Add demo scene button, thread config, user expansion

app.py

@@ -36,6 +36,7 @@ from module.qdrant_manager import (
 )
 from module.qdrant_manager import qdrant_search_similar, pt_id
 from module.text_classifier import classify_scene
+from module.demo_images import ensure_demo_images
 
 BASE_DIR = os.path.dirname(os.path.abspath(__file__))
 FRONTEND_PATH = os.path.join(BASE_DIR, 'index.html')
@@ -105,6 +106,9 @@ async def lifespan(app):
     total_startup = time.perf_counter() - startup_t0
     print(f'[启动] 总启动时间: {total_startup:.1f}s')
 
+    # 生成示例场景图片（用于 /api/demo 端点）
+    ensure_demo_images(BASE_DIR)
+
     yield
 
     session_data['shutdown_stats'] = _snapshot()
@@ -684,6 +688,177 @@ async def status():
     }
 
 
+@app.post("/api/demo")
+async def demo(scene_type: str = Form("auto")):
+    """示例场景检测：自动读取 test_imgs 中的示例图片执行检测
+
+    Args:
+        scene_type: 'text' / 'complex' / 'auto'（auto 表示自动选择）
+    """
+    import shutil
+    demo_map = ensure_demo_images(BASE_DIR)
+    if not demo_map:
+        return JSONResponse({"error": "示例图片生成失败"}, status_code=500)
+
+    # 选择场景类型
+    if scene_type == 'auto':
+        scene_type = 'text'
+    if scene_type not in demo_map:
+        return JSONResponse({"error": f"不支持的场景类型: {scene_type}"}, status_code=400)
+
+    img_paths = demo_map[scene_type]
+    if len(img_paths) < 2:
+        return JSONResponse({"error": "示例图片不足"}, status_code=500)
+
+    # 复制到临时目录执行检测（不污染任何用户的历史数据）
+    import tempfile
+    tmp_dir = tempfile.mkdtemp(prefix='demo_')
+    tmp_imgs = []
+    for i, p in enumerate(img_paths[:2]):
+        dst = os.path.join(tmp_dir, f"demo_{i}.jpg")
+        shutil.copy(p, dst)
+        tmp_imgs.append(dst)
+
+    try:
+        loop = asyncio.get_event_loop()
+        result = await loop.run_in_executor(
+            executor,
+            _run_demo_pipeline,
+            tmp_imgs, scene_type
+        )
+        return result
+    except Exception as e:
+        import traceback
+        traceback.print_exc()
+        return JSONResponse({"error": f"示例检测失败: {str(e)}"}, status_code=500)
+    finally:
+        # 清理临时目录
+        try:
+            shutil.rmtree(tmp_dir, ignore_errors=True)
+        except Exception:
+            pass
+
+
+def _run_demo_pipeline(img_paths, expected_scene):
+    """执行示例检测流程（不涉及用户历史数据，纯两张图比对）"""
+    import cv2
+    import numpy as np
+
+    path1, path2 = img_paths[0], img_paths[1]
+
+    # 场景分类（用第一张图）
+    ocr_result = None
+    full_text = ''
+    try:
+        scene_type, detail = classify_scene(path1, ocr_engine)
+        ocr_result = detail.get('ocr_result')
+        full_text = detail.get('full_text', '')
+    except Exception as e:
+        scene_type = expected_scene
+        print(f"[demo] 场景分类失败: {e}")
+
+    is_same = False
+    similarity = 0.0
+    log_entry = {
+        'query_path': path1,
+        'history_path': path2,
+        'scene_type': scene_type,
+        'is_same_scene': False,
+        'query_image': os.path.basename(path1),
+        'history_image': os.path.basename(path2),
+        'query_text': full_text,
+    }
+    visualization_data = []
+
+    if scene_type == 'text':
+        # 文本场景：OCR + BGE 比对
+        from module.text_matcher import compute_text_similarity
+        try:
+            # OCR 第二张图
+            scene2, detail2 = classify_scene(path2, ocr_engine)
+            text2 = detail2.get('full_text', '')
+            log_entry['history_text'] = text2
+
+            # BGE 编码比对
+            text_sim = compute_text_similarity(full_text, text2, bge_tokenizer, bge_model)
+            similarity = text_sim
+            is_same = text_sim >= 0.85
