feat: 2.18.10 靶规格复检与置信度门槛

- 检出靶框裁剪放大后二次 detect 复检(conf 0.25)，以复检结果为准 - 全图结果上报门槛 MIN_REPORT_CONF=0.5，复检结果 >0.25 即采纳 - 0 结果时按 RETRY_CONF_TH 自动重检；class_id=1 解析阶段过滤 - 换用 model_326162，去除调试 print 改走 logger
2026-10-23 15:13:48 +08:00
parent d1a1d43bd8
commit 8d35bf3068
6 changed files with 154 additions and 39 deletions
@@ -1,6 +1,6 @@
 id: t11
 name: t11
-version: 2.18.9
+version: 2.18.10
 author: t11
 icon: ''
 desc: t11
@@ -18,8 +18,8 @@ files:
  - laser_manager.py
  - logger_manager.py
  - main.py
-  - model_325964.cvimodel
-  - model_325964.mud
+  - model_326162.cvimodel
+  - model_326162.mud
  - network.py
  - ota_curl.sh
  - ota_manager.py
@@ -264,7 +264,7 @@ TRIANGLE_YOLO_PRELOAD_ON_BOOT = False

 # YOLO target size classification: class_id → label，下标即 class_id：0=40cm, 1=0(靶心), 2=20cm.
 TARGET_CLASS_YOLO_ENABLE = True
-TARGET_CLASS_YOLO_MODEL_PATH = APP_DIR + "/model_325964.mud"
+TARGET_CLASS_YOLO_MODEL_PATH = APP_DIR + "/model_326162.mud"
 TARGET_CLASS_YOLO_LABELS = (40, 0, 20)
 # 只有映射结果落在该集合内才用于上报（如 class 1→label 0 直接忽略）
 TARGET_CLASS_YOLO_VALID_LABELS = (20, 40)
@@ -273,6 +273,13 @@ TARGET_CLASS_YOLO_IGNORE_CLASS_IDS = (1,)
 TARGET_CLASS_YOLO_CONF_TH = 0.5
 # 首次检测 0 结果时用该阈值重新检测一次（≥CONF_TH 时按 CONF_TH 同参重跑）
 TARGET_CLASS_YOLO_RETRY_CONF_TH = 0.25
+# 检出靶框后裁剪放大再 detect 一次，与全图结果互相印证（不一致以复检为准）
+TARGET_CLASS_YOLO_RECHECK_CROP = True
+TARGET_CLASS_YOLO_RECHECK_MARGIN_FRAC = 0.1
+# 复检的 detect 阈值与接受门槛：置信度大于该值即采纳复检结果
+TARGET_CLASS_YOLO_RECHECK_CONF_TH = 0.25
+# 全图结果（无复检时）的上报置信度门槛，None/0 表示不启用
+TARGET_CLASS_YOLO_MIN_REPORT_CONF = 0.5
 TARGET_CLASS_YOLO_IOU_TH = 0.45
 TARGET_CLASS_YOLO_PRELOAD_ON_BOOT = True

@@ -1,7 +1,7 @@

 [basic]
 type = cvimodel
-model = model_325964.cvimodel
+model = model_326162.cvimodel

 [extra]
 model_type = yolov5
@@ -312,7 +312,8 @@ def _det_obj_score(o):
 def try_get_target_class_from_yolo(maix_frame, logger=None):
    """调用 YOLO 检测：class_id∈IGNORE_CLASS_IDS(默认 1) 的框在解析时直接过滤；
    detections 输出其余全部框（不按置信度过滤）；
-    上报字段 label/class_id/confidence 仅在有效靶规格(20/40)中取置信度最高的一条。"""
+    上报以「检出框裁剪放大后的复检结果」为主，复检不可用时用全图结果；
+    有效靶规格(20/40)中取置信度最高的一条，且低于 MIN_REPORT_CONF 不上报类别。"""
    try:
        import config as cfg
    except Exception:
@@ -365,7 +366,6 @@ def try_get_target_class_from_yolo(maix_frame, logger=None):
        try:
            return detector.detect(maix_frame, conf_th=threshold, iou_th=iou_th)
        except Exception as exc:
-            print(f"[YOLO-TARGET] detect 异常: {exc}")
            if logger:
                logger.warning(f"[YOLO-TARGET] detect 异常: {exc}")
            return None
@@ -408,57 +408,165 @@ def try_get_target_class_from_yolo(maix_frame, logger=None):
    # 第 1 次检测
    raw = _run(conf_th)
    detections = _parse(raw)
-    print(f"[YOLO-TARGET] 第1次 conf_th={conf_th} iou_th={iou_th} raw={raw!r}")
-    print(
-        f"[YOLO-TARGET] 检测结果={detections}"
-        f"（已忽略 class_id∈{tuple(sorted(ignore_class_ids))} 共{n_ignored[0]} 条）"
-    )

