diff --git a/Storage/customs/huzesama/onnx-detect-IoU/README.md b/Storage/customs/huzesama/onnx-detect-IoU/README.md
new file mode 100644
index 0000000..c59d27c
--- /dev/null
+++ b/Storage/customs/huzesama/onnx-detect-IoU/README.md
@@ -0,0 +1,38 @@
+# ONNX Detect IoU
+
+## 功能
+
+- 基于IoU的神经网络检测，更易于检测被大目标覆盖的小目标
+
+## 文件说明
+
+- `onnxDetect.py`：Recognition 实现
+- `pipeline.json`：pipeline 示例
+
+## 使用方式
+
+使用方式与内置的recognition相似
+
+`custom_recognition_param` 参数：
+
+- `model`：基于py文件所在目录为根目录（可选）：`../model/detect.onnx`
+- `expected`: 支持输入label字符串（必填）：`[int,"str"]`
+- `conf_threshold`: 置信度（可选）。默认：`0.75`
+- `iou_threshold`: 交并比（可选）。默认：`0.25`
+- `label`: 默认从onnx文件读取，以数组的形式输入（可选）：`["label1","label2",...]`
+- `order_by`: 排序方式（可选）。支持`Horizontal|Vertical|Score|Area|Random|Expected`，与内置神经网络检测类似的排序。默认`Score`
+- `index`: 输出索引结果（可选）。默认：`0`
+
+示例可见 `pipeline.json`。
+
+## 依赖
+
+- Python
+- MaaFramework Agent SDK（`maa.custom_recognition`）
+- onnx runtime
+
+## Files
+
+- `maahub_meta.json`: website metadata
+- `README.md`: contributor-facing overview
+- `main.py`: example entry file
diff --git a/Storage/customs/huzesama/onnx-detect-IoU/maahub_meta.json b/Storage/customs/huzesama/onnx-detect-IoU/maahub_meta.json
new file mode 100644
index 0000000..64ca3b3
--- /dev/null
+++ b/Storage/customs/huzesama/onnx-detect-IoU/maahub_meta.json
@@ -0,0 +1,20 @@
+{
+  "id": "huzesama/onnx-detect-IoU",
+  "title": "ONNX Detect IoU",
+  "description": "根据MaaFramework内置NeuralNetworkDetect修改的神经网络检测(IoU)。解决检测存在全屏大目标时，细分的小目标少甚至没有的问题。适用于需要搭配用户自定义优先级排序进行细分检测的场景",
+  "author": "huzesama",
+  "source": "MaaWoA",
+  "sourceGithub": "https://github.com/huzesama/MaaHub",
+  "tags": ["custom", "recognition", "onnx","IoU", "python"],
+  "createdAt": "2026-08-20",
+  "updatedAt": "2026-08-20",
+  "version": "0.0.1",
+  "mfwVersion": "5.12.2",
+  "entry": "onnxDetect.py",
+  "readme": "./README.md",
+  "status": "stable",
+  "type": "custom",
+  "language": "python",
+  "runtime": "python 3.14.6",
+  "dependencies": ["maa framework sdk","onnx runtime"]
+}
diff --git a/Storage/customs/huzesama/onnx-detect-IoU/onnxDetect.py b/Storage/customs/huzesama/onnx-detect-IoU/onnxDetect.py
new file mode 100644
index 0000000..b85edb4
--- /dev/null
+++ b/Storage/customs/huzesama/onnx-detect-IoU/onnxDetect.py
@@ -0,0 +1,253 @@
+import sys
+import os
+import re
+import json
+import numpy as np
+import onnxruntime as ort
+import random
+#增加了与内置相当的order_by
+from maa.agent.agent_server import AgentServer
+from maa.custom_recognition import CustomRecognition
+from maa.context import Context
+from maa.tasker import Tasker
+
+AGENT_DIR = os.path.dirname(os.path.abspath(__file__))
+
+def order_boxes(boxes: list[dict], order_by: str, expected_indices=None) -> list[dict]:
+	if order_by == "Horizontal":
+		# x 升序，同 x 按 y 升序
+		return sorted(boxes, key=lambda b: (b["x"], b["y"]))
+	elif order_by == "Vertical":
+		# y 升序，同 y 按 x 升序
+		return sorted(boxes, key=lambda b: (b["y"], b["x"]))
+	elif order_by == "Score":
+		return sorted(boxes, key=lambda b: b["score"], reverse=True)
+	elif order_by == "Area":
+		return sorted(boxes, key=lambda b: b["w"] * b["h"], reverse=True)
+	elif order_by == "Random":
+		boxes = boxes.copy()
+		random.shuffle(boxes)
+		return boxes
+	elif order_by == "Expected" and expected_indices:
+		# 按 expected 列表里给出的顺序排序，未匹配项排最后
+		order_map = {idx: i for i, idx in enumerate(expected_indices)}
+		return sorted(boxes, key=lambda b: order_map.get(b["cls_index"], len(order_map)))
+	else:
+		return boxes  # 不支持的 order_by 或未设置，保持原顺序
+
+def parse_labels_from_metadata(session: ort.InferenceSession) -> list:
+	#从 ONNX 模型元数据解析 labels，等价于框架内置的 parse_labels_from_metadata
+	meta = session.get_modelmeta()
