New quad algorithm: identify edges from contour orientation (#130)

Goal: improve precision of automatic document cropping by switching:
- from Douglas-Peucker algorithm (OpenCV's approxPolyDP) + a heuristic for documents missing a corner
- to an algorithm that looks for edges

* New quad algorithm: identify edges from contour orientation
* Performance optimization: reduce number of calls to trigonometric functions
* Performance: use a single threshold for live analysis
* Fix orientation of debug mask and compute it only if required
* Exclude quads that go out of the frame

app/src/main/java/org/fairscan/app/domain/ImageSegmentation.kt

@@ -29,6 +29,7 @@ import kotlinx.coroutines.sync.Mutex
 import kotlinx.coroutines.sync.withLock
 import kotlinx.coroutines.withContext
 import org.fairscan.app.data.Logger
+import org.fairscan.imageprocessing.ImageSize
 import org.fairscan.imageprocessing.Mask
 import org.opencv.core.CvType
 import org.opencv.core.Mat
@@ -39,7 +40,6 @@ import org.tensorflow.lite.support.common.ops.NormalizeOp
 import org.tensorflow.lite.support.image.ImageProcessor
 import org.tensorflow.lite.support.image.TensorImage
 import org.tensorflow.lite.support.image.ops.ResizeOp
-import org.tensorflow.lite.support.image.ops.Rot90Op
 import java.nio.ByteBuffer
 import java.nio.ByteOrder
 
@@ -73,13 +73,11 @@ class ImageSegmentationService(private val context: Context, private val logger:
     private fun runSegmentation(interpreter: Interpreter, bitmap: Bitmap, rotationDegrees: Int): SegmentationResult {
         val startTime = SystemClock.uptimeMillis()
 
-        val rotation = -rotationDegrees / 90
         val (_, h, w, _) = interpreter.getOutputTensor(0).shape()
         val imageProcessor =
             ImageProcessor
                 .Builder()
                 .add(ResizeOp(h, w, ResizeOp.ResizeMethod.BILINEAR))
-                .add(Rot90Op(rotation))
                 .add(NormalizeOp(127.5f, 127.5f)) // TODO check if it's correct
                 .build()
         val tensorImage = TensorImage(DataType.FLOAT32)
@@ -88,7 +86,11 @@ class ImageSegmentationService(private val context: Context, private val logger:
         val segmentResult = segment(interpreter, processedImage)
 
         val inferenceTime = SystemClock.uptimeMillis() - startTime
-        return SegmentationResult(segmentResult, inferenceTime)
+        return SegmentationResult(
+            segmentResult,
+            ImageSize(bitmap.width, bitmap.height),
+            rotationDegrees,
+            inferenceTime)
     }
 
     suspend fun runSegmentationAndReturn(bitmap: Bitmap, rotationDegrees: Int): SegmentationResult? {
@@ -163,10 +165,14 @@ class ImageSegmentationService(private val context: Context, private val logger:
             mask.put(0, 0, data)
             return mask
         }
+
+        fun maskSize() = ImageSize(width, height)
     }
 
     data class SegmentationResult(
         val segmentation: Segmentation,
+        val originalSize: ImageSize,
+        val rotationDegrees: Int,
         val inferenceTime: Long
     )
 }

app/src/main/java/org/fairscan/app/ui/screens/camera/CameraPreview.kt

@@ -211,7 +211,7 @@ fun bindCameraUseCases(
 
 @Composable
 fun AnalysisOverlay(liveAnalysisState: LiveAnalysisState, debugMode: Boolean) {
-    val binaryMask = liveAnalysisState.binaryMask ?: return
+    val maskSize = liveAnalysisState.maskSize ?: return
     val targetQuad = liveAnalysisState.stableQuad
     var displayedQuad by remember { mutableStateOf<Quad?>(null) }
     val quadColor = MaterialTheme.colorScheme.primary
@@ -233,14 +233,15 @@ fun AnalysisOverlay(liveAnalysisState: LiveAnalysisState, debugMode: Boolean) {
 
     Canvas(modifier = Modifier.fillMaxSize()) {
         if (debugMode) {
-            drawMask(this, binaryMask)
+            val binaryMask = liveAnalysisState.binaryMaskProvider.invoke()
+            binaryMask?.let { drawMask(this, it) }
         }
         displayedQuad?.let { quad ->
             val scaledQuad = quad.scaledTo(
-                fromWidth = binaryMask.width,
+                fromWidth = maskSize.width,
-                fromHeight = binaryMask.height,
+                fromHeight = maskSize.height,
-                toWidth = size.width.toInt(),
+                toWidth = size.width.toDouble(),
-                toHeight = size.height.toInt()
+                toHeight = size.height.toDouble()
             )
             scaledQuad.edges().forEach {
                 drawLine(quadColor, it.from.toOffset(), it.to.toOffset(), 10.0f)

app/src/main/java/org/fairscan/app/ui/screens/camera/CameraUiState.kt

@@ -16,12 +16,14 @@ package org.fairscan.app.ui.screens.camera
 
 import android.graphics.Bitmap
 import androidx.compose.runtime.Immutable
+import org.fairscan.imageprocessing.ImageSize
 import org.fairscan.imageprocessing.Quad
 
