first time
	gocv.IMWrite("debug_columns.png", vis)

	// Return early if we've printed before (use static-like behavior via file mod time)
	fmt.Printf("Digit runs (minGap=%d): ", minGap)
	for i, run := range runs {
		fmt.Printf("Run%d[%d-%d] ", i, run.Start, run.End)
	}
	fmt.Println()

	// Return SpO2 and HR boundaries
	if len(runs) >= 2 {
		return runs[0].Start, runs[0].End, runs[1].Start, runs[1].End
	} else if len(runs) == 1 {
		mid := (runs[0].Start + runs[0].End) / 2
		return runs[0].Start, mid, mid, runs[0].End
	}
	return 0, cols/2, cols/2, cols
}

// findDigitRegions analyzes bar image and returns SpO2 and HR regions
// Uses column classification: empty, full-height (digits), bottom-only (labels)
func findDigitRegions(img gocv.Mat) (spo2 image.Rectangle, hr image.Rectangle) {
	rows, cols := img.Rows(), img.Cols()

	// Classify each column
	colTypes := make([]ColumnType, cols)
	for x := 0; x < cols; x++ {
		colTypes[x] = classifyColumn(img, x)
	}

	// Find contiguous regions of full-height columns (digit regions)
	type Region struct {
		Start, End int
	}
	var digitRegions []Region

	inRegion := false
	regionStart := 0
	for x := 0; x < cols; x++ {
		if colTypes[x] == ColFullHeight {
			if !inRegion {
				inRegion = true
				regionStart = x
			}
		} else {
			if inRegion {
				// End of region - must be at least 10% of width to count
				if x - regionStart > cols/10 {
					digitRegions = append(digitRegions, Region{regionStart, x})
				}
				inRegion = false
			}
		}
	}
	// Don't forget last region
	if inRegion && cols - regionStart > cols/10 {
		digitRegions = append(digitRegions, Region{regionStart, cols})
	}

	// Debug output
	fmt.Printf("Found %d digit regions: ", len(digitRegions))
	for _, r := range digitRegions {
		pctStart := float64(r.Start) * 100 / float64(cols)
		pctEnd := float64(r.End) * 100 / float64(cols)
		fmt.Printf("[%d-%d (%.0f%%-%.0f%%)] ", r.Start, r.End, pctStart, pctEnd)
	}
	fmt.Println()

	// We expect 2 digit regions: SpO2 (left) and HR (right)
	if len(digitRegions) >= 2 {
		// First region = SpO2, Second region = HR
		spo2 = image.Rect(digitRegions[0].Start, 0, digitRegions[0].End, rows)
		hr = image.Rect(digitRegions[1].Start, 0, digitRegions[1].End, rows)
	} else if len(digitRegions) == 1 {
		// Only one region found - split it in half
		mid := (digitRegions[0].Start + digitRegions[0].End) / 2
		spo2 = image.Rect(digitRegions[0].Start, 0, mid, rows)
		hr = image.Rect(mid, 0, digitRegions[0].End, rows)
	} else {
		// Fallback - split whole image
		spo2 = image.Rect(0, 0, cols/2, rows)
		hr = image.Rect(cols/2, 0, cols, rows)
	}

	return spo2, hr
}

// CropRegion extracts a region from a Mat
func CropRegion(mat gocv.Mat, r Region) gocv.Mat {
	rect := image.Rect(r.X, r.Y, r.X+r.W, r.Y+r.H)
	region := mat.Region(rect)
	cropped := gocv.NewMat()
	region.CopyTo(&cropped)
	region.Close()
	return cropped
}

// DigitPattern holds a contour's direction pattern and position
type DigitPattern struct {
	X       int    // leftmost X for sorting
	Pattern string // direction chain
}

// extractDigitPatterns extracts digit patterns from contours, filtering small ones and sorting left-to-right
// Also tracks hole positions (as percentage of width) and appends to the pattern
// filterRightEdge: if true, ignore contours starting past 80% of width (for right side noise)
func extractDigitPatterns(contours gocv.PointsVector, imgHeight, imgWidth int, filterRightEdge bool) []string {
	var patterns []DigitPattern
	var holes []image.Rectangle  // bounding boxes of potential holes
	var digitBounds image.Rectangle // combined bounding box of all digits

	// Size thresholds
	minDigitHeight := imgHeight * 25 / 100 // outer digit: at least 25% height
	minDigitWidth := imgWidth * 15 / 100   // outer digit: at least 15% width
	minHoleSize := imgHeight * 10 / 100    // hole: at least 10% in either dimension

	// First pass: find digits and potential holes
	for i := 0; i < contours.Size(); i++ {
		contour := contours.At(i)
		rect := gocv.BoundingRect(contour)

		isDigit := rect.Dy() >= minDigitHeight && rect.Dx() >= minDigitWidth
		if filterRightEdge && rect.Min.X >= imgWidth*80/100 {
			isDigit = false
		}
		isHole := rect.Dy() >= minHoleSize && rect.Dx() >= minHoleSize && rect.Min.Y < imgHeight*60/100

		if isHole && !isDigit {
			holes = append(holes, rect)
			continue
		}

		if !isDigit {
			continue
		}

		// Filter contours that start too low (labels are at bottom)
		if rect.Min.Y > imgHeight*60/100 {
			continue
		}

		// Expand digit bounds
		if digitBounds.Empty() {
			digitBounds = rect
		} else {
			digitBounds = digitBounds.Union(rect)
		}

		// Simplify and get direction pattern
		epsilon := 0.01 * gocv.ArcLength(contour, true)
		approx := gocv.ApproxPolyDP(contour, epsilon, true)
		pts := approx.ToPoints()

		if len(pts) > 2 {
			dirs := pointsToDirections(pts)
			pattern := strings.Join(dirs, "")
			if len(pattern) > 0 {
				patterns = append(patterns, DigitPattern{X: rect.Min.X, Pattern: pattern})
			}
		}
	}

	// Sort by X position (left to right)
	for i := 0; i < len(patterns)-1; i++ {
		for j := i + 1; j < len(patterns); j++ {
			if patterns[j].X < patterns[i].X {
				patterns[i], patterns[j] = patterns[j], patterns[i]
			}
		}
	}

	// Return patterns in order, with hole positions appended
	result := make([]string, len(patterns))
	for i, p := range patterns {
		result[i] = p.Pattern
	}

	// Filter holes: only keep those inside the digit bounding box
	var holePositions []int
	for _, hole := range holes {
		// Check if hole center is inside digit bounds
		centerX := hole.Min.X + hole.Dx()/2
		centerY := hole.Min.Y + hole.Dy()/2
		if centerX >= digitBounds.Min.X && centerX <= digitBounds.Max.X &&
			centerY >= digitBounds.Min.Y && centerY <= digitBounds.Max.Y {
			// Record X position as percentage of digit width (0-9)
			relX := centerX - digitBounds.Min.X
			pct := relX * 10 / digitBounds.Dx()
			if pct > 9 {
				pct = 9
			}
			holePositions = append(holePositions, pct)
		}
	}