    # 0 结果时重新检测一次（阈值取更小值，仍为 0 则同参数重跑）
    if not detections:
        retry_th = float(getattr(cfg, "TARGET_CLASS_YOLO_RETRY_CONF_TH", 0.25) or 0)
        if not (0 < retry_th < conf_th):
            retry_th = conf_th
-        print(f"[YOLO-TARGET] 0 结果，重新检测 第2次 conf_th={retry_th}")
+        if logger:
+            logger.info(f"[YOLO-TARGET] 0 结果，重新检测 conf_th={retry_th}")
        raw_retry = _run(retry_th)
        detections_retry = _parse(raw_retry)
-        print(f"[YOLO-TARGET] 第2次 raw={raw_retry!r}")
-        print(
-            f"[YOLO-TARGET] 检测结果={detections_retry}"
-            f"（已忽略 class_id∈{tuple(sorted(ignore_class_ids))} 共{n_ignored[0]} 条）"
-        )
        if detections_retry:
-            raw, detections = raw_retry, detections_retry
+            detections = detections_retry

    if logger:
-        logger.info(f"[YOLO-TARGET] 检测结果={detections}")
+        logger.info(
+            f"[YOLO-TARGET] 检测结果={detections}"
+            f"（已忽略 class_id∈{tuple(sorted(ignore_class_ids))} 共{n_ignored[0]} 条）"
+        )

    if not detections:
        if logger:
            logger.warning("[YOLO-TARGET] 未检测到目标（已重检 1 次）")
        return None

-    # 上报口径：只在有效靶规格(20/40)里选置信度最高的一条，忽略 label=0 等无效类
-    valid = [d for d in detections if d["label"] in valid_labels]
-    if valid:
-        picked = max(valid, key=lambda d: d["confidence"])
-    else:
-        picked = {
-            "class_id": None,
-            "label": None,
-            "confidence": None,
-            "box_xywh": None,
-        }
+    def _best_valid(dets):
+        valid = [d for d in dets if d["label"] in valid_labels]
+        if not valid:
+            return None
+        return max(valid, key=lambda d: d["confidence"])

