一、前言

? ? 了解到Sahi，是通過切圖，實現提高小目標的檢測效果。sahi 目前支持yolo5\yolo8\mmdet\detection2 等等算法，本篇主要通過實驗onnx加載模型的方式使sahi支持yolov10。

二、代碼

（1）轉換模型

? ? ? 首先使用 ?conda創建虛擬環境，配置好yolov10環境，然后 pip 將sahi 安裝上

pip install sahi

? ? ?將yolov10 模型導出為 onnx格式文件，命令窗cd 至 模型文件所在目錄 執行

 yolo export model=vitrolite_best.pt  format=onnx opset=11 simplify

? ? ?轉換成功，將在目錄下生成同名 onnx格式文件

?（2）加載onnx 模型推理代碼

參考 sahi 提供的demo 文件 inference_for_yolov8_onnx.ipynb?，分塊大小和重疊比例可設置

from sahi import AutoDetectionModel
from sahi.utils.cv import read_image
from sahi.utils.file import download_from_url
from sahi.predict import get_prediction, get_sliced_prediction, predictimport timeif __name__ == '__main__':yolov8_onnx_model_path = "runs\\detect\\train_v102\\weights\\vitrolite_best.onnx"  #加載自己的onnx模型文件#yolov8_onnx_model_path = "D:\\Project\\yolov8\\weights\\yolov8x.onnx"category_mapping = {'0': 'p0', '1': 'p1', '2': 'p2', '3': 'p3'  }  #類別映射，換成你的detection_model = AutoDetectionModel.from_pretrained(model_type='yolov8onnx',model_path=yolov8_onnx_model_path,confidence_threshold=0.3,category_mapping=category_mapping,device='cuda:0', # or 'cuda:0'  #這里要使用GPU)#img_path = "datasets\\vitrolite\\images\\val\\c144846_5_10.png"   #推導圖片路徑#result = get_prediction(read_image(img_path), detection_model)  #第一次啟動GOU ，時間比較慢,必須先啟動一次#分塊檢測result = get_sliced_prediction("D:\\Project\\vitroliteDefect\\tile_round1_train_20201231\\train_imgs\\197_2_t20201119084924170_CAM1.jpg",detection_model,slice_height=640,slice_width=640,overlap_height_ratio=0.05,overlap_width_ratio=0.05)result.export_visuals(export_dir="demo_data/")

（3）修改sahi的接口文件??

找到pip安裝的sahi 位置 ， 修改?yolov8onnx.py

修改 _post_process 函數 ， 主要是由于yolov8 和yolov10 輸出 shape有所不同，yolov10沒有nms

    def _post_process(self, outputs: np.ndarray, input_shape: Tuple[int, int], image_shape: Tuple[int, int]) -> List[torch.Tensor]:image_h, image_w = image_shapeinput_w, input_h = input_shapepredictions = np.squeeze(outputs[0])  # 不用.T 轉置  ,  #( 300,6)# 在下面這個地方改動，by zjy ,  self.confidence_threshold 是0.3#for row in predictions:#print(  "row:" ,  row.shape   )scores =  predictions[: , 4]  #shape 為 ( 300,6) ,第5個，下標為4#self.confidence_threshold = 0.9predictions = predictions[ scores > self.confidence_threshold, : ]scores = scores[scores > self.confidence_threshold]boxes = predictions[:, :4]boxes = boxes.astype(np.int32)class_ids = predictions[:,  5 ].astype(np.int32)# Format the resultsprediction_result = []for bbox, score, label in zip(boxes , scores , class_ids ):bbox = bbox.tolist()cls_id = int(label)prediction_result.append([bbox[0], bbox[1], bbox[2], bbox[3], score, cls_id])"""  # Filter out object confidence scores below thresholdscores = np.max(predictions[:, 4:], axis=1)predictions = predictions[scores > self.confidence_threshold, :]scores = scores[scores > self.confidence_threshold]class_ids = np.argmax(predictions[:, 4:], axis=1)boxes = predictions[:, :4]# Scale boxes to original dimensionsinput_shape = np.array([input_w, input_h, input_w, input_h])boxes = np.divide(boxes, input_shape, dtype=np.float32)boxes *= np.array([image_w, image_h, image_w, image_h])boxes = boxes.astype(np.int32)# Convert from xywh two xyxyboxes = xywh2xyxy(boxes).round().astype(np.int32)# Perform non-max supressionsindices = non_max_supression(boxes, scores, self.iou_threshold)# Format the resultsprediction_result = []for bbox, score, label in zip(boxes[indices], scores[indices], class_ids[indices]):bbox = bbox.tolist()cls_id = int(label)prediction_result.append([bbox[0], bbox[1], bbox[2], bbox[3], score, cls_id])"""prediction_result = [torch.tensor(prediction_result)]# prediction_result = [prediction_result]return prediction_result

三、結果

運行推理代碼，在 demo_data 文件夾下生成結果圖片, 實驗圖片是一張高分辨率瓷磚圖片