	// Sort and append hole positions (e.g., "h35" means holes at 30%, 50% of digit width)
	// Skip if >2 holes (likely alarm bell or other icon, not a digit)
	if len(holePositions) > 2 {
		return nil
	}
	if len(holePositions) > 0 {
		for i := 0; i < len(holePositions)-1; i++ {
			for j := i + 1; j < len(holePositions); j++ {
				if holePositions[j] < holePositions[i] {
					holePositions[i], holePositions[j] = holePositions[j], holePositions[i]
				}
			}
		}
		h := "h"
		for _, p := range holePositions {
			h += fmt.Sprintf("%d", p)
		}
		result = append(result, h)
	}
	return result
}

// pointsToDirections converts contour points to compass directions (normalized - no consecutive duplicates)
// Skips segments shorter than minDist pixels
func pointsToDirections(pts []image.Point) []string {
	const minDist = 10 // minimum distance between points
	dirs := make([]string, 0, len(pts))
	var lastDir string
	for i := 0; i < len(pts); i++ {
		p1 := pts[i]
		p2 := pts[(i+1)%len(pts)] // wrap to first point

		dx := p2.X - p1.X
		dy := p2.Y - p1.Y

		// Skip segments that are too short
		if abs(dx) < minDist && abs(dy) < minDist {
			continue
		}

		dir := getDirection(dx, dy)
		// Normalize: skip consecutive duplicates
		if dir != lastDir {
			dirs = append(dirs, dir)
			lastDir = dir
		}
	}
	return dirs
}

// getDirection returns direction as digit 1-4 (cardinal directions only)
// 1=N, 2=E, 3=S, 4=W
func getDirection(dx, dy int) string {
	if dx == 0 && dy == 0 {
		return ""
	}

	adx, ady := abs(dx), abs(dy)

	// Pick dominant axis
	if adx >= ady {
		// Horizontal dominant
		if dx > 0 {
			return "2" // E
		}
		return "4" // W
	}
	// Vertical dominant
	if dy > 0 {
		return "3" // S
	}
	return "1" // N
}

/* OLD 8-direction version (keep for easy revert):
// getDirection returns direction as digit 1-8 (clockwise from N)
// 1=N, 2=NE, 3=E, 4=SE, 5=S, 6=SW, 7=W, 8=NW
func getDirection8(dx, dy int) string {
	if dx == 0 && dy == 0 {
		return ""
	}

	adx, ady := abs(dx), abs(dy)

	// Cardinal directions (one axis dominates >2x)
	if adx > ady*2 {
		if dx > 0 {
			return "3" // E
		}
		return "7" // W
	}
	if ady > adx*2 {
		if dy > 0 {
			return "5" // S
		}
		return "1" // N
	}

	// Diagonal
	if dx > 0 && dy > 0 {
		return "4" // SE
	}
	if dx > 0 && dy < 0 {
		return "2" // NE
	}
	if dx < 0 && dy > 0 {
		return "6" // SW
	}
	return "8" // NW
}
*/

// zhangSuenThinning performs skeletonization using the Zhang-Suen algorithm
// Input: binary image (0 = background, 255 = foreground)
// Returns: thinned image (1-pixel wide skeleton)
func zhangSuenThinning(img gocv.Mat) gocv.Mat {
	rows, cols := img.Rows(), img.Cols()

	// Convert to binary array for easier manipulation
	data := make([][]byte, rows)
	for y := 0; y < rows; y++ {
		data[y] = make([]byte, cols)
		for x := 0; x < cols; x++ {
			if img.GetUCharAt(y, x) > 127 {
				data[y][x] = 1
			}
		}
	}

	// Neighbor offsets (P2,P3,P4,P5,P6,P7,P8,P9 clockwise from top)
	// P9 P2 P3
	// P8 P1 P4
	// P7 P6 P5
	dy := []int{-1, -1, 0, 1, 1, 1, 0, -1}
	dx := []int{0, 1, 1, 1, 0, -1, -1, -1}

	getP := func(y, x, i int) byte {
		ny, nx := y+dy[i], x+dx[i]
		if ny < 0 || ny >= rows || nx < 0 || nx >= cols {
			return 0
		}
		return data[ny][nx]
	}

	// Count 0->1 transitions in clockwise order
	transitions := func(y, x int) int {
		count := 0
		for i := 0; i < 8; i++ {
			if getP(y, x, i) == 0 && getP(y, x, (i+1)%8) == 1 {
				count++
			}
		}
		return count
	}

	// Count non-zero neighbors
	neighbors := func(y, x int) int {
		count := 0
		for i := 0; i < 8; i++ {
			count += int(getP(y, x, i))
		}
		return count
	}

	changed := true
	for changed {
		changed = false

		// Sub-iteration 1
		var toRemove []image.Point
		for y := 1; y < rows-1; y++ {
			for x := 1; x < cols-1; x++ {
				if data[y][x] != 1 {
					continue
				}
				n := neighbors(y, x)
				if n < 2 || n > 6 {
					continue
				}
				if transitions(y, x) != 1 {
					continue
				}
				// P2 * P4 * P6 == 0
				if getP(y, x, 0)*getP(y, x, 2)*getP(y, x, 4) != 0 {
					continue
				}
				// P4 * P6 * P8 == 0
				if getP(y, x, 2)*getP(y, x, 4)*getP(y, x, 6) != 0 {
					continue
				}
				toRemove = append(toRemove, image.Point{x, y})
			}
		}
		for _, p := range toRemove {
			data[p.Y][p.X] = 0
			changed = true
		}

		// Sub-iteration 2
		toRemove = nil
		for y := 1; y < rows-1; y++ {
			for x := 1; x < cols-1; x++ {
				if data[y][x] != 1 {
					continue
				}
				n := neighbors(y, x)
				if n < 2 || n > 6 {
					continue
				}
				if transitions(y, x) != 1 {
					continue
				}
				// P2 * P4 * P8 == 0
				if getP(y, x, 0)*getP(y, x, 2)*getP(y, x, 6) != 0 {
					continue
				}
				// P2 * P6 * P8 == 0
				if getP(y, x, 0)*getP(y, x, 4)*getP(y, x, 6) != 0 {
					continue
				}
				toRemove = append(toRemove, image.Point{x, y})
			}
		}
		for _, p := range toRemove {
			data[p.Y][p.X] = 0
			changed = true
		}
	}

	// Convert back to Mat
	result := gocv.NewMatWithSize(rows, cols, gocv.MatTypeCV8U)
	for y := 0; y < rows; y++ {
		for x := 0; x < cols; x++ {
			if data[y][x] == 1 {
				result.SetUCharAt(y, x, 255)
			}
		}
	}
	return result
}

// traceSkeletonStrokes traces the skeleton and returns stroke directions
// Returns directions like "RDLU" (Right, Down, Left, Up) for a "9"
func traceSkeletonStrokes(skeleton gocv.Mat) string {
	rows, cols := skeleton.Rows(), skeleton.Cols()