+	custom_meta = meta.custom_metadata_map or {}
+
+	names_str = None
+	for key in ("names", "name", "labels", "class_names"):
+		if key in custom_meta:
+			names_str = custom_meta[key]
+			break
+
+	if not names_str:
+		return []
+
+	# 解析 {0: 'cat', 1: 'dog', ...} 格式（支持单/双引号）
+	pattern = re.compile(r"(\d+)\s*:\s*['\"]([^'\"]+)['\"]")
+	label_map = {int(idx): label for idx, label in pattern.findall(names_str) if label}
+
+	if not label_map:
+		return []
+
+	max_index = max(label_map.keys())
+	if max_index < 0 or max_index > 10000:
+		return []
+
+	labels = [""] * (max_index + 1)
+	for idx, label in label_map.items():
+		labels[idx] = label
+	return labels
+
+def nms_iou(boxes: list[dict], threshold: float = 0.7) -> list[dict]:
+	#IoU NMS，替代 MaaFramework 内置 NeuralNetworkDetector 的 IoM 版本
+	boxes = sorted(boxes, key=lambda b: b["score"], reverse=True)
+	kept = []
+
+	def iou(a, b):
+		ax1, ay1, ax2, ay2 = a["x"], a["y"], a["x"] + a["w"], a["y"] + a["h"]
+		bx1, by1, bx2, by2 = b["x"], b["y"], b["x"] + b["w"], b["y"] + b["h"]
+
+		inter_x1, inter_y1 = max(ax1, bx1), max(ay1, by1)
+		inter_x2, inter_y2 = min(ax2, bx2), min(ay2, by2)
+		inter_w = max(0, inter_x2 - inter_x1)
+		inter_h = max(0, inter_y2 - inter_y1)
+		inter_area = inter_w * inter_h
+
+		area_a = a["w"] * a["h"]
+		area_b = b["w"] * b["h"]
+		union_area = area_a + area_b - inter_area  # 关键差异：除以并集而非单框面积
+
+		return inter_area / union_area if union_area > 0 else 0.0
+
+	while boxes:
+		best = boxes.pop(0)
+		kept.append(best)
+		boxes = [b for b in boxes if iou(best, b) < threshold]
+
+	return kept
+
+
+@AgentServer.custom_recognition("onnxDetect")
+class onnxDetect(CustomRecognition):
+	_session = None
+	_labels = None
+
+	def _load(self, model_path, labels):
+		if self._session is None:
+			self._session = ort.InferenceSession(model_path, providers=["CPUExecutionProvider"])
+			self._labels = labels if labels else parse_labels_from_metadata(self._session)
+
+	def analyze(
+		self,
+		context: Context,
+		argv: CustomRecognition.AnalyzeArg,
+	) -> CustomRecognition.AnalyzeResult:
+		param = json.loads(argv.custom_recognition_param or "{}")
+		model_path = os.path.join(AGENT_DIR, param.get("model", "../model/detect.onnx"))##AGENT_DIR是基于该py文件所在目录为根目录
+		raw_expected = param.get("expected", [])
+		conf_threshold = param.get("conf_threshold", 0.75)
+		iou_threshold = param.get("iou_threshold", 0.25)
+		labels = param.get("labels", [])
+		self._load(model_path, param.get("labels", []))
+		order_by = param.get("order_by", "Score")
+		index = param.get("index", 0) 
+
+		image = argv.image
+		roi = argv.roi
+
+		input_tensor, scale, pad_left, pad_top = self._letterbox(image, roi)
+
+		input_name = self._session.get_inputs()[0].name
+		outputs = self._session.run(None, {input_name: input_tensor})
+
+		candidates = self._parse_outputs(outputs, scale, pad_left, pad_top, conf_threshold, roi)
+
+		#expected label 字符串或下标 int 混合 ----
+		expected_indices = self._resolve_expected(raw_expected, self._labels)
+
+		#expected 非空但全部未匹配到任何 label/下标，视为无匹配结果 ----  
+		if raw_expected and not expected_indices:  
+			return CustomRecognition.AnalyzeResult(  
+			    box=None,  
+			    detail={"msg": "expected labels not matched", "raw_expected": raw_expected},  
+
+		)  
+		if expected_indices:
+			candidates = [c for c in candidates if c["cls_index"] in expected_indices]
+
+		boxes = nms_iou(candidates, threshold=iou_threshold)
+		boxes = order_boxes(boxes, order_by, list(expected_indices) if order_by == "Expected" else None)
+
+		if not boxes:
+			return CustomRecognition.AnalyzeResult(box=None, detail={"msg": "no detection"})
+
+		n = len(boxes)
+		idx = index if index >= 0 else n + index
+		if idx < 0 or idx >= n:
