基于改进YOLOv5m的水电厂工器具识别系统研究OA

To address the need for rapid and accurate identification of tools during the borrowing and returning process in hydropower plants,as well as to prevent the issues of incorrect or missed borrowing,a tool dataset named Tool-Data is established,and a lightweight detection algorithm based on an improved YOLOv5m is proposed.This method replaces the original feature extraction network with MobileNetV3 and substitutes the conventional convolutional modules in the original network with optimized cross-stage depthwise separable convolutional modules,aiming to reduce the number of parameters and computational load of the network.Meanwhile,the SE attention mechanism