	// Find all skeleton pixels
	var points []image.Point
	for y := 0; y < rows; y++ {
		for x := 0; x < cols; x++ {
			if skeleton.GetUCharAt(y, x) > 0 {
				points = append(points, image.Point{x, y})
			}
		}
	}

	if len(points) == 0 {
		return ""
	}

	// Build adjacency - 8-connected neighbors
	isSet := func(x, y int) bool {
		if x < 0 || x >= cols || y < 0 || y >= rows {
			return false
		}
		return skeleton.GetUCharAt(y, x) > 0
	}

	neighbors := func(p image.Point) []image.Point {
		var n []image.Point
		for dy := -1; dy <= 1; dy++ {
			for dx := -1; dx <= 1; dx++ {
				if dx == 0 && dy == 0 {
					continue
				}
				if isSet(p.X+dx, p.Y+dy) {
					n = append(n, image.Point{p.X + dx, p.Y + dy})
				}
			}
		}
		return n
	}

	// Find endpoint (pixel with exactly 1 neighbor) - prefer topmost, then leftmost
	var startPoint image.Point
	foundEndpoint := false
	for _, p := range points {
		if len(neighbors(p)) == 1 {
			if !foundEndpoint || p.Y < startPoint.Y || (p.Y == startPoint.Y && p.X < startPoint.X) {
				startPoint = p
				foundEndpoint = true
			}
		}
	}

	// If no endpoint (closed loop), start from topmost-leftmost point
	if !foundEndpoint {
		startPoint = points[0]
		for _, p := range points {
			if p.Y < startPoint.Y || (p.Y == startPoint.Y && p.X < startPoint.X) {
				startPoint = p
			}
		}
	}

	// Trace the skeleton from startPoint
	visited := make(map[image.Point]bool)
	var path []image.Point

	var trace func(p image.Point)
	trace = func(p image.Point) {
		if visited[p] {
			return
		}
		visited[p] = true
		path = append(path, p)

		for _, n := range neighbors(p) {
			if !visited[n] {
				trace(n)
			}
		}
	}
	trace(startPoint)

	// Convert path to directions - only emit when moved 10+ pixels from checkpoint
	if len(path) < 2 {
		return ""
	}

	const threshold = 20
	var result strings.Builder
	var lastDir string
	checkX, checkY := path[0].X, path[0].Y

	for _, p := range path {
		dx := p.X - checkX
		dy := p.Y - checkY

		// Determine dominant direction if we've moved enough
		var dir string
		if abs(dx) >= threshold || abs(dy) >= threshold {
			// Pick the dominant direction
			if abs(dy) > abs(dx) {
				if dy > 0 {
					dir = "D"
				} else {
					dir = "U"
				}
			} else {
				if dx > 0 {
					dir = "R"
				} else {
					dir = "L"
				}
			}
			// Reset checkpoint to current position
			checkX, checkY = p.X, p.Y
		}

		if dir != "" && dir != lastDir {
			result.WriteString(dir)
			lastDir = dir
		}
	}

	return result.String()
}

// analyzeHalf analyzes a half of a digit skeleton and returns shape descriptors for left/middle/right
// Returns a 3-char string like "|-|" (vertical, horizontal, vertical) or "|^|" (vertical, arch, vertical)
func analyzeHalf(skeleton gocv.Mat) string {
	rows, cols := skeleton.Rows(), skeleton.Cols()
	if rows < 3 || cols < 3 {
		return "..."
	}

	// Divide into left/middle/right thirds
	third := cols / 3

	// Count pixels in each vertical slice and their distribution
	analyzeThird := func(startX, endX int) rune {
		var pixels []image.Point
		for y := 0; y < rows; y++ {
			for x := startX; x < endX; x++ {
				if skeleton.GetUCharAt(y, x) > 0 {
					pixels = append(pixels, image.Point{x, y})
				}
			}
		}

		if len(pixels) < 2 {
			return '.' // empty
		}

		// Find Y span and X span
		minY, maxY := rows, 0
		minX, maxX := cols, 0
		for _, p := range pixels {
			if p.Y < minY { minY = p.Y }
			if p.Y > maxY { maxY = p.Y }
			if p.X < minX { minX = p.X }
			if p.X > maxX { maxX = p.X }
		}

		ySpan := maxY - minY
		xSpan := maxX - minX

		// Vertical stroke: tall and narrow
		if ySpan > rows/2 && xSpan < cols/4 {
			return '|'
		}

		// Horizontal stroke: wide and short
		if xSpan > (endX-startX)/2 && ySpan < rows/3 {
			return '-'
		}

		// Has content but not clearly vertical or horizontal
		return 'o' // curve/other
	}

	// For middle section, also detect arch vs bowl
	analyzeMiddle := func() rune {
		startX, endX := third, cols-third
		var pixels []image.Point
		for y := 0; y < rows; y++ {
			for x := startX; x < endX; x++ {
				if skeleton.GetUCharAt(y, x) > 0 {
					pixels = append(pixels, image.Point{x, y})
				}
			}
		}

		if len(pixels) < 2 {
			return '.' // empty
		}

		// Check if pixels are concentrated at top (arch) or bottom (bowl)
		topCount, bottomCount := 0, 0
		midY := rows / 2
		for _, p := range pixels {
			if p.Y < midY {
				topCount++
			} else {
				bottomCount++
			}
		}

		// Find horizontal span
		minX, maxX := cols, 0
		minY, maxY := rows, 0
		for _, p := range pixels {
			if p.X < minX { minX = p.X }
			if p.X > maxX { maxX = p.X }
			if p.Y < minY { minY = p.Y }
			if p.Y > maxY { maxY = p.Y }
		}
		xSpan := maxX - minX
		ySpan := maxY - minY

		// Horizontal line
		if xSpan > (endX-startX)/2 && ySpan < rows/3 {
			return '-'
		}

		// Arch (pixels at top, spans width)
		if topCount > bottomCount*2 && xSpan > (endX-startX)/3 {
			return '^'
		}

		// Bowl (pixels at bottom, spans width)
		if bottomCount > topCount*2 && xSpan > (endX-startX)/3 {
			return 'u'
		}

		return 'o' // other/curve
	}

	left := analyzeThird(0, third)
	middle := analyzeMiddle()
	right := analyzeThird(cols-third, cols)

	return string([]rune{left, middle, right})
}

// classifyDigit analyzes a digit skeleton and returns a fingerprint based on top/bottom half shapes
func classifyDigit(skeleton gocv.Mat) string {
	rows := skeleton.Rows()
	if rows < 6 {
		return ""
	}

	// Split into top and bottom halves
	midY := rows / 2
	topHalf := skeleton.Region(image.Rect(0, 0, skeleton.Cols(), midY))
	bottomHalf := skeleton.Region(image.Rect(0, midY, skeleton.Cols(), rows))

	topPattern := analyzeHalf(topHalf)
	bottomPattern := analyzeHalf(bottomHalf)

	topHalf.Close()
	bottomHalf.Close()

	return topPattern + "/" + bottomPattern
}

// isVerticalStroke checks if a skeleton is just a simple vertical stroke (for detecting "1")
func isVerticalStroke(skeleton gocv.Mat) bool {
	rows, cols := skeleton.Rows(), skeleton.Cols()
	if rows < 5 {
		return false
	}