+            log_entry['text_similarity'] = text_sim
+            log_entry['is_same_scene'] = is_same
+        except Exception as e:
+            print(f"[demo] 文本场景比对失败: {e}")
+
+    else:
+        # 复杂场景：DINOv2 比对
+        try:
+            if dinov2_extractor is not None:
+                dino_sim = dinov2_extractor.compute_similarity(path1, path2)
+                similarity = dino_sim if dino_sim else 0.0
+                is_same = similarity >= 0.5
+                log_entry['dinov2_similarity'] = round(similarity, 4)
+                log_entry['is_same_scene'] = is_same
+
+                # 计算 patch 匹配（用于可视化）
+                if is_same:
+                    patch_info = dinov2_extractor.compute_patch_matches(path1, path2, top_k=50)
+                    log_entry['patch_match_info'] = patch_info
+        except Exception as e:
+            print(f"[demo] 复杂场景比对失败: {e}")
+
+    # 构建可视化（两张图都还在，直接构建）
+    _image_cache = {}
+    for p in [path1, path2]:
+        try:
+            data = np.fromfile(p, dtype=np.uint8)
+            img = cv2.imdecode(data, cv2.IMREAD_COLOR)
+            if img is not None:
+                _image_cache[p] = img
+        except Exception:
+            pass
+
+    try:
+        if scene_type == 'text':
+            vis = _build_text_vis_from_cache(log_entry, _image_cache)
+            if vis:
+                visualization_data.append(vis)
+        else:
+            vis = _build_complex_vis_from_cache(log_entry, _image_cache)
+            if vis:
+                visualization_data.append(vis)
+    except Exception as e:
+        print(f"[demo] 可视化生成失败: {e}")
+
+    # 将原图也转为 base64 返回（前端无需再请求 /api/image）
+    def _img_to_b64(path):
+        try:
+            data = np.fromfile(path, dtype=np.uint8)
+            img = cv2.imdecode(data, cv2.IMREAD_COLOR)
+            if img is None:
+                return None
+            _, buf = cv2.imencode('.jpg', img, [cv2.IMWRITE_JPEG_QUALITY, 85])
+            return f'data:image/jpeg;base64,{base64.b64encode(buf).decode("utf-8")}'
+        except Exception:
+            return None
+
+    return {
+        "scene_type": scene_type,
+        "is_same_scene": is_same,
+        "similarity": round(similarity, 4),
+        "query_image": os.path.basename(path1),
+        "history_image": os.path.basename(path2),
+        "query_image_b64": _img_to_b64(path1),
+        "history_image_b64": _img_to_b64(path2),
+        "visualizations": visualization_data,
+        "evaluation_logs": [{
+            "query_image": os.path.basename(path1),
+            "history_image": os.path.basename(path2),
+            "scene_type": scene_type,
+            "is_same_scene": is_same,
+            "similarity": round(similarity, 4),
+        }],
+    }
+
+
 if __name__ == "__main__":
     import uvicorn
     port = int(os.environ.get("PORT", 7860))

index.html

@@ -99,6 +99,13 @@ tr.row-fail td{color:#721c24}
     <input type="text" id="usernameInput" placeholder="请输入用户名" maxlength="50">
     <div id="loginError" class="error-msg"></div>
     <button class="btn btn-primary" style="width:100%;margin-top:8px" onclick="doLogin()">确认登录</button>
+    <div style="margin-top:16px;border-top:1px solid #eee;padding-top:14px">
+      <p style="font-size:12px;color:#999;margin-bottom:8px">无需登录，一键体验示例场景：</p>
+      <div style="display:flex;gap:8px">
+        <button class="btn btn-warning" style="flex:1" onclick="runDemo('text')">📄 文本场景示例</button>
+        <button class="btn btn-warning" style="flex:1" onclick="runDemo('complex')">🏞️ 复杂场景示例</button>
+      </div>
+    </div>
   </div>
 </div>
 