 @Immutable
 data class LiveAnalysisState(
     val inferenceTime: Long = 0L,
-    val binaryMask: Bitmap? = null,
+    val maskSize: ImageSize? = null,
+    val binaryMaskProvider: () -> Bitmap? = { -> null },
     val documentQuad: Quad? = null,
     val stableQuad: Quad? = null,
 )

app/src/main/java/org/fairscan/app/ui/screens/camera/CameraViewModel.kt

@@ -15,6 +15,7 @@
 package org.fairscan.app.ui.screens.camera
 
 import android.graphics.Bitmap
+import android.graphics.Matrix
 import androidx.camera.core.ImageProxy
 import androidx.core.graphics.createBitmap
 import androidx.lifecycle.ViewModel
@@ -72,16 +73,29 @@ class CameraViewModel(appContainer: AppContainer): ViewModel() {
             imageSegmentationService.segmentation
                 .filterNotNull()
                 .collect { result ->
-                    // TODO Should we really call toBinaryMask if it's used only in debug mode?
+                    val binaryMaskProvider = { ->
-                    val binaryMask = result.segmentation.toBinaryMask()
+                        var binaryMask: Bitmap = result.segmentation.toBinaryMask()
+                        if (result.rotationDegrees != 0) {
+                            binaryMask = rotateBitmap(binaryMask, result.rotationDegrees.toFloat())
+                        }
+                        binaryMask
+                    }
+
                     val rawQuad = detectDocumentQuad(
                         result.segmentation,
+                        result.originalSize,
                         isLiveAnalysis = true
+                    )?.rotate90(
+                        result.rotationDegrees / 90,
+                        result.segmentation.width,
+                        result.segmentation.height
                     )
+
                     val stableQuad = quadStabilizer.update(rawQuad)
                     _liveAnalysisState.value = LiveAnalysisState(
                         inferenceTime = result.inferenceTime,
-                        binaryMask = binaryMask,
+                        binaryMaskProvider = binaryMaskProvider,
+                        maskSize = result.segmentation.maskSize(),
                         documentQuad = rawQuad,
                         stableQuad = stableQuad,
                     )
@@ -145,13 +159,13 @@ class CameraViewModel(appContainer: AppContainer): ViewModel() {
 
     private suspend fun processCapturedImage(
         source: Bitmap,
-        rotationDegrees: Int
+        rotationDegrees: Int,
     ): CapturedPage? = withContext(Dispatchers.IO) {
         var result: CapturedPage? = null
         val segmentation = imageSegmentationService.runSegmentationAndReturn(source, 0)
         if (segmentation != null) {
             val mask = segmentation.segmentation
-            val quad = detectDocumentQuad(mask, isLiveAnalysis = false)
+            val quad = detectDocumentQuad(mask, segmentation.originalSize, isLiveAnalysis = false)
             if (quad != null) {
                 val resizedQuad = quad.scaledTo(mask.width, mask.height, source.width, source.height)
                 result = extractDocumentFromBitmap(source, resizedQuad, rotationDegrees, mask)
@@ -230,3 +244,9 @@ fun toBitmap(bgr: Mat): Bitmap {
     rgba.release()
     return bmp
 }
+
+fun rotateBitmap(source: Bitmap, angle: Float): Bitmap {
+    val matrix = Matrix()
+    matrix.postRotate(angle)
+    return Bitmap.createBitmap(source, 0, 0, source.getWidth(), source.getHeight(), matrix, true)
+}

evaluation/src/main/java/org/fairscan/evaluation/ColorDetectionEvaluator.kt

@@ -18,6 +18,7 @@ import org.fairscan.imageprocessing.detectDocumentQuad
 import org.fairscan.imageprocessing.extractDocument
 import org.fairscan.imageprocessing.isColoredDocument
 import org.fairscan.imageprocessing.scaledTo
+import org.fairscan.imageprocessing.toImageSize
 import org.opencv.imgcodecs.Imgcodecs
 import java.io.File
 
@@ -57,7 +58,7 @@ object ColorDetectionEvaluator {
 
             val mask = MatMask(maskMat)
 
-            val quad = detectDocumentQuad(mask, isLiveAnalysis = false)
+            val quad = detectDocumentQuad(mask, mat.size().toImageSize(), isLiveAnalysis = false)
                 ?.scaledTo(mask.width, mask.height, mat.width(), mat.height())
 
             if (quad == null) continue

evaluation/src/main/java/org/fairscan/evaluation/DatasetEvaluator.kt

@@ -19,6 +19,7 @@ import org.fairscan.imageprocessing.detectDocumentQuad
 import org.fairscan.imageprocessing.extractDocument
 import org.fairscan.imageprocessing.isColoredDocument
 import org.fairscan.imageprocessing.scaledTo
+import org.fairscan.imageprocessing.toImageSize
 import org.opencv.core.Mat
 import org.opencv.imgcodecs.Imgcodecs
 import java.io.File
@@ -68,7 +69,8 @@ object DatasetEvaluator {
 
             val mask = MatMask(maskMat)
 
-            val quad = detectDocumentQuad(mask, isLiveAnalysis = false)
+            val originalSize = inputMat.size().toImageSize()
+            val quad = detectDocumentQuad(mask, originalSize, isLiveAnalysis = false)
                 ?.scaledTo(mask.width, mask.height, inputMat.width(), inputMat.height())
 
             val corrected: Mat? = if (quad != null) {

evaluation/src/main/java/org/fairscan/evaluation/ExportQualityEvaluator.kt

@@ -18,6 +18,7 @@ import org.fairscan.imageprocessing.detectDocumentQuad
 import org.fairscan.imageprocessing.extractDocument
 import org.fairscan.imageprocessing.isColoredDocument
 import org.fairscan.imageprocessing.scaledTo
+import org.fairscan.imageprocessing.toImageSize
 import org.opencv.core.MatOfInt
 import org.opencv.imgcodecs.Imgcodecs
 import java.io.File
@@ -56,7 +57,8 @@ object ExportQualityEvaluator {
 
             val mask = MatMask(maskMat)
 