	// Count pixels and check distribution
	var pixels []image.Point
	for y := 0; y < rows; y++ {
		for x := 0; x < cols; x++ {
			if skeleton.GetUCharAt(y, x) > 0 {
				pixels = append(pixels, image.Point{x, y})
			}
		}
	}

	if len(pixels) < 3 {
		return false
	}

	// Check X variance - should be small for vertical stroke
	minX, maxX := cols, 0
	minY, maxY := rows, 0
	for _, p := range pixels {
		if p.X < minX { minX = p.X }
		if p.X > maxX { maxX = p.X }
		if p.Y < minY { minY = p.Y }
		if p.Y > maxY { maxY = p.Y }
	}

	xSpan := maxX - minX
	ySpan := maxY - minY

	// Vertical: tall and narrow
	return ySpan > rows/2 && xSpan < cols/3
}

// getDigitBounds finds bounding box of components reaching top 20% (the actual digits, not noise)
func getDigitBounds(img gocv.Mat) image.Rectangle {
	rows, cols := img.Rows(), img.Cols()
	topThreshold := rows * 20 / 100

	// Find connected components on raw image
	labels := gocv.NewMat()
	numLabels := gocv.ConnectedComponents(img, &labels)
	defer labels.Close()

	// Check which labels reach the top
	reachesTop := make([]bool, numLabels)
	for y := 0; y < topThreshold; y++ {
		for x := 0; x < cols; x++ {
			label := int(labels.GetIntAt(y, x))
			if label > 0 {
				reachesTop[label] = true
			}
		}
	}

	// Find bounding box of all top-reaching components
	minX, minY := cols, rows
	maxX, maxY := 0, 0
	found := false
	for y := 0; y < rows; y++ {
		for x := 0; x < cols; x++ {
			label := int(labels.GetIntAt(y, x))
			if label > 0 && reachesTop[label] {
				found = true
				if x < minX { minX = x }
				if x > maxX { maxX = x }
				if y < minY { minY = y }
				if y > maxY { maxY = y }
			}
		}
	}

	if !found {
		return image.Rect(0, 0, cols, rows) // fallback to full image
	}

	// Add padding on all sides to avoid cutting off edges
	pad := 3
	minX = max(0, minX-pad)
	minY = max(0, minY-pad)
	maxX = min(cols, maxX+pad)
	maxY = min(rows, maxY+pad)

	return image.Rect(minX, minY, maxX, maxY)
}

func max(a, b int) int {
	if a > b { return a }
	return b
}

// HoleInfo holds information about a detected hole
type HoleInfo struct {
	CenterX, CenterY int     // center position
	Area             int     // pixel count
	BoundsW, BoundsH int     // bounding box size
}

// findHoles finds enclosed regions (holes) and returns their properties
// Uses connected components on inverted image, excluding border-touching regions
func findHoles(img gocv.Mat) []HoleInfo {
	rows, cols := img.Rows(), img.Cols()
	if rows < 3 || cols < 3 {
		return nil
	}

	// Invert image (holes become white)
	inverted := gocv.NewMat()
	gocv.BitwiseNot(img, &inverted)
	defer inverted.Close()

	// Find connected components
	labels := gocv.NewMat()
	numLabels := gocv.ConnectedComponents(inverted, &labels)
	defer labels.Close()

	if numLabels <= 1 {
		return nil
	}

	// Check which labels touch the border (those are background, not holes)
	touchesBorder := make([]bool, numLabels)
	for x := 0; x < cols; x++ {
		touchesBorder[int(labels.GetIntAt(0, x))] = true
		touchesBorder[int(labels.GetIntAt(rows-1, x))] = true
	}
	for y := 0; y < rows; y++ {
		touchesBorder[int(labels.GetIntAt(y, 0))] = true
		touchesBorder[int(labels.GetIntAt(y, cols-1))] = true
	}

	// Analyze each hole
	var holes []HoleInfo
	for label := 1; label < numLabels; label++ {
		if touchesBorder[label] {
			continue
		}

		// Find bounds and count pixels
		minX, maxX := cols, 0
		minY, maxY := rows, 0
		sumX, sumY, count := 0, 0, 0

		for y := 0; y < rows; y++ {
			for x := 0; x < cols; x++ {
				if int(labels.GetIntAt(y, x)) == label {
					if x < minX { minX = x }
					if x > maxX { maxX = x }
					if y < minY { minY = y }
					if y > maxY { maxY = y }
					sumX += x
					sumY += y
					count++
				}
			}
		}

		if count < 4 { // ignore tiny specks
			continue
		}

		holes = append(holes, HoleInfo{
			CenterX: sumX / count,
			CenterY: sumY / count,
			Area:    count,
			BoundsW: maxX - minX + 1,
			BoundsH: maxY - minY + 1,
		})
	}

	return holes
}

// WhiteRun represents a contiguous white region at a scan line
type WhiteRun struct {
	Start, End int
}

func (r WhiteRun) Width() int { return r.End - r.Start }
func (r WhiteRun) Mid() int   { return (r.Start + r.End) / 2 }

// findWhiteRuns scans a row and returns all white runs
// Merges runs separated by tiny gaps (< 3px) which are likely rendering artifacts
func findWhiteRuns(img gocv.Mat, y int) []WhiteRun {
	cols := img.Cols()
	var runs []WhiteRun

	inWhite := false
	runStart := 0

	for x := 0; x <= cols; x++ {
		isWhite := x < cols && img.GetUCharAt(y, x) > 128
		if isWhite && !inWhite {
			runStart = x
			inWhite = true
		} else if !isWhite && inWhite {
			runs = append(runs, WhiteRun{runStart, x})
			inWhite = false
		}
	}

	// Merge runs separated by tiny gaps (< 3 pixels)
	if len(runs) > 1 {
		merged := []WhiteRun{runs[0]}
		for i := 1; i < len(runs); i++ {
			gap := runs[i].Start - merged[len(merged)-1].End
			if gap < 3 {
				// Merge with previous run
				merged[len(merged)-1].End = runs[i].End
			} else {
				merged = append(merged, runs[i])
			}
		}
		runs = merged
	}

	return runs
}

// recognizeNumber uses scan-line analysis to find and classify digits
// Strategy: Find "1"s first (narrow at both scan lines), then divide rest equally
func recognizeNumber(img gocv.Mat) string {
	rows, cols := img.Rows(), img.Cols()
	if rows < 10 || cols < 5 {
		return ""
	}

	y2 := rows * 2 / 5 // upper scan
	y4 := rows * 4 / 5 // lower scan

	runs2 := findWhiteRuns(img, y2)
	runs4 := findWhiteRuns(img, y4)

	// Debug
	fmt.Printf("    y=%d: ", y2)
	for _, r := range runs2 {
		fmt.Printf("[%d-%d](%d) ", r.Start, r.End, r.Width())
	}
	fmt.Printf("\n    y=%d: ", y4)
	for _, r := range runs4 {
		fmt.Printf("[%d-%d](%d) ", r.Start, r.End, r.Width())
	}
	fmt.Println()

