@jaak.ai/stamps 2.0.0-dev.47 → 2.0.0-dev.48

package/dist/esm/jaak-stamps.entry.js

@@ -728,7 +728,7 @@ class DetectionService {
     modelLoaded = false;
     deviceStrategy;
     MODEL_PATH = "https://storage.googleapis.com/jaak-static/web/component/stamps/ddmyp-v2.onnx";
-    MOBILENET_MODEL_PATH = "https://storage.googleapis.com/jaak-static/web/component/stamps/cdmmp-v1.onnx";
+    MOBILENET_MODEL_PATH = "https://storage.googleapis.com/jaak-static/web/component/stamps/squeezenet.onnx";
     MOBILENET_CLASSES_PATH = "https://storage.googleapis.com/jaak-static/web/component/stamps/cdmmp-v1.json";
     INPUT_SIZE = 320;
     CONFIDENCE_THRESHOLD = 0.6;
@@ -790,15 +790,27 @@ class DetectionService {
         }
         try {
             this.logger.state('CARGANDO_MODELO_MOBILENET', { path: this.MOBILENET_MODEL_PATH });
-            // Load class map
-            const classResponse = await fetch(this.MOBILENET_CLASSES_PATH);
-            if (!classResponse.ok) {
-                throw new Error(`Failed to load class map: ${this.MOBILENET_CLASSES_PATH}`);
+            // Try to load class map (optional - SqueezeNet may not need it for basic classification)
+            try {
+                const classResponse = await fetch(this.MOBILENET_CLASSES_PATH);
+                if (classResponse.ok) {
+                    this.mobileNetClassMap = await classResponse.json();
+                    this.logger.state('CLASES_MOBILENET_CARGADAS', {
+                        classCount: Object.keys(this.mobileNetClassMap).length
+                    });
+                }
+            }
+            catch (error) {
+                this.logger.warn('No se pudo cargar el mapa de clases de MobileNet, usando clasificación básica');
+                // Create a basic class map for document types
+                this.mobileNetClassMap = {
+                    "0": "document",
+                    "1": "id_card",
+                    "2": "passport",
+                    "3": "license",
+                    "4": "other"
+                };
             }
-            this.mobileNetClassMap = await classResponse.json();
-            this.logger.state('CLASES_MOBILENET_CARGADAS', {
-                classCount: Object.keys(this.mobileNetClassMap).length
-            });
             // Load model
             const sessionOptions = this.deviceStrategy.getSessionOptions(this.debug);
             try {
@@ -887,7 +899,7 @@ class DetectionService {
         try {
             this.logger.state('CLASIFICANDO_DOCUMENTO', { timestamp: Date.now() });
             const inputTensor = this.preprocessMobileNet(canvas);
-            const feeds = { input: inputTensor };
+            const feeds = { [this.mobileNetSession.inputNames[0]]: inputTensor };
             const results = await this.mobileNetSession.run(feeds);
             const output = results[Object.keys(results)[0]].data;
             const maxIdx = output.reduce((bestIdx, val, idx, arr) => val > arr[bestIdx] ? idx : bestIdx, 0);
@@ -897,7 +909,8 @@ class DetectionService {
                 class: className,
                 confidence: confidence.toFixed(3),
                 classIndex: maxIdx,
-                timestamp: Date.now()
+                timestamp: Date.now(),
+                results
             });
             return {
                 class: className,
@@ -936,7 +949,7 @@ class DetectionService {
             displayedVideoHeight = INPUT_SIZE / videoAspectRatio;
         }
         else {
-            // Video is taller - pillarboxed in display  
+            // Video is taller - pillarboxed in display
             displayedVideoHeight = INPUT_SIZE;
             displayedVideoWidth = INPUT_SIZE * videoAspectRatio;
         }
@@ -987,7 +1000,7 @@ class DetectionService {
         const leftAligned = Math.abs(docLeft - maskLeft) <= tolerance;
         return {
             top: topAligned && leftAligned, // Esquina superior izquierda: borde top Y left alineados
-            right: topAligned && rightAligned, // Esquina superior derecha: borde top Y right alineados  
+            right: topAligned && rightAligned, // Esquina superior derecha: borde top Y right alineados
             bottom: bottomAligned && leftAligned, // Esquina inferior izquierda: borde bottom Y left alineados
             left: bottomAligned && rightAligned // Esquina inferior derecha: borde bottom Y right alineados
         };
@@ -1072,15 +1085,20 @@ class DetectionService {
         tempCanvas.width = 224;
         tempCanvas.height = 224;
         const tempCtx = tempCanvas.getContext('2d');
+        // Resize image to 224x224 for SqueezeNet (using MobileNet interface)
         tempCtx.drawImage(canvas, 0, 0, 224, 224);
         const imageData = tempCtx.getImageData(0, 0, 224, 224);
         const data = imageData.data;
         const hw = 224 * 224;
         const arr = new Float32Array(3 * hw);
+        // SqueezeNet preprocessing: normalize to [0,1] range and arrange in CHW format
         for (let i = 0; i < hw; i++) {
-            arr[i] = (data[i * 4 + 0] / 255 - 0.5) / 0.5;
-            arr[hw + i] = (data[i * 4 + 1] / 255 - 0.5) / 0.5;
-            arr[2 * hw + i] = (data[i * 4 + 2] / 255 - 0.5) / 0.5;
+            // Red channel
+            arr[i] = data[i * 4 + 0] / 255.0;
+            // Green channel
+            arr[hw + i] = data[i * 4 + 1] / 255.0;
+            // Blue channel
+            arr[2 * hw + i] = data[i * 4 + 2] / 255.0;
         }
         return new window.ort.Tensor('float32', arr, [1, 3, 224, 224]);
     }