-            val quad = detectDocumentQuad(mask, isLiveAnalysis = false)
+            val originalSize = sourceMat.size().toImageSize()
+            val quad = detectDocumentQuad(mask, originalSize, isLiveAnalysis = false)
                 ?.scaledTo(mask.width, mask.height, sourceMat.width(), sourceMat.height())
             if (quad == null) {
                 println("Failed to detect quad for $imgName")

evaluation/src/main/java/org/fairscan/evaluation/QuadDetectionEvaluator.kt

@@ -18,6 +18,7 @@ import nu.pattern.OpenCV
 import org.fairscan.imageprocessing.detectDocumentQuad
 import org.fairscan.imageprocessing.scaledTo
 import org.fairscan.imageprocessing.toCv
+import org.fairscan.imageprocessing.toImageSize
 import org.opencv.core.Core
 import org.opencv.core.Mat
 import org.opencv.core.Scalar
@@ -63,7 +64,8 @@ object QuadDetectionEvaluator {
 
             val mask = MatMask(maskMat)
 
-            val quad = detectDocumentQuad(mask, isLiveAnalysis = false)
+            val originalSize = inputMat.size().toImageSize()
+            val quad = detectDocumentQuad(mask, originalSize, isLiveAnalysis = false)
                 ?.scaledTo(mask.width, mask.height, inputMat.width(), inputMat.height())
 
             val inputOut = File(outputDir, "${e.name}_input.jpg")

imageprocessing/src/main/java/org/fairscan/imageprocessing/DocumentDetection.kt

@@ -14,10 +14,11 @@
  */
 package org.fairscan.imageprocessing
 
-import org.fairscan.imageprocessing.quad.detectDocumentQuadFromProbmap
+import org.fairscan.imageprocessing.quad.findQuadFromContourOrientation
-import org.fairscan.imageprocessing.quad.findQuadFromRightAngles
 import org.fairscan.imageprocessing.quad.minAreaRect
+import org.fairscan.imageprocessing.quad.scoreQuadAgainstProbmap
 import org.opencv.core.Core
+import org.opencv.core.CvType
 import org.opencv.core.Mat
 import org.opencv.core.MatOfPoint
 import org.opencv.core.MatOfPoint2f
@@ -31,35 +32,78 @@ interface Mask {
     fun toMat(): Mat
 }
 
-fun detectDocumentQuad(mask: Mask, isLiveAnalysis: Boolean, minQuadAreaRatio: Double = 0.02): Quad? {
+fun detectDocumentQuad(mask: Mask, originalSize: ImageSize, isLiveAnalysis: Boolean): Quad? {
     val mat = mask.toMat()
-    val (biggest: MatOfPoint2f?, area) = biggestContour(mat)
+    // Best thresholds on test dataset: {0.95=146, 0.85=39, 0.75=35, 0.90=8, 0.70=1, 0.35=1}
-    var vertices: List<Point>?
-    if (biggest != null && biggest.total() == 4L && area > mask.width * mask.height * minQuadAreaRatio) {
-        vertices = biggest.toList()?.map { Point(it.x, it.y) }
-    } else {
-
-        // Fallback 1: adjust threshold
     val thresholds =
-            if (isLiveAnalysis) listOf(25.0, 50.0, 75.0) else (0..12).map { 0.2 + it * 0.05 }
+        if (isLiveAnalysis) listOf(0.9) else listOf(0.5, 0.7, 0.75, 0.8, 0.85, 0.9, 0.95)
-        vertices = detectDocumentQuadFromProbmap(mat, thresholds)
+    var vertices = findQuadFromOrientationWithAdaptiveThreshold(mat, originalSize, thresholds)
         ?.map { Point(it.x, it.y) }
-        if (vertices == null && biggest != null && biggest.total() > 4) {
-
-            // Fallback 2: look for right angles
-            val polygon = biggest.toList().map { Point(it.x, it.y) }
-            vertices = findQuadFromRightAngles(polygon, mask.width, mask.height)
     if (vertices == null && !isLiveAnalysis) {
-
+        // Fallback: bounding rectangle
-                // Fallback 3: bounding rectangle
+        val biggest = biggestContour(mat)
+        if (biggest != null) {
+            val polygon = biggest.toList().map { Point(it.x, it.y) }
             vertices = minAreaRect(polygon, mask.width, mask.height)
         }
     }
-    }
     return if (vertices?.size == 4) createQuad(vertices) else null
 }
 