	// Find "1"s: narrow runs at both scan lines at similar X position
	// A "1" is narrow (<20px) at both levels
	type OneDigit struct {
		X, W int // position and width
	}
	var ones []OneDigit

	for _, r2 := range runs2 {
		if r2.Width() > 20 {
			continue // not narrow enough for "1"
		}
		// Find matching narrow run at y4
		for _, r4 := range runs4 {
			if r4.Width() > 20 {
				continue
			}
			// Check if they overlap in X (same digit)
			overlap := min(r2.End, r4.End) - max(r2.Start, r4.Start)
			if overlap > 5 {
				// Found a "1"
				x := min(r2.Start, r4.Start)
				w := max(r2.End, r4.End) - x
				ones = append(ones, OneDigit{x, w})
				break
			}
		}
	}

	// Find content bounds (first and last white pixel columns)
	contentStart, contentEnd := cols, 0
	for _, r := range runs2 {
		if r.Start < contentStart {
			contentStart = r.Start
		}
		if r.End > contentEnd {
			contentEnd = r.End
		}
	}
	for _, r := range runs4 {
		if r.Start < contentStart {
			contentStart = r.Start
		}
		if r.End > contentEnd {
			contentEnd = r.End
		}
	}

	// Build digit list
	type Digit struct {
		Start, End   int
		IsOne        bool
		Runs2, Runs4 []WhiteRun
	}
	var digits []Digit

	// Sort ones by X
	for i := 0; i < len(ones); i++ {
		for j := i + 1; j < len(ones); j++ {
			if ones[j].X < ones[i].X {
				ones[i], ones[j] = ones[j], ones[i]
			}
		}
	}

	// Find digit boundaries using GAPS at BOTH scan lines
	// A gap >15px at EITHER line suggests a digit boundary
	gapThreshold := 15

	// Special case: if y=58 (4/5) has one merged run >80px but y=29 (2/5) has multiple runs,
	// the digits are touching at the bottom. Use a smarter approach for y=29 gaps.
	y4Merged := len(runs4) == 1 && runs4[0].Width() > 80

	// Collect split points from both scan lines
	splitSet := make(map[int]bool)

	for i := 1; i < len(runs2); i++ {
		gap := runs2[i].Start - runs2[i-1].End
		threshold := gapThreshold

		if y4Merged && len(runs2) >= 3 {
			// When digits merge at bottom, use width analysis:
			// A narrow run following a gap is likely "other side of hole", not a new digit
			// Don't split before narrow runs (they're continuations of previous digit)
			prevWidth := runs2[i-1].Width()
			nextWidth := runs2[i].Width()

			if nextWidth < prevWidth*2/3 {
				// Next run is significantly narrower - likely part of same digit (hole)
				continue // Skip this gap
			}
			threshold = 8 // Lower threshold for digit boundaries
		}

		if gap > threshold {
			splitSet[(runs2[i-1].End+runs2[i].Start)/2] = true
		}
	}
	for i := 1; i < len(runs4); i++ {
		gap := runs4[i].Start - runs4[i-1].End
		// Use lower threshold for y=58 gaps - they often represent real digit boundaries
		// even when the gap is small (like between "9" and "2")
		if gap > 2 {
			splitSet[(runs4[i-1].End+runs4[i].Start)/2] = true
		}
	}

	// Convert to sorted list
	var splits []int
	for sp := range splitSet {
		splits = append(splits, sp)
	}
	for i := 0; i < len(splits); i++ {
		for j := i + 1; j < len(splits); j++ {
			if splits[j] < splits[i] {
				splits[i], splits[j] = splits[j], splits[i]
			}
		}
	}

	// Remove splits that are too close together (<20px)
	var filteredSplits []int
	for _, sp := range splits {
		if len(filteredSplits) == 0 || sp-filteredSplits[len(filteredSplits)-1] > 20 {
			filteredSplits = append(filteredSplits, sp)
		}
	}
	splits = filteredSplits

	// Create digit boundaries
	totalWidth := contentEnd - contentStart
	pos := contentStart
	for _, sp := range splits {
		if sp > pos && sp < contentEnd {
			digits = append(digits, Digit{Start: pos, End: sp})
			pos = sp
		}
	}
	digits = append(digits, Digit{Start: pos, End: contentEnd})

	// SIMPLIFIED APPROACH: Numbers are either 2 digits OR 3 digits starting with "1"
	// Detect leading "1" using multiple strategies:
	// 1. Narrow region at BOTH scan lines followed by gap at both (ideal case)
	// 2. Narrow first run at y=58 followed by gap (when "1" merges with next digit at y=29)

	hasLeadingOne := false
	oneEndPos := contentStart

	// Strategy 1: Scan columns looking for narrow both-white region followed by both-dark gap
	scanLimit := contentStart + totalWidth*3/10
	inWhite := false
	whiteStart := 0

	for x := contentStart; x < scanLimit; x++ {
		isWhite := img.GetUCharAt(y2, x) > 128 && img.GetUCharAt(y4, x) > 128

		if isWhite && !inWhite {
			whiteStart = x
			inWhite = true
		} else if !isWhite && inWhite {
			whiteWidth := x - whiteStart
			if whiteWidth > 0 && whiteWidth < totalWidth/4 {
				minGapWidth := totalWidth / 20
				trueGapWidth := 0
				for x2 := x; x2 < contentEnd; x2++ {
					bothDark := img.GetUCharAt(y2, x2) <= 128 && img.GetUCharAt(y4, x2) <= 128
					if bothDark {
						trueGapWidth++
					} else {
						break
					}
				}
				if trueGapWidth >= minGapWidth {
					hasLeadingOne = true
					oneEndPos = x
				}
			}
			break
		}
	}

	// Strategy 2: If not found, check if first run at y=58 is narrow with gap after
	// This catches "1" that merges with next digit at y=29 but separates at y=58
	// IMPORTANT: Also verify y=29 has a narrow first run - this distinguishes "1" from "9"
	// "9" has narrow stem at y=58 but WIDE top with hole at y=29
	// "1" is narrow at BOTH scan lines
	if !hasLeadingOne && len(runs4) >= 2 && len(runs2) >= 1 {
		firstRun4 := runs4[0]
		firstRunWidth4 := firstRun4.Width()
		gapAfterFirst := runs4[1].Start - firstRun4.End

		// Check y=29 first run is also narrow (< 25% of total)
		// "1" should be narrow at both lines; "9" has wide top
		firstRun2 := runs2[0]
		firstRunWidth2 := firstRun2.Width()

		// "1" at y=58 is narrow (< 20% of total) and positioned in first third
		// Relaxed position check to 33% to handle merged displays
		if firstRunWidth4 < totalWidth/5 && firstRun4.End < contentStart+totalWidth/3 {
			// Additional check: y=29 first run must also be narrow (< 30% of total)
			// This prevents "9" (which has wide top) from being detected as "1"
			// Relaxed from 25% to 30% because "1" can merge slightly with next digit
			if firstRunWidth2 < totalWidth*3/10 {
				if gapAfterFirst > totalWidth/30 || len(runs4) >= 3 {
					hasLeadingOne = true
					// Use second run's start as boundary (where digit 2 begins)
					oneEndPos = runs4[1].Start
				}
			}
		}
	}