@@ -114,6 +121,8 @@ tr.row-fail td{color:#721c24}
       <span>📁 历史图片: <b id="dispCount">0</b> 张</span>
     </div>
     <div class="status-actions">
+      <button class="btn btn-warning btn-sm" onclick="runDemo('text')">📄 文本示例</button>
+      <button class="btn btn-warning btn-sm" onclick="runDemo('complex')">🏞️ 复杂示例</button>
       <button class="btn btn-success btn-sm" onclick="openHistory()">📁 历史图片</button>
       <button class="btn btn-info btn-sm" onclick="switchUser()">🔄 更换用户</button>
     </div>
@@ -432,6 +441,90 @@ function hideLoading() {
 document.getElementById('historyPanel').addEventListener('click', function(e) {
   if (e.target === this) closeHistory();
 });
+
+// ========== 示例场景演示 ==========
+async function runDemo(sceneType) {
+  showLoading('正在运行示例场景检测，请稍候...');
+  // 清空之前的结果
+  const resultArea = document.getElementById('resultArea');
+  if (resultArea) resultArea.innerHTML = '';
+
+  try {
+    const fd = new FormData();
+    fd.append('scene_type', sceneType);
+    const res = await fetch('/api/demo', { method: 'POST', body: fd });
+    const data = await res.json();
+    if (data.error) {
+      alert('示例检测失败: ' + data.error);
+      hideLoading();
+      return;
+    }
+    renderDemoResult(data, sceneType);
+  } catch (e) {
+    alert('网络错误: ' + e.message);
+  } finally {
+    hideLoading();
+  }
+}
+
+function renderDemoResult(data, sceneType) {
+  // 如果在登录页，切到主页视图展示结果
+  document.getElementById('loginView').classList.add('hidden');
+  const mainView = document.getElementById('mainView');
+  mainView.classList.remove('hidden');
+  document.getElementById('dispUser').textContent = '示例演示';
+  document.getElementById('dispCount').textContent = '0';
+
+  const resultArea = document.getElementById('resultArea');
+  const isSame = data.is_same_scene;
+  const sceneLabel = data.scene_type === 'text' ? '文本场景' : '复杂场景';
+  const simPercent = (data.similarity * 100).toFixed(1);
+
+  let html = '';
+  html += `<div class="card"><h3>🎯 示例场景检测结果（${sceneLabel}）</h3>`;
+  html += `<div style="margin:12px 0;padding:16px;border-radius:8px;background:${isSame ? '#d4edda' : '#f8d7da'};border:2px solid ${isSame ? '#28a745' : '#dc3545'};text-align:center">`;
+  html += `<h2 style="color:${isSame ? '#155724' : '#721c24'};font-size:20px;margin-bottom:4px">${isSame ? '✅ 同一场景' : '❌ 不同场景'}</h2>`;
+  html += `<p style="color:${isSame ? '#155724' : '#721c24'};font-size:14px">相似度: ${simPercent}% | 场景类型: ${sceneLabel}</p>`;
+  html += `</div>`;
+
+  // 图片对比
+  html += `<div style="display:flex;gap:12px;margin:12px 0">`;
+  html += `<div style="flex:1;text-align:center"><p style="font-size:12px;color:#666;margin-bottom:4px">Query 图</p>${data.query_image_b64 ? `<img src="${data.query_image_b64}" style="max-width:100%;border-radius:6px;border:1px solid #eee">` : '<span style="color:#999">无</span>'}</div>`;
+  html += `<div style="flex:1;text-align:center"><p style="font-size:12px;color:#666;margin-bottom:4px">History 图</p>${data.history_image_b64 ? `<img src="${data.history_image_b64}" style="max-width:100%;border-radius:6px;border:1px solid #eee">` : '<span style="color:#999">无</span>'}</div>`;
+  html += `</div>`;
+
+  // 可视化
+  if (data.visualizations && data.visualizations.length > 0) {
+    for (const vis of data.visualizations) {
+      const visType = vis.scene_type === 'text' ? '文本场景' : '复杂场景';
+      html += `<div style="margin-top:12px">`;
+      html += `<div class="vis-toggle" onclick="this.nextElementSibling.classList.toggle('collapsed');this.nextElementSibling.classList.toggle('expanded')"><span>▶</span> 🖼️ 点击查看可视化证据（${visType}）</div>`;
+      html += `<div class="vis-content collapsed">`;
+      html += `<img src="${vis.image_base64}" style="width:100%;border-radius:6px;border:1px solid #eee">`;
+      if (vis.scene_type === 'complex') {
+        html += `<div class="vis-stats">`;
+        html += `<span>DINOv2 相似度: <b>${(vis.dinov2_similarity * 100).toFixed(1)}%</b></span>`;
+        html += `<span>Patch 匹配数: <b>${vis.patch_match_count}</b></span>`;
+        html += `<span>平均相似度: <b>${(vis.avg_patch_similarity * 100).toFixed(1)}%</b></span>`;
+        html += `</div>`;
+      } else {
+        html += `<div class="vis-stats">`;
+        html += `<span>BGE 语义相似度: <b>${(vis.text_similarity * 100).toFixed(1)}%</b></span>`;
+        html += `<span>共同关键词: <b>${vis.common_keyword_count}</b> 个</span>`;
+        if (vis.common_keywords && vis.common_keywords.length > 0) {
+          html += `<div style="margin-top:4px;font-size:12px;color:#666">关键词: ${vis.common_keywords.join('、')}</div>`;
+        }
+        html += `</div>`;
+      }
+      html += `</div></div>`;
+    }
+  }
+
+  html += `<div style="margin-top:16px;text-align:center;font-size:12px;color:#999">这是示例演示结果。如需检测您自己的图片，请先登录。</div>`;
+  html += `</div>`;
+
+  resultArea.innerHTML = html;
+}
 </script>
 </body>
 </html>