-private fun biggestContour(mat: Mat): Pair<MatOfPoint2f?, Double> {
+fun findQuadFromOrientationWithAdaptiveThreshold(
+    maskMat: Mat, originalSize: ImageSize, thresholds: List<Double>
+): List<org.opencv.core.Point>? {
+    val probmapU8 = Mat()
+    val probmap = maskMat
+    probmap.convertTo(probmapU8, CvType.CV_8U, 255.0)
+    val probmapSmooth = Mat()
+    Imgproc.GaussianBlur(probmapU8, probmapSmooth, Size(3.0, 3.0), 0.0)
+
+    var bestQuad: List<org.opencv.core.Point>? = null
+    var bestScore = 0.0
+    for (thr in thresholds) {
+        val bin = Mat()
+        Imgproc.threshold(probmapSmooth, bin, thr * 255.0, 255.0, Imgproc.THRESH_BINARY)
+        val kernel = Imgproc.getStructuringElement(Imgproc.MORPH_ELLIPSE, Size(5.0, 5.0))
+        Imgproc.morphologyEx(bin, bin, Imgproc.MORPH_CLOSE, kernel)
+        val quad = findQuadFromOrientation(bin, originalSize)
+        if (quad != null && isValidQuad(quad, originalSize)) {
+            val probFloat = Mat()
+            probmap.convertTo(probFloat, CvType.CV_32F)
+            val score = scoreQuadAgainstProbmap(quad, probFloat, minQuadAreaRatio = 0.02)
+            if (score > bestScore) {
+                bestScore = score
+                bestQuad = quad
+            }
+        }
+        bin.release()
+    }
+
+    probmapSmooth.release()
+    probmapU8.release()
+    return bestQuad
+}
+
+fun isValidQuad(quad: List<org.opencv.core.Point>, originalSize: ImageSize): Boolean {
+    return quad.all {
+          it.x >= 0 && it.x <= originalSize.width
+       && it.y >= 0 && it.y <= originalSize.height
+    }
+}
+
+fun findQuadFromOrientation(maskMat: Mat, originalSize: ImageSize): List<org.opencv.core.Point>? {
+    val contour = biggestContour(maskMat)
+    contour?:return null
+
+    val scaleX = originalSize.width / maskMat.size().width
+    val scaleY = originalSize.height / maskMat.size().height
+
+    return findQuadFromContourOrientation(
+        contour.toList().map { org.opencv.core.Point(it.x * scaleX, it.y * scaleY) }
+    )?.map { org.opencv.core.Point(it.x / scaleX, it.y / scaleY) }
+}
+
+fun biggestContour(mat: Mat): MatOfPoint? {
     val refinedMask = refineMask(mat)
 
     val blurred = Mat()
@@ -70,24 +114,19 @@ private fun biggestContour(mat: Mat): Pair<MatOfPoint2f?, Double> {
 
     val contours = mutableListOf<MatOfPoint>()
     val hierarchy = Mat()
-    Imgproc.findContours(edges, contours, hierarchy, Imgproc.RETR_LIST, Imgproc.CHAIN_APPROX_SIMPLE)
+    Imgproc.findContours(edges, contours, hierarchy, Imgproc.RETR_LIST, Imgproc.CHAIN_APPROX_NONE)
 
-    var biggest: MatOfPoint2f? = null
+    var biggest: MatOfPoint? = null
     var maxArea = 0.0
 
     for (contour in contours) {
-        val contour2f = MatOfPoint2f(*contour.toArray())
+        val area = abs(Imgproc.contourArea(contour))
-        val peri = Imgproc.arcLength(contour2f, true)
-        val approx = MatOfPoint2f()
-        Imgproc.approxPolyDP(contour2f, approx, 0.02 * peri, true)
-
-        val area = abs(Imgproc.contourArea(approx))
         if (area > maxArea) {
             maxArea = area
-            biggest = approx
+            biggest = contour
         }
     }
-    return Pair(biggest, maxArea)
+    return biggest
 }
 
 /**
@@ -171,3 +210,6 @@ fun Point.toCv(): org.opencv.core.Point {
     return org.opencv.core.Point(x, y)
 }
 
+fun Size.toImageSize(): ImageSize {
+    return ImageSize(width, height)
+}

imageprocessing/src/main/java/org/fairscan/imageprocessing/Geometry.kt

@@ -103,3 +103,7 @@ fun Quad.scaledTo(fromWidth: Int, fromHeight: Int, toWidth: Int, toHeight: Int):
 fun Point.scaled(scaleX: Double, scaleY: Double): Point {
     return Point((x * scaleX), (y * scaleY))
 }
+
+data class ImageSize(val width: Double, val height: Double) {
+    constructor(width: Int, height: Int) : this (width.toDouble(), height.toDouble())
+}

imageprocessing/src/main/java/org/fairscan/imageprocessing/quad/AdaptiveThreshold.kt