	// Strategy 3: Detect 3-digit numbers when "1" and "0" merge at y=58
	// This happens when the "1" stem connects with the "0" at the bottom
	// Indicators: wide total content (>100px), 3+ runs at y=29, first y=58 run is very wide
	if !hasLeadingOne && totalWidth > 100 && len(runs2) >= 3 && len(runs4) >= 2 {
		// For "104": y=58 shows [1+0 merged ~50%] [4 ~35%]
		// The first run being very wide (> 40% of total) suggests merged digits
		firstRun4 := runs4[0]
		firstRunWidth := firstRun4.Width()

		// If first run is very wide (> 40% of content), it's likely "1"+"0" merged
		// This distinguishes from 2-digit numbers where each digit is ~50%
		if firstRunWidth > totalWidth*40/100 && firstRunWidth < totalWidth*65/100 {
			hasLeadingOne = true
			// For merged "1"+"0", split at 1/3 of total width for the "1"
			oneEndPos = contentStart + totalWidth/3
		}
	}

	// Build digit list based on detection
	digits = nil
	if hasLeadingOne {
		// 3 digits: "1" + 2 more digits
		// For 3 runs at y=58: run0="1", run1="0", run2="2" (each run = one digit)
		// Use run starts as digit boundaries
		digits = append(digits, Digit{Start: contentStart, End: oneEndPos, IsOne: true})

		secondEnd := (oneEndPos + contentEnd) / 2 // fallback
		if len(runs4) >= 3 {
			// Digit 2 ends where digit 3's run begins
			secondEnd = runs4[2].Start
		} else if len(runs4) == 2 {
			// Strategy 3 case: 2 runs at y=58 where first is merged "1"+"0"
			// Use last run's start as boundary (that's where digit 3 begins)
			lastRunStart := runs4[1].Start
			// Only use if it's in the right 40% of the content (digit 3 area)
			if lastRunStart > contentStart+totalWidth*55/100 {
				secondEnd = lastRunStart
			}
		}

		digits = append(digits, Digit{Start: oneEndPos, End: secondEnd})
		digits = append(digits, Digit{Start: secondEnd, End: contentEnd})
	} else {
		// 2 digits - split at the LARGEST gap at y=51 that's in the middle region (30-70%)
		// This is the digit boundary; gaps outside this range are likely artifacts
		if len(runs4) >= 2 {
			// Find the largest gap between consecutive runs that's in the middle region
			maxGap := 0
			mid := (contentStart + contentEnd) / 2
			splitPos := mid // fallback to center
			minSplit := contentStart + totalWidth*30/100
			maxSplit := contentStart + totalWidth*70/100

			for i := 1; i < len(runs4); i++ {
				gapMid := (runs4[i-1].End + runs4[i].Start) / 2
				// Only consider gaps in the middle region
				if gapMid >= minSplit && gapMid <= maxSplit {
					gap := runs4[i].Start - runs4[i-1].End
					if gap > maxGap {
						maxGap = gap
						splitPos = gapMid
					}
				}
			}
			digits = append(digits, Digit{Start: contentStart, End: splitPos})
			digits = append(digits, Digit{Start: splitPos, End: contentEnd})
		} else {
			// Equal split
			mid := (contentStart + contentEnd) / 2
			digits = append(digits, Digit{Start: contentStart, End: mid})
			digits = append(digits, Digit{Start: mid, End: contentEnd})
		}
	}

	// NOTE: Previously we checked for detected "1"s and marked digits as IsOne.
	// This caused false positives when narrow runs from other digits (like "4", "5")
	// were mistakenly identified as "1"s. Now we rely solely on the leading "1"
	// detection strategies (1/2/3) which only mark the FIRST digit as IsOne when
	// it's a 3-digit number starting with "1". The scan-line classification
	// handles all other "1" detection.

	// Assign runs to digits with CLIPPING to digit boundaries
	// This handles runs that span multiple digits (e.g., merged "1" + "0")
	// Convert run coordinates to be RELATIVE to digit start
	// Use different thresholds: y=29 needs 8px (avoid boundary artifacts),
	// y=58 needs only 4px (to catch narrow hole edges like right side of "0")
	minRunWidth2 := 8 // for y=29
	minRunWidth4 := 4 // for y=58

	for i := range digits {
		d := &digits[i]

		for _, r := range runs2 {
			overlapStart := max(r.Start, d.Start)
			overlapEnd := min(r.End, d.End)
			if overlapEnd-overlapStart >= minRunWidth2 {
				relRun := WhiteRun{Start: overlapStart - d.Start, End: overlapEnd - d.Start}
				d.Runs2 = append(d.Runs2, relRun)
			}
		}

		for _, r := range runs4 {
			overlapStart := max(r.Start, d.Start)
			overlapEnd := min(r.End, d.End)
			if overlapEnd-overlapStart >= minRunWidth4 {
				relRun := WhiteRun{Start: overlapStart - d.Start, End: overlapEnd - d.Start}
				d.Runs4 = append(d.Runs4, relRun)
			}
		}
	}

	// Debug image
	vis := gocv.NewMat()
	gocv.CvtColor(img, &vis, gocv.ColorGrayToBGR)
	gocv.Line(&vis, image.Point{0, y2}, image.Point{cols, y2}, color.RGBA{255, 255, 0, 255}, 1)
	gocv.Line(&vis, image.Point{0, y4}, image.Point{cols, y4}, color.RGBA{255, 0, 255, 255}, 1)
	for _, d := range digits {
		c := color.RGBA{0, 255, 0, 255}
		if d.IsOne {
			c = color.RGBA{255, 0, 0, 255} // red for "1"s
		}
		gocv.Line(&vis, image.Point{d.Start, 0}, image.Point{d.Start, rows}, c, 1)
		gocv.Line(&vis, image.Point{d.End, 0}, image.Point{d.End, rows}, c, 1)
	}
	if cols < 100 {
		gocv.IMWrite("debug_cuts_left.png", vis)
	} else {
		gocv.IMWrite("debug_cuts_right.png", vis)
	}
	vis.Close()

	// Determine side for labeling
	side := "left"
	if cols > 100 {
		side = "right"
	}

	// Classify each digit
	var result string
	for i, d := range digits {
		var digit string
		if d.IsOne {
			digit = "1"
		} else {
			// Calculate fingerprint
			white2 := 0
			for _, r := range d.Runs2 {
				white2 += r.Width()
			}
			white4 := 0
			for _, r := range d.Runs4 {
				white4 += r.Width()
			}
			width := d.End - d.Start
			p2 := ((white2 * 100 / width) + 2) / 5 * 5
			p4 := ((white4 * 100 / width) + 2) / 5 * 5
			fingerprint := fmt.Sprintf("%d_%d", p2, p4)

			// Check pattern store first
			if patternStore != nil {
				if val, ok := patternStore.Lookup(fingerprint); ok {
					if val >= 0 && val <= 9 {
						digit = fmt.Sprintf("%d", val)
						fmt.Printf("    [%d-%d] fp=%s -> %s (learned)\n", d.Start, d.End, fingerprint, digit)
					} else {
						// val == -1 means ignore/unknown
						digit = "?"
						fmt.Printf("    [%d-%d] fp=%s -> ? (marked ignore)\n", d.Start, d.End, fingerprint)
					}
				}
			}