+			return CustomRecognition.AnalyzeResult(box=None, detail={"msg": "index out of range"})
+
+		best = boxes[idx]
+		return CustomRecognition.AnalyzeResult(
+			box=(best["x"], best["y"], best["w"], best["h"]),
+			detail={"all_boxes": boxes},
+		)
+
+
+	def _resolve_expected(self, raw_expected: list, labels: list) -> set:
+		#将 expected 中的 label 字符串解析为下标，int 原样保留 
+		resolved = set()
+		for item in raw_expected:
+			if isinstance(item, int):
+				resolved.add(item)
+			elif isinstance(item, str):
+				if item in labels:
+					resolved.add(labels.index(item))
+				else:
+					# 找不到对应 label，记录日志但不中断
+					print(f"[MyNNDetect] Warning: label '{item}' not found in labels list")
+			else:
+				print(f"[MyNNDetect] Warning: invalid expected item type: {item!r}")
+		return resolved
+
+
+	def _letterbox(self, image, roi):
+		x, y, w, h = roi.x, roi.y, roi.w, roi.h
+		cropped = image[y : y + h, x : x + w]
+
+		input_shape = self._session.get_inputs()[0].shape  # [1, 3, H, W]
+		input_h, input_w = input_shape[2], input_shape[3]
+
+		raw_h, raw_w = cropped.shape[:2]
+		scale = min(input_w / raw_w, input_h / raw_h, 1.0)
+		resized_w, resized_h = int(raw_w * scale), int(raw_h * scale)
+
+		import cv2
+		resized = cv2.resize(cropped, (resized_w, resized_h), interpolation=cv2.INTER_AREA)
+
+		pad_w, pad_h = input_w - resized_w, input_h - resized_h
+		pad_left, pad_top = pad_w // 2, pad_h // 2
+		padded = cv2.copyMakeBorder(
+			resized, pad_top, pad_h - pad_top, pad_left, pad_w - pad_left,
+			cv2.BORDER_CONSTANT, value=(114, 114, 114),
+		)
+
+		blob = padded[:, :, ::-1].transpose(2, 0, 1).astype(np.float32) / 255.0
+		blob = np.expand_dims(blob, axis=0)
+		return blob, scale, pad_left, pad_top
+
+	def _parse_outputs(self, outputs, scale, pad_left, pad_top, conf_threshold, roi):
+		raw_output = outputs[0][0]  # shape: [5+nc, N]
+		candidates = []
+
+		for i in range(raw_output.shape[1]):
+			cls_scores = raw_output[4:, i]
+			cls_index = int(np.argmax(cls_scores))
+			score = float(cls_scores[cls_index])
+			if score < conf_threshold:
+				continue
+
+			cx, cy, w, h = raw_output[0, i], raw_output[1, i], raw_output[2, i], raw_output[3, i]
+			x = (cx - w / 2 - pad_left) / scale
+			y = (cy - h / 2 - pad_top) / scale
+
+			# ---- 关键修改：加入 label 信息 ----
+			label = (
+				self._labels[cls_index]
+				if self._labels and cls_index < len(self._labels)
+				else f"Unknown_{cls_index}"
+			)
+
+			candidates.append({
+				"x": int(x) + roi.x,
+				"y": int(y) + roi.y,
+				"w": int(w / scale),
+				"h": int(h / scale),
+				"score": score,
+				"cls_index": cls_index,
+				"label": label,
+			})
+
+		return candidates
+
+
+def main():
+	if len(sys.argv) < 2:
+		print("Usage: python my_agent.py <socket_id>")
+		exit(1)
+
+	socket_id = sys.argv[-1]
+	Tasker.set_log_dir("./debug")
+	AgentServer.start_up(socket_id)
+	AgentServer.join()
+	AgentServer.shut_down()
+
+
+if __name__ == "__main__":
+	main()
\ No newline at end of file
diff --git a/Storage/customs/huzesama/onnx-detect-IoU/pipeline.json b/Storage/customs/huzesama/onnx-detect-IoU/pipeline.json
new file mode 100644
index 0000000..98b73f7
--- /dev/null
+++ b/Storage/customs/huzesama/onnx-detect-IoU/pipeline.json
@@ -0,0 +1,26 @@
+{
+	"神经网络检测目标": {
+		"recognition": {
+			"type": "Custom",
+			"param": {
+				"custom_recognition": "onnxDetect",
+				"custom_recognition_param": {
+          "model":"../model/detect.onnx",
+					"expected": [
+						"cat",
+            0,
+            1
+					],
+          "conf_threshold": 0.25,
+          "iou_threshold": 0.7,
+          "label": ["label1","label2"],
+          "order_by": "Score",
+          "index": 0
+				}
+			}
+		},
+		"next": [
+			"下一个节点"
+		]
+	}
+}