@@ -1,129 +0,0 @@
-/*
- * Copyright 2025-2026 Pierre-Yves Nicolas
- *
- * This program is free software: you can redistribute it and/or modify it
- * under the terms of the GNU General Public License as published by the Free
- * Software Foundation, either version 3 of the License, or (at your option)
- * any later version.
- * This program is distributed in the hope that it will be useful, but WITHOUT
- * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
- * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
- * more details.
- * You should have received a copy of the GNU General Public License along with
- * this program. If not, see <https://www.gnu.org/licenses/>.
- */
-package org.fairscan.imageprocessing.quad
-
-import org.opencv.core.Mat
-import org.opencv.core.CvType
-import org.opencv.core.Size
-import org.opencv.core.Point
-import org.opencv.core.Scalar
-import org.opencv.core.MatOfPoint
-import org.opencv.core.MatOfPoint2f
-import org.opencv.core.Core
-import org.opencv.imgproc.Imgproc
-import kotlin.math.abs
-
-// Look for a threshold for which we find a quad in the mask
-fun detectDocumentQuadFromProbmap(
-    probmap: Mat,
-    thresholds: List<Double>,
-    useOtsu: Boolean = true,
-    minQuadAreaRatio: Double = 0.02
-): List<Point>? {
-    val probmapU8 = Mat()
-    probmap.convertTo(probmapU8, CvType.CV_8U, 255.0)
-    val probmapSmooth = Mat()
-    Imgproc.GaussianBlur(probmapU8, probmapSmooth, Size(3.0, 3.0), 0.0)
-
-    var bestScore = 0.0
-    var bestQuad: List<Point>? = null
-
-    // 1) Otsu
-    if (useOtsu) {
-        val otsu = Mat()
-        Imgproc.threshold(probmapSmooth, otsu, 0.0, 255.0, Imgproc.THRESH_BINARY + Imgproc.THRESH_OTSU)
-        val quad = findQuadFromBinaryMask(otsu, minQuadAreaRatio)
-        if (quad != null) {
-            val probFloat = Mat()
-            probmap.convertTo(probFloat, CvType.CV_32F)
-            val sc = scoreQuadAgainstProbmap(quad, probFloat)
-            if (sc > bestScore) {
-                bestScore = sc
-                bestQuad = quad
-            }
-        }
-    }
-
-    // 2) Threshold sweep
-    for (thr in thresholds) {
-        val bin = Mat()
-        Imgproc.threshold(probmapSmooth, bin, thr * 255.0, 255.0, Imgproc.THRESH_BINARY)
-        val kernel = Imgproc.getStructuringElement(Imgproc.MORPH_ELLIPSE, Size(5.0, 5.0))
-        Imgproc.morphologyEx(bin, bin, Imgproc.MORPH_CLOSE, kernel)
-        val quad = findQuadFromBinaryMask(bin, minQuadAreaRatio)
-        if (quad != null) {
-            val probFloat = Mat()
-            probmap.convertTo(probFloat, CvType.CV_32F)
-            val sc = scoreQuadAgainstProbmap(quad, probFloat)
-            if (sc > bestScore) {
-                bestScore = sc
-                bestQuad = quad
-            }
-        }
-    }
-
-    return bestQuad
-}
-
-// Fill polygon and return binary mask (0/1)
-fun makePolygonMask(size: Size, polygon: List<Point>): Mat {
-    val mask = Mat.zeros(size, CvType.CV_8U)
-    val pts = MatOfPoint(*polygon.toTypedArray())
-    Imgproc.fillPoly(mask, listOf(pts), Scalar(1.0))
-    return mask
-}
-
-// Compute score between quad and probmap
-fun scoreQuadAgainstProbmap(quad: List<Point>, probmap: Mat): Double {
-    val mask = makePolygonMask(probmap.size(), quad)
-    val maskFloat = Mat()
-    mask.convertTo(maskFloat, CvType.CV_32F)
-    val masked = Mat()
-    Core.multiply(probmap, maskFloat, masked)
-    val meanProb = Core.sumElems(masked).`val`[0] / Core.sumElems(maskFloat).`val`[0]
-    val areaRatio = Core.sumElems(maskFloat).`val`[0] / (probmap.rows() * probmap.cols())
-    return meanProb * (0.7 + 0.3 * areaRatio)
-}
-
-// Find largest quadrilateral in a binary mask
-fun findQuadFromBinaryMask(binMask: Mat, minQuadAreaRatio: Double = 0.02): List<Point>? {
-    val blurred = Mat()
-    Imgproc.GaussianBlur(binMask, blurred, Size(5.0, 5.0), 0.0)
-    val edges = Mat()
-    Imgproc.Canny(blurred, edges, 75.0, 200.0)
-
-    val contours = mutableListOf<MatOfPoint>()
-    Imgproc.findContours(edges, contours, Mat(), Imgproc.RETR_LIST, Imgproc.CHAIN_APPROX_SIMPLE)
-
-    var biggest: MatOfPoint2f? = null
-    var maxArea = 0.0
-    for (cnt in contours) {
-        val cnt2f = MatOfPoint2f(*cnt.toArray())
-        val peri = Imgproc.arcLength(cnt2f, true)
-        val approx = MatOfPoint2f()
-        Imgproc.approxPolyDP(cnt2f, approx, 0.02 * peri, true)
-        if (approx.rows() == 4) {
-            val area = abs(Imgproc.contourArea(approx))
-            if (area > maxArea) {
-                maxArea = area
-                biggest = approx
-            }
-        }
-    }
-    val totalArea = binMask.rows() * binMask.cols().toDouble()
-    return if (maxArea > totalArea * minQuadAreaRatio && biggest != null) {
-        biggest.toList()
-    } else null
-}

imageprocessing/src/main/java/org/fairscan/imageprocessing/quad/ContourOrientation.kt