module/config.py

@@ -9,6 +9,9 @@ DEVICE = "cpu"
 USE_HALF = False
 IMAGE_SIZE = 512
 
+# 匹配 HF Spaces 免费 2vCPU 配置，显式设置线程数避免过度并发开销
+torch.set_num_threads(2)
+
 # MASt3R 已禁用（CPU 部署，耗时过长）
 MAST3R_ROOT = None
 LOCAL_WEIGHTS = None

module/demo_images.py

@@ -0,0 +1,135 @@
+"""示例场景图片生成 — 在启动时自动生成演示用图片
+
+生成两组示例：
+- 文本场景：模拟合同文档（白底 + 文字行 + 标题），两张内容相同但亮度/偏移不同
+- 复杂场景：模拟实景照片（彩色渐变 + 几何形状），两张构图相似但有差异
+"""
+import os
+from PIL import Image, ImageDraw, ImageFont, ImageFilter
+
+
+def _get_font(size):
+    try:
+        return ImageFont.truetype("/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf", size)
+    except Exception:
+        try:
+            return ImageFont.truetype("DejaVuSans.ttf", size)
+        except Exception:
+            return ImageFont.load_default()
+
+
+def _generate_text_scene(path1, path2):
+    """生成文本场景示例：模拟合同文档
+
+    两张图内容相同，但图2做了亮度调整和水平偏移，模拟不同角度拍摄
+    """
+    w, h = 600, 800
+
+    for path, brightness, offset_x in [(path1, 0, 0), (path2, 30, 5)]:
+        img = Image.new('RGB', (w, h), (255, 255, 255))
+        draw = ImageDraw.Draw(img)
+
+        # 标题区域
+        draw.rectangle([50 + offset_x, 40, 550 + offset_x, 80], fill=(30, 30, 30))
+        # 正文行
+        for i in range(25):
+            y = 120 + i * 25
+            x_start = 50 + offset_x
+            x_end = 550 - (i % 5) * 20 + offset_x
+            draw.rectangle([x_start, y, x_end, y + 12], fill=(60, 60, 60))
+
+        # 模拟印章（红色圆形）
+        draw.ellipse([400 + offset_x, 650, 500 + offset_x, 750],
+                     outline=(180, 30, 30), width=3)
+        draw.text((420 + offset_x, 680), "SEAL", fill=(180, 30, 30),
+                  font=_get_font(20))
+
+        # 亮度调整
+        if brightness > 0:
+            import numpy as np
+            arr = np.array(img).astype(int) + brightness
+            arr = np.clip(arr, 0, 255).astype('uint8')
+            img = Image.fromarray(arr)
+
+        img.save(path, 'JPEG', quality=85)
+
+
+def _generate_complex_scene(path1, path2):
+    """生成复杂场景示例：模拟实景照片
+
+    两张图构图相似（同一场景不同角度），使用彩色渐变和几何形状
+    """
+    w, h = 600, 450
+
+    for path, shift_x, color_shift in [(path1, 0, 0), (path2, 15, 20)]:
+        img = Image.new('RGB', (w, h))
+        draw = ImageDraw.Draw(img)
+
+        # 天空渐变
