Update index.html

index.html

@@ -338,13 +338,12 @@
         </div>
     </div>
 </div>
-
-  <script>
-    // モデル定義（各モデルの入力サイズを設定）
+<script>
+    // モデル定義（JSONの閾値 threshold: 0.19671 を追加）
     const models = [
-        { file: "tinymodelV9s.onnx", type: "hsc", size: 256 },  // inf.py 仕様
-        { file: "tinymodelV3.onnx",  type: "hsc", size: 224 },  // 異常検知 (224x224)
-        { file: "tinymodelV2.onnx",  type: "softmax", size: 224 } // 分類 (224x224)
+        { file: "tinymodelV9s.onnx", type: "hsc", size: 256, threshold: 0.3 }, // 最適化閾値を適用
+        { file: "tinymodelV3.onnx",  type: "hsc", size: 224, threshold: 0.5 },
+        { file: "tinymodelV2.onnx",  type: "softmax", size: 224, threshold: 0.5 }
     ];
 
     let session = null;
@@ -384,7 +383,7 @@
             }
             session = await ort.InferenceSession.create('./' + modelConfig.file);
             isReady = true;
-            statusText.innerText = "✅ " + modelConfig.file + " の準備完了";
+            statusText.innerText = "✅ " + modelConfig.file + " の準備完了 (閾値: " + (modelConfig.threshold * 100).toFixed(1) + "%)";
             
             if (previewImg.src && previewImg.style.display !== 'none') {
                 submitBtn.disabled = false;
@@ -410,7 +409,6 @@
 
     fileInput.addEventListener('change', (e) => handleFile(e.target.files[0]));
 
-    // モデルごとの targetSize に合わせてリサイズと標準化処理を行う
     async function preprocess(imgElement, targetSize) {
         const canvas = document.createElement('canvas');
         canvas.width = targetSize;
@@ -419,7 +417,6 @@
         ctx.drawImage(imgElement, 0, 0, targetSize, targetSize);
         const data = ctx.getImageData(0, 0, targetSize, targetSize).data;
         
-        // inf.py と同じ mean / std
         const mean = [0.485, 0.456, 0.406];
         const std = [0.229, 0.224, 0.225];
         
@@ -445,10 +442,11 @@
             const origW = previewImg.naturalWidth;
             const origH = previewImg.naturalHeight;
             
-            let probNeed, probTrash;
+            let probNeed, probTrash, rawTrashScore;
 
-            // inf.py の仕様: 元画像が 256px 以下の場合は推論をスキップして即 trash 判定
+            // 256px 以下の低解像度画像の除外判定 (inf.py 仕様)
             if (currentModelConfig.file === "tinymodelV9s.onnx" && (origW <= 256 || origH <= 256)) {
+                rawTrashScore = 1.0;
                 probTrash = 100.0;
                 probNeed = 0.0;
                 statusText.innerText = "⚠️ 低解像度のため判定をスキップ (Trash)";
@@ -464,26 +462,28 @@
                 const output = results[session.outputNames[0]].data;
 
                 if (currentModelConfig.type === "hsc") {
-                    // 1次元出力 (Trashスコア)
-                    const trashScore = output[0];
-                    probTrash = trashScore * 100;
-                    probNeed = (1.0 - trashScore) * 100;
+                    rawTrashScore = output[0];
+                    probTrash = rawTrashScore * 100;
+                    probNeed = (1.0 - rawTrashScore) * 100;
                 } else if (currentModelConfig.type === "softmax") {
-                    // 2次元出力 Logits
                     const exp0 = Math.exp(output[0]);
                     const exp1 = Math.exp(output[1]);
                     const sum = exp0 + exp1;
                     probNeed = (exp0 / sum) * 100;
                     probTrash = (exp1 / sum) * 100;
+                    rawTrashScore = probTrash / 100.0;
                 }
                 statusText.innerText = "✅ 完了 (" + duration + "ms)";
             }
 
+            // 判定閾値 (threshold) を使用した判定ロジック
+            const threshold = currentModelConfig.threshold ?? 0.5;
+            const topClass = rawTrashScore > threshold ? "trash" : "need";
+            const topConf = (topClass === "trash" ? probTrash : probNeed).toFixed(1);
+
             // UI描画
             emptyOutput.style.display = 'none';
             resultOutput.style.display = 'flex';
-            const topClass = probNeed >= probTrash ? "need" : "trash";
-            const topConf = Math.max(probNeed, probTrash).toFixed(1);
             
             document.getElementById('topClassName').innerText = topClass;
             document.getElementById('topClassConf').innerText = "信頼度: " + topConf + "%";