			// If not found in store, save for labeling and use best guess
			if digit == "" {
				// Save digit image for labeling
				digitImg := img.Region(image.Rect(d.Start, 0, d.End, rows))
				imgPath := fmt.Sprintf("digit_%s_%d_%s.png", side, i, fingerprint)
				gocv.IMWrite(imgPath, digitImg)

				// Add to unlabeled queue
				AddUnlabeled(fingerprint, imgPath, side)

				// Use fingerprint matcher as fallback
				digit = classifyByFingerprint(p2, p4)
				fmt.Printf("    [%d-%d] fp=%s -> %s (guessed, queued for labeling)\n", d.Start, d.End, fingerprint, digit)
			}
		}

		fmt.Printf("    [%d-%d] runs2=%d runs4=%d -> %s\n",
			d.Start, d.End, len(d.Runs2), len(d.Runs4), digit)
		if digit == "" || digit == "?" {
			return ""
		}
		result += digit
	}

	return result
}

// classifyByFingerprint uses p2/p4 percentages to guess the digit
// This is a fallback when the pattern isn't in the learned store
func classifyByFingerprint(p2, p4 int) string {
	// Fingerprint table: [p2, p4] -> digit
	// Based on 7-segment display patterns:
	//   _      At 2/5: middle area (left+right segments or holes)
	//  |_|     At 4/5: lower area (bottom segments)
	//  |_|
	//
	// Approximate expected percentages (will need calibration):
	// "0": ~40-50% at both (left+right segments, no middle bar visible)
	// "1": ~15-25% at both (just right stem)
	// "2": ~50-60% at 2/5 (top curve), ~70-80% at 4/5 (bottom bar)
	// "3": ~50-60% at both (curves on right)
	// "4": ~50-60% at 2/5 (left+right+middle), ~20-30% at 4/5 (just right stem)
	// "5": ~50-60% at 2/5 (top+left), ~40-50% at 4/5 (bottom curve)
	// "6": ~25-35% at 2/5 (left stem), ~60-70% at 4/5 (bottom loop)
	// "7": ~50-60% at 2/5 (top bar), ~20-30% at 4/5 (diagonal stem)
	// "8": ~40-50% at both (holes visible at both)
	// "9": ~40-50% at 2/5 (top loop), ~40-50% at 4/5 (stem + bottom)

	// Match by finding closest pattern
	// Format: {p2, p4, digit}
	patterns := []struct {
		p2, p4 int
		digit  string
	}{
		{45, 70, "0"}, // wide at both, slightly wider at bottom
		{20, 20, "1"}, // narrow at both
		{55, 75, "2"}, // medium top, wide bottom bar
		{55, 50, "3"}, // medium at both, right-heavy
		{55, 25, "4"}, // wide top (hole), narrow bottom (stem)
		{55, 45, "5"}, // medium top, medium-right bottom
		{30, 65, "6"}, // narrow top (stem), wide bottom (loop)
		{55, 25, "7"}, // wide top (bar), narrow bottom (stem) - same as 4
		{45, 45, "8"}, // medium at both (holes)
		{45, 50, "9"}, // medium top (loop), medium bottom (stem+bottom)
	}

	// Find best match using Manhattan distance
	bestDist := 1000
	bestDigit := "?"
	for _, pat := range patterns {
		dist := abs(p2-pat.p2) + abs(p4-pat.p4)
		if dist < bestDist {
			bestDist = dist
			bestDigit = pat.digit
		}
	}

	return bestDigit
}

// classifyDigitByFeatures classifies a single digit using hole and stroke analysis
// Returns the digit as a string ("0"-"9") or "" if uncertain
func classifyDigitByFeatures(img gocv.Mat) string {
	rows, cols := img.Rows(), img.Cols()
	if rows < 5 || cols < 3 {
		return ""
	}

	holes := findHoles(img)
	imgArea := rows * cols

	// Debug output
	fmt.Printf("  [Digit %dx%d] holes=%d", cols, rows, len(holes))
	for i, h := range holes {
		fmt.Printf(" h%d(y=%d,area=%d)", i, h.CenterY, h.Area)
	}

	// === HOLE-BASED CLASSIFICATION ===
	if len(holes) == 2 {
		fmt.Println(" -> 8 (2 holes)")
		return "8"
	}

	if len(holes) == 1 {
		hole := holes[0]
		midY := rows / 2

		// Hole in top half → 9
		if hole.CenterY < midY {
			fmt.Printf(" -> 9 (hole in top, centerY=%d < midY=%d)\n", hole.CenterY, midY)
			return "9"
		}
		// Hole in bottom half → 6
		if hole.CenterY > midY {
			fmt.Printf(" -> 6 (hole in bottom, centerY=%d > midY=%d)\n", hole.CenterY, midY)
			return "6"
		}
		// Hole exactly at middle - check size
		// Large hole (>10% of image area) → 0
		// Small hole → 4
		if hole.Area > imgArea*10/100 {
			fmt.Printf(" -> 0 (large hole at middle, area=%d > %d)\n", hole.Area, imgArea*10/100)
			return "0"
		}
		fmt.Printf(" -> 4 (small hole at middle, area=%d <= %d)\n", hole.Area, imgArea*10/100)
		return "4"
	}
	fmt.Print(" no-hole:")

	// === NO-HOLE CLASSIFICATION ===
	// First find the bounding box of white pixels
	minX, maxX := cols, 0
	minY, maxY := rows, 0
	for y := 0; y < rows; y++ {
		for x := 0; x < cols; x++ {
			if img.GetUCharAt(y, x) > 0 {
				if x < minX { minX = x }
				if x > maxX { maxX = x }
				if y < minY { minY = y }
				if y > maxY { maxY = y }
			}
		}
	}
	digitWidth := maxX - minX + 1
	digitHeight := maxY - minY + 1

	// Check for narrow vertical (1) FIRST - before other checks
	// "1" is tall and narrow (height/width ratio > 2)
	if digitWidth > 0 && digitHeight > digitWidth*2 {
		fmt.Printf(" narrow ratio h/w=%d/%d=%.1f -> 1\n", digitHeight, digitWidth, float64(digitHeight)/float64(digitWidth))
		return "1"
	}

	// Check for bar across top (5 or 7)
	topRegion := rows / 5
	topMinX, topMaxX := cols, 0
	for y := 0; y < topRegion; y++ {
		for x := 0; x < cols; x++ {
			if img.GetUCharAt(y, x) > 0 {
				if x < topMinX { topMinX = x }
				if x > topMaxX { topMaxX = x }
			}
		}
	}
	topSpan := topMaxX - topMinX
	hasTopBar := topSpan > cols*50/100 // spans more than 50% of width

	if hasTopBar {
		// Distinguish 5 vs 7: check bottom-left quadrant
		bottomLeftPixels := 0
		for y := rows * 2 / 3; y < rows; y++ {
			for x := 0; x < cols/2; x++ {
				if img.GetUCharAt(y, x) > 0 {
					bottomLeftPixels++
				}
			}
		}
		// 5 has content in bottom-left, 7 doesn't
		if bottomLeftPixels > rows*cols/50 {
			fmt.Printf(" topBar+bottomLeft=%d -> 5\n", bottomLeftPixels)
			return "5"
		}
		fmt.Printf(" topBar, no bottomLeft=%d -> 7\n", bottomLeftPixels)
		return "7"
	}