@@ -0,0 +1,308 @@
+/*
+ * Copyright 2025-2026 Pierre-Yves Nicolas
+ *
+ * This program is free software: you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License as published by the Free
+ * Software Foundation, either version 3 of the License, or (at your option)
+ * any later version.
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
+ * more details.
+ * You should have received a copy of the GNU General Public License along with
+ * this program. If not, see <https://www.gnu.org/licenses/>.
+ */
+package org.fairscan.imageprocessing.quad
+
+import org.opencv.core.Point
+import kotlin.math.abs
+import kotlin.math.atan2
+import kotlin.math.cos
+import kotlin.math.hypot
+import kotlin.math.sin
+
+/**
+Instead of detecting corners (like Douglas-Peucker), this algorithm detects
+the four dominant sides of the document by segmenting the contour according
+to stable edge orientations, then fits lines and intersects them to
+reconstruct the quadrilateral.
+*/
+fun findQuadFromContourOrientation(
+    contour: List<Point>,
+    smoothWindow: Int = 5,
+    maxAngleVar: Double = Math.toRadians(5.0),
+    mergeAngle: Double = Math.toRadians(7.0),
+    minSideLengthRatio: Double = 0.02
+): List<Point>? {
+
+    if (contour.size < 20) return null
+
+    val angles = computeSmoothedAngles(contour, smoothWindow)
+
+    val perimeter = contour.zipWithNext { a, b -> hypot(b.x - a.x, b.y - a.y) }.sum()
+
+    val minLength = perimeter * minSideLengthRatio
+    val segments = extractSegments(contour, angles, maxAngleVar, minLength)
+    val mergedSegments = mergeSegments(segments, mergeAngle)
+    val dominantSegments = selectDominantSegments(
+        mergedSegments,
+        maxCount = 4,
+        minAngleSeparation = Math.toRadians(25.0)
+    )
+
+    if (dominantSegments.size != 4) return null
+
+    val lines = dominantSegments.map {
+        val points = if (it.start < it.end)
+            contour.subList(it.start, it.end)
+        else
+            contour.subList(it.start, contour.size) + contour.subList(0, it.end)
+        fitLine(points)
+    }
+
+    val corners = mutableListOf<Point>()
+    for (i in 0 until 4) {
+        val p = intersectLines(lines[i], lines[(i + 1) % 4])
+            ?: return null
+        corners += p
+    }
+    return corners
+}
+
+private fun normalizeAngle(a: Double): Double {
+    var x = a
+    while (x <= -Math.PI) x += 2 * Math.PI
+    while (x > Math.PI) x -= 2 * Math.PI
+    return x
+}
+
+private fun angleDiff(a: Double, b: Double): Double =
+    abs(normalizeAngle(a - b))
+
+private data class Line(
+    val p: Point,
+    val d: Point
+)
+
+private fun fitLine(points: List<Point>): Line {
+    val cx = points.map { it.x }.average()
+    val cy = points.map { it.y }.average()
+
+    var xx = 0.0
+    var xy = 0.0
+    var yy = 0.0
+
+    for (p in points) {
+        val dx = p.x - cx
+        val dy = p.y - cy
+        xx += dx * dx
+        xy += dx * dy
+        yy += dy * dy
+    }
+
+    val theta = 0.5 * atan2(2 * xy, xx - yy)
+    val dir = Point(cos(theta), sin(theta))
+
+    return Line(Point(cx, cy), dir)
+}
+
+private fun intersectLines(l1: Line, l2: Line): Point? {
+    val x1 = l1.p.x
+    val y1 = l1.p.y
+    val x2 = x1 + l1.d.x
+    val y2 = y1 + l1.d.y
+
+    val x3 = l2.p.x
+    val y3 = l2.p.y
+    val x4 = x3 + l2.d.x
+    val y4 = y3 + l2.d.y
+
+    val denom = (x1 - x2) * (y3 - y4) - (y1 - y2) * (x3 - x4)
+    if (abs(denom) < 1e-6) return null
+
+    val px =
+        ((x1*y2 - y1*x2)*(x3 - x4) - (x1 - x2)*(x3*y4 - y3*x4)) / denom
+    val py =
+        ((x1*y2 - y1*x2)*(y3 - y4) - (y1 - y2)*(x3*y4 - y3*x4)) / denom
+
+    return Point(px, py)
+}
+
+private data class ContourSegment(
+    val start: Int,
+    val end: Int,
+    val angle: Double,
+    val length: Double
+)
+
+private fun extractSegments(
+    contour: List<Point>,
+    angles: DoubleArray,
+    maxAngleVar: Double,
+    minLength: Double
+): List<ContourSegment> {
+
+    val n = contour.size
+    val result = mutableListOf<ContourSegment>()
+
+    val startIndex = findBestStartIndex(angles)
+
+    var start = startIndex
+    var refAngle = angles[startIndex]
+
+    fun segmentLength(s: Int, e: Int): Double {
+        var len = 0.0
+        var i = s
+        while (i != e) {
+            val j = (i + 1) % n
+            len += hypot(
+                contour[j].x - contour[i].x,
+                contour[j].y - contour[i].y
+            )
+            i = j
+        }
+        return len
+    }
+
+    var steps = 1
+    while (steps <= n) {
+        val idx = (startIndex + steps) % n
+
+        if (steps < n && angleDiff(angles[idx], refAngle) < maxAngleVar) {
+            refAngle = angleMean(refAngle, angles[idx])
+        } else {
+            val len = segmentLength(start, idx)
+            if (len >= minLength) {
+                result += ContourSegment(start, idx, refAngle, len)
+            }
+            start = idx
+            refAngle = angles[idx]
+        }
+
+        steps++
+    }
+
+    return result
+}
+
+private fun findBestStartIndex(angles: DoubleArray): Int {
+    val n = angles.size
+    var bestIndex = 0
+    var bestDelta = 0.0
+
+    for (i in 0 until n) {
+        val j = (i + 1) % n
+        val d = angleDiff(angles[i], angles[j])
+        if (d > bestDelta) {
+            bestDelta = d
+            bestIndex = j
+        }
+    }
+    return bestIndex
+}
+
+private fun angleMean(a: Double, b: Double): Double {
+    val x = cos(a) + cos(b)
+    val y = sin(a) + sin(b)
+    return atan2(y, x)
+}
+
+private fun computeSmoothedAngles(
+    contour: List<Point>,
+    window: Int
+): DoubleArray {
+    val n = contour.size
+
+    // --- Step 1: raw angles ---
+    val angles = DoubleArray(n)
+    for (i in 0 until n) {
+        val p0 = contour[(i - 1 + n) % n]
+        val p1 = contour[(i + 1) % n]
+        angles[i] = atan2(p1.y - p0.y, p1.x - p0.x)
+    }
+
+    // --- Step 2: precompute cos/sin ---
+    val cosA = DoubleArray(n)
+    val sinA = DoubleArray(n)
+    for (i in 0 until n) {
+        cosA[i] = cos(angles[i])
+        sinA[i] = sin(angles[i])
+    }
+
+    // --- Step 3: sliding window smoothing ---
+    val smooth = DoubleArray(n)
+
+    var sx = 0.0
+    var sy = 0.0
+
+    // initial window centered on index 0
+    for (k in -window..window) {
+        val idx = (k + n) % n
+        sx += cosA[idx]
+        sy += sinA[idx]
+    }
+
+    smooth[0] = atan2(sy, sx)
+
+    for (i in 1 until n) {
+        val outIdx = (i - window - 1 + n) % n
+        val inIdx  = (i + window) % n
+        sx -= cosA[outIdx]
+        sy -= sinA[outIdx]
+        sx += cosA[inIdx]
+        sy += sinA[inIdx]
+        smooth[i] = atan2(sy, sx)
+    }
+    return smooth
+}
+
+private fun mergeSegments(
+    segments: List<ContourSegment>,
+    angleThreshold: Double
+): List<ContourSegment> {
+    if (segments.isEmpty()) return emptyList()
+    if (segments.size <= 4) return segments
+
+    val merged = mutableListOf<ContourSegment>()
+    var cur = segments[0]
+
+    for (i in 1 until segments.size) {
+        val p = segments[i]
+        if (angleDiff(p.angle, cur.angle) < angleThreshold) {
+            cur = ContourSegment(
+                cur.start,
+                p.end,
+                angleMean(cur.angle, p.angle),
+                cur.length + p.length
+            )
+        } else {
+            merged += cur
+            cur = p
+        }
+    }
+    merged += cur
+    return merged
+}
+
+private fun selectDominantSegments(
+    segments: List<ContourSegment>,
+    maxCount: Int,
+    minAngleSeparation: Double
+): List<ContourSegment> {
+
+    val sorted = segments.sortedByDescending { it.length }
+    val selected = mutableListOf<ContourSegment>()
+
+    for (p in sorted) {
+        val tooClose = selected.any { s ->
+            angleDiff(p.angle, s.angle) < minAngleSeparation
+        }
+
+        if (!tooClose) {
+            selected += p
+            if (selected.size == maxCount) break
+        }
+    }
+
+    return selected.sortedBy { it.start }
+}