-    print(
-        f"[YOLO-TARGET] 上报={picked['label']}, class_id={picked['class_id']}, "
-        f"conf={picked['confidence']}, 有效候选={[(d['label'], d['confidence']) for d in valid]}"
-    )
+    def _crop_target(box, margin_frac):
+        """按检出框(带 margin)裁出放大 ROI；返回 Maix image，失败返回 None。"""
+        if not box or len(box) != 4:
+            return None
+        x, y, w, h = [float(v) for v in box]
+        try:
+            fw, fh = int(maix_frame.width()), int(maix_frame.height())
+        except Exception:
+            fw = fh = 0
+        mx, my = w * margin_frac, h * margin_frac
+        x0, y0 = int(max(0, x - mx)), int(max(0, y - my))
+        x1 = int(min(fw, x + w + mx)) if fw else int(x + w + mx)
+        y1 = int(min(fh, y + h + my)) if fh else int(y + h + my)
+        if (x1 - x0) < 32 or (y1 - y0) < 32:
+            if logger:
+                logger.warning(
+                    f"[YOLO-TARGET] 复检裁剪过小 [{x0},{y0},{x1},{y1}]"
+                )
+            return None
+        try:
+            cropped = maix_frame.crop(x0, y0, x1 - x0, y1 - y0)
+            if cropped is not None:
+                return cropped
+        except Exception:
+            pass
+        try:
+            from maix import image as maix_image
+
+            arr = maix_frame.to_numpy()
+            slab = np.ascontiguousarray(arr[y0:y1, x0:x1], dtype=np.uint8)
+            if slab.size == 0:
+                return None
+            return maix_image.cv2image(slab, False, True)
+        except Exception as exc:
+            if logger:
+                logger.warning(f"[YOLO-TARGET] 复检裁剪失败: {exc}")
+            return None
+
+    # 上报口径：只在有效靶规格(20/40)里取置信度最高的一条，class_id=1 已在解析阶段过滤
+    full_best = _best_valid(detections)
+
+    # 二次裁剪复检：把全图检出的靶框裁剪放大后再 detect 一次，与全图结果互相印证
+    # 复检用更低阈值 RECHECK_CONF_TH，置信度大于该值即接受
+    recheck_th = float(getattr(cfg, "TARGET_CLASS_YOLO_RECHECK_CONF_TH", 0.25))
+    crop_best = None
+    recheck_status = "off"
+    if full_best is not None and bool(
+        getattr(cfg, "TARGET_CLASS_YOLO_RECHECK_CROP", True)
+    ):
+        recheck_status = "fail"
+        margin_frac = float(
+            getattr(cfg, "TARGET_CLASS_YOLO_RECHECK_MARGIN_FRAC", 0.1)
+        )
+        crop_img = _crop_target(full_best["box_xywh"], margin_frac)
+        if crop_img is not None:
+            try:
+                raw_crop = detector.detect(
+                    crop_img, conf_th=recheck_th, iou_th=iou_th
+                )
+            except Exception as exc:
+                raw_crop = None
+                if logger:
+                    logger.warning(f"[YOLO-TARGET] 复检 detect 异常: {exc}")
+            if raw_crop is not None:
+                crop_dets = _parse(raw_crop)
+                recheck_status = "ok"
+                crop_best = _best_valid(crop_dets)
+                if crop_best is not None and crop_best["confidence"] <= recheck_th:
+                    crop_best = None
+                if logger:
+                    logger.info(
+                        f"[YOLO-TARGET] 复检(conf_th={recheck_th}, "
+                        f"裁剪框={full_best['box_xywh']}) 结果={crop_dets}"
+                    )
+
+    # 合并：以复检（裁剪放大）结果为主；复检无结果才用全图
+    picked = full_best
+    picked_from_crop = False
+    note = ""
+    if recheck_status == "ok":
+        if crop_best is None:
+            note = f"复检无 conf>{recheck_th} 的有效框，沿用全图"
+        else:
+            picked = crop_best
+            picked_from_crop = True
+            if crop_best["label"] == full_best["label"]:
+                note = f"复检一致({crop_best['label']})，以复检为准"
+            else:
+                note = (
+                    f"复检与全图不一致(全图={full_best['label']}@{full_best['confidence']}, "
+                    f"裁剪={crop_best['label']}@{crop_best['confidence']})，以复检为准"
+                )
+
+    # 置信度门槛：低于阈值不上报类别，避免硬猜（复检结果只需 > RECHECK_CONF_TH）
+    if picked is not None:
+        if picked_from_crop:
+            min_report_conf = recheck_th
+            below = picked["confidence"] <= min_report_conf
+        else:
+            min_report_conf = float(
+                getattr(cfg, "TARGET_CLASS_YOLO_MIN_REPORT_CONF", 0.5)
+            )
+            below = picked["confidence"] < min_report_conf
+        if below:
+            note = (note + "; " if note else "") + (
+                f"conf={picked['confidence']} 未达门槛{min_report_conf}，置信度不足"
+            )
+            picked = None
+
+    if picked is None:
+        label, class_id, confidence = None, None, None
+    else:
+        label = picked["label"]
+        class_id = picked["class_id"]
+        confidence = picked["confidence"]
+
+    if logger:
+        logger.info(
+            f"[YOLO-TARGET] 上报={label}, class_id={class_id}, conf={confidence}, "
+            f"全图候选={[(d['label'], d['confidence']) for d in detections if d['label'] in valid_labels]}, "
+            f"复检={recheck_status} {note}"
+        )
+        if note:
+            logger.warning(f"[YOLO-TARGET] {note}")

    return {
-        "label": picked["label"],
-        "class_id": picked["class_id"],
-        "confidence": picked["confidence"],
+        "label": label,
+        "class_id": class_id,
+        "confidence": confidence,
        "detections": detections,
    }

@@ -4,6 +4,6 @@
 应用版本号
 每次 OTA 更新时，只需要更新这个文件中的版本号
 """
-VERSION = '2.18.9'
+VERSION = '2.18.10'