	// Check for horizontal bar in center (3)
	// 3 has content spanning the middle horizontally
	midY := rows / 2
	centerRegion := rows / 6
	centerMinX, centerMaxX := cols, 0
	for y := midY - centerRegion; y < midY+centerRegion; y++ {
		for x := 0; x < cols; x++ {
			if img.GetUCharAt(y, x) > 0 {
				if x < centerMinX { centerMinX = x }
				if x > centerMaxX { centerMaxX = x }
			}
		}
	}
	centerSpan := centerMaxX - centerMinX
	// Also check that 3 has right-side content in center
	rightCenterPixels := 0
	for y := midY - centerRegion; y < midY+centerRegion; y++ {
		for x := cols * 2 / 3; x < cols; x++ {
			if img.GetUCharAt(y, x) > 0 {
				rightCenterPixels++
			}
		}
	}
	if centerSpan > cols*40/100 && rightCenterPixels > rows*cols/100 {
		fmt.Printf(" centerSpan=%d, rightCenter=%d -> 3\n", centerSpan, rightCenterPixels)
		return "3"
	}

	// Leftover → 2
	fmt.Printf(" width=%d, centerSpan=%d -> 2 (leftover)\n", digitWidth, centerSpan)
	return "2"
}

// findHolePositions finds enclosed regions (holes) and returns their X positions as percentages (0-9)
// Uses connected components on inverted image, excluding border-touching regions
func findHolePositions(img gocv.Mat) []int {
	rows, cols := img.Rows(), img.Cols()
	if rows < 3 || cols < 3 {
		return nil
	}

	// Invert image (holes become white)
	inverted := gocv.NewMat()
	gocv.BitwiseNot(img, &inverted)
	defer inverted.Close()

	// Find connected components
	labels := gocv.NewMat()
	numLabels := gocv.ConnectedComponents(inverted, &labels)
	defer labels.Close()

	if numLabels <= 1 {
		return nil
	}

	// Check which labels touch the border (those are background, not holes)
	touchesBorder := make([]bool, numLabels)
	for x := 0; x < cols; x++ {
		touchesBorder[int(labels.GetIntAt(0, x))] = true
		touchesBorder[int(labels.GetIntAt(rows-1, x))] = true
	}
	for y := 0; y < rows; y++ {
		touchesBorder[int(labels.GetIntAt(y, 0))] = true
		touchesBorder[int(labels.GetIntAt(y, cols-1))] = true
	}

	// Find center X of each hole and convert to percentage (0-9)
	var positions []int
	for label := 1; label < numLabels; label++ {
		if touchesBorder[label] {
			continue
		}
		// Find bounding box of this hole
		minX, maxX := cols, 0
		for y := 0; y < rows; y++ {
			for x := 0; x < cols; x++ {
				if int(labels.GetIntAt(y, x)) == label {
					if x < minX { minX = x }
					if x > maxX { maxX = x }
				}
			}
		}
		// Center X as percentage (0-9)
		centerX := (minX + maxX) / 2
		pct := centerX * 10 / cols
		if pct > 9 {
			pct = 9
		}
		positions = append(positions, pct)
	}

	// Sort positions
	for i := 0; i < len(positions)-1; i++ {
		for j := i + 1; j < len(positions); j++ {
			if positions[j] < positions[i] {
				positions[i], positions[j] = positions[j], positions[i]
			}
		}
	}

	return positions
}

// extractSkeletonPatterns extracts patterns using spatial division approach:
// 1. Crop to digit bounds (exclude noise not reaching top)
// 2. Check for leading "1" (narrow stroke on left)
// 3. Split remaining area in half for the two main digits
// 4. Analyze each with top/bottom half shape method
// 5. Count holes and append to fingerprint
func extractSkeletonPatterns(img gocv.Mat) []string {
	// First find digit bounds (components reaching top), then crop
	bounds := getDigitBounds(img)
	croppedRegion := img.Region(bounds)
	cropped := gocv.NewMat()
	croppedRegion.CopyTo(&cropped)
	croppedRegion.Close()
	defer cropped.Close()

	// Find hole positions before skeletonization (on the filled shapes)
	holePositions := findHolePositions(cropped)

	// Now skeletonize the cropped digit area
	skeleton := zhangSuenThinning(cropped)
	defer skeleton.Close()

	rows, cols := skeleton.Rows(), skeleton.Cols()
	var patterns []string

	// Check for leading "1": look at leftmost 12% of image
	// A "1" must: have vertical content AND have a gap after it
	leadingWidth := cols * 12 / 100
	gapStart := leadingWidth
	gapEnd := cols * 20 / 100
	hasLeadingOne := false

	// Count pixels in leading region
	leadingPixels := 0
	for y := 0; y < rows; y++ {
		for x := 0; x < leadingWidth; x++ {
			if skeleton.GetUCharAt(y, x) > 0 {
				leadingPixels++
			}
		}
	}

	// Count pixels in gap region (should be empty for a true "1")
	gapPixels := 0
	for y := 0; y < rows; y++ {
		for x := gapStart; x < gapEnd; x++ {
			if skeleton.GetUCharAt(y, x) > 0 {
				gapPixels++
			}
		}
	}

	// If there are pixels in leading region AND gap is mostly empty
	if leadingPixels > rows/2 && gapPixels < rows/4 {
		// Check if pixels are vertically aligned (narrow X spread)
		minX, maxX := cols, 0
		for y := 0; y < rows; y++ {
			for x := 0; x < leadingWidth; x++ {
				if skeleton.GetUCharAt(y, x) > 0 {
					if x < minX { minX = x }
					if x > maxX { maxX = x }
				}
			}
		}
		if maxX-minX < leadingWidth {
			hasLeadingOne = true
			patterns = append(patterns, "1")
		}
	}

	// Determine where the two main digits start
	startX := 0
	if hasLeadingOne {
		startX = leadingWidth
	}

	// Split remaining area in half for two digits
	remainingWidth := cols - startX
	midX := startX + remainingWidth/2

	// Extract left digit region
	leftDigit := skeleton.Region(image.Rect(startX, 0, midX, rows))
	leftPattern := classifyDigit(leftDigit)
	leftDigit.Close()
	if leftPattern != "" {
		patterns = append(patterns, leftPattern)
	}

	// Extract right digit region
	rightDigit := skeleton.Region(image.Rect(midX, 0, cols, rows))
	rightPattern := classifyDigit(rightDigit)
	rightDigit.Close()
	if rightPattern != "" {
		patterns = append(patterns, rightPattern)
	}

	// Append hole positions (e.g., "h27" means holes at 20% and 70% of width)
	if len(holePositions) > 0 && len(holePositions) <= 4 {
		h := "h"
		for _, p := range holePositions {
			h += fmt.Sprintf("%d", p)
		}
		patterns = append(patterns, h)
	}

	return patterns
}