imageprocessing/src/main/java/org/fairscan/imageprocessing/quad/QuadScore.kt

@@ -0,0 +1,44 @@
+/*
+ * Copyright 2025-2026 Pierre-Yves Nicolas
+ *
+ * This program is free software: you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License as published by the Free
+ * Software Foundation, either version 3 of the License, or (at your option)
+ * any later version.
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
+ * more details.
+ * You should have received a copy of the GNU General Public License along with
+ * this program. If not, see <https://www.gnu.org/licenses/>.
+ */
+package org.fairscan.imageprocessing.quad
+
+import org.opencv.core.Core
+import org.opencv.core.CvType
+import org.opencv.core.Mat
+import org.opencv.core.MatOfPoint
+import org.opencv.core.Point
+import org.opencv.core.Scalar
+import org.opencv.core.Size
+import org.opencv.imgproc.Imgproc
+
+// Fill polygon and return binary mask (0/1)
+fun makePolygonMask(size: Size, polygon: List<Point>): Mat {
+    val mask = Mat.zeros(size, CvType.CV_8U)
+    val pts = MatOfPoint(*polygon.toTypedArray())
+    Imgproc.fillPoly(mask, listOf(pts), Scalar(1.0))
+    return mask
+}
+
+// Compute score between quad and probmap
+fun scoreQuadAgainstProbmap(quad: List<Point>, probmap: Mat, minQuadAreaRatio: Double): Double {
+    val mask = makePolygonMask(probmap.size(), quad)
+    val maskFloat = Mat()
+    mask.convertTo(maskFloat, CvType.CV_32F)
+    val masked = Mat()
+    Core.multiply(probmap, maskFloat, masked)
+    val meanProb = Core.sumElems(masked).`val`[0] / Core.sumElems(maskFloat).`val`[0]
+    val areaRatio = Core.sumElems(maskFloat).`val`[0] / (probmap.rows() * probmap.cols())
+    return if (areaRatio < minQuadAreaRatio) 0.0 else meanProb * (0.7 + 0.3 * areaRatio)
+}