+        for y in range(h // 2):
+            r = int(100 + y * 0.3 + color_shift)
+            g = int(150 + y * 0.2)
+            b = int(200 + y * 0.15)
+            draw.line([(0, y), (w, y)], fill=(min(r, 255), min(g, 255), min(b, 255)))
+
+        # 地面渐变
+        for y in range(h // 2, h):
+            r = int(80 + (y - h // 2) * 0.3)
+            g = int(120 + (y - h // 2) * 0.2)
+            b = int(60 + (y - h // 2) * 0.1)
+            draw.line([(0, y), (w, y)], fill=(min(r, 255), min(g, 255), min(b, 255)))
+
+        # 太阳
+        sun_x = 450 + shift_x
+        draw.ellipse([sun_x - 40, 50, sun_x + 40, 130],
+                     fill=(255, 220, 100))
+
+        # 山脉
+        draw.polygon([(0, 225), (150 + shift_x, 150), (300 + shift_x, 200),
+                      (450 + shift_x, 160), (600, 220), (600, 225)],
+                     fill=(80, 100, 80))
+
+        # 建筑
+        draw.rectangle([100 + shift_x, 250, 200 + shift_x, 400],
+                       fill=(120, 100, 90))
+        draw.rectangle([250 + shift_x, 280, 350 + shift_x, 400],
+                       fill=(100, 110, 100))
+        draw.rectangle([400 + shift_x, 260, 480 + shift_x, 400],
+                       fill=(90, 80, 100))
+
+        # 轻微模糊模拟不同焦距
+        if shift_x > 0:
+            img = img.filter(ImageFilter.GaussianBlur(radius=0.8))
+
+        img.save(path, 'JPEG', quality=85)
+
+
+def ensure_demo_images(base_dir):
+    """确保示例图片存在，不存在则生成
+
+    Args:
+        base_dir: 项目根目录
+    Returns:
+        dict: {'text': [path1, path2], 'complex': [path1, path2]}
+    """
+    demo_dir = os.path.join(base_dir, 'test_imgs')
+    os.makedirs(demo_dir, exist_ok=True)
+
+    text1 = os.path.join(demo_dir, 'demo_text_1.jpg')
+    text2 = os.path.join(demo_dir, 'demo_text_2.jpg')
+    complex1 = os.path.join(demo_dir, 'demo_complex_1.jpg')
+    complex2 = os.path.join(demo_dir, 'demo_complex_2.jpg')
+
+    if not os.path.exists(text1):
+        try:
+            _generate_text_scene(text1, text2)
+            _generate_complex_scene(complex1, complex2)
+            print("[示例] 已生成演示图片")
+        except Exception as e:
+            print(f"[示例] 生成失败: {e}")
+            return None
+
+    return {
+        'text': [text1, text2],
+        'complex': [complex1, complex2],
+    }