imageprocessing/src/main/java/org/fairscan/imageprocessing/quad/RightAngles.kt

@@ -1,133 +0,0 @@
-/*
- * Copyright 2025-2026 Pierre-Yves Nicolas
- *
- * This program is free software: you can redistribute it and/or modify it
- * under the terms of the GNU General Public License as published by the Free
- * Software Foundation, either version 3 of the License, or (at your option)
- * any later version.
- * This program is distributed in the hope that it will be useful, but WITHOUT
- * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
- * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
- * more details.
- * You should have received a copy of the GNU General Public License along with
- * this program. If not, see <https://www.gnu.org/licenses/>.
- */
-package org.fairscan.imageprocessing.quad
-
-import org.fairscan.imageprocessing.Point
-import kotlin.math.abs
-import kotlin.math.acos
-import kotlin.math.sqrt
-import kotlin.math.sign
-
-// Look for 3 consecutive angles that are (almost) right angles
-fun findQuadFromRightAngles(
-    points: List<Point>,
-    imgWidth: Int,
-    imgHeight: Int,
-    angleMin: Float = 60f,
-    angleMax: Float = 120f
-): List<Point>? {
-    if (points.size < 4) return null
-    val n = points.size
-
-    val angles = mutableListOf<Double>()
-    for (i in 0 until n) {
-        val a = points[(i + n - 1) % n]
-        val b = points[i]
-        val c = points[(i + 1) % n]
-        angles.add(orientedAngle(a, b, c))
-    }
-
-    var bestQuad: List<Point>? = null
-    var bestScore = Double.POSITIVE_INFINITY
-
-    for (i in 0 until n) {
-        val triplet = listOf(angles[i % n], angles[(i + 1) % n], angles[(i + 2) % n])
-        if (triplet.all { it in angleMin..angleMax }) {
-            val a = points[(i + n - 1) % n]
-            val b = points[i]
-            val c = points[(i + 1) % n]
-            val d = points[(i + 2) % n]
-            val e = points[(i + 3) % n]
-
-            val inter = lineIntersection2(a, b, d, e) ?: continue
-
-            val quad = listOf(b, c, d, inter)
-
-            // ensure inside image bounds
-            if (quad.any { it.x < 0 || it.x >= imgWidth || it.y < 0 || it.y >= imgHeight }) continue
-
-            // ensure convex
-            if (!isConvex(quad)) continue
-
-            val score = quadAngleError(quad)
-            if (score < bestScore) {
-                bestScore = score
-                bestQuad = quad
-            }
-        }
-    }
-    return bestQuad
-}
-
-fun angleBetween(v1: Point, v2: Point): Float {
-    val norm1 = sqrt(v1.x * v1.x + v1.y * v1.y) + 1e-9f
-    val norm2 = sqrt(v2.x * v2.x + v2.y * v2.y) + 1e-9f
-    val dot = (v1.x * v2.x + v1.y * v2.y) / (norm1 * norm2)
-    val cosAngle = dot.coerceIn(-1.0, 1.0)
-    return Math.toDegrees(acos(cosAngle).toDouble()).toFloat()
-}
-
-fun orientedAngle(a: Point, b: Point, c: Point): Double {
-    val v1 = Point(a.x - b.x, a.y - b.y)
-    val v2 = Point(c.x - b.x, c.y - b.y)
-    val norm1 = sqrt(v1.x * v1.x + v1.y * v1.y) + 1e-9f
-    val norm2 = sqrt(v2.x * v2.x + v2.y * v2.y) + 1e-9f
-    val dot = ((v1.x * v2.x + v1.y * v2.y) / (norm1 * norm2)).coerceIn(-1.0, 1.0)
-    val cross = v1.x * v2.y - v1.y * v2.x
-    var angle = Math.toDegrees(acos(dot))
-    if (cross < 0) angle = 360.0 - angle
-    return angle
-}
-
-fun lineIntersection2(p1: Point, p2: Point, p3: Point, p4: Point): Point? {
-    val denom = (p1.x - p2.x) * (p3.y - p4.y) - (p1.y - p2.y) * (p3.x - p4.x)
-    if (abs(denom) < 1e-6f) return null
-    val numX = (p1.x * p2.y - p1.y * p2.x)
-    val numY = (p3.x * p4.y - p3.y * p4.x)
-    val px = (numX * (p3.x - p4.x) - (p1.x - p2.x) * numY) / denom
-    val py = (numX * (p3.y - p4.y) - (p1.y - p2.y) * numY) / denom
-    return Point(px, py)
-}
-
-fun quadAngleError(quad: List<Point>): Double {
-    var err = 0.0
-    for (i in 0 until 4) {
-        val a = quad[(i + 3) % 4]
-        val b = quad[i]
-        val c = quad[(i + 1) % 4]
-        val ang = angleBetween(Point(a.x - b.x, a.y - b.y), Point(c.x - b.x, c.y - b.y))
-        err += abs(ang - 90.0)
-    }
-    return err
-}
-
-fun isConvex(quad: List<Point>): Boolean {
-    if (quad.size != 4) return false
-    var sign = 0
-    for (i in quad.indices) {
-        val a = quad[i]
-        val b = quad[(i + 1) % 4]
-        val c = quad[(i + 2) % 4]
-        val cross = (b.x - a.x) * (c.y - b.y) - (b.y - a.y) * (c.x - b.x)
-        val currentSign = cross.sign.toInt()
-        if (sign == 0 && currentSign != 0) {
-            sign = currentSign
-        } else if (currentSign != 0 && currentSign != sign) {
-            return false
-        }
-    }
-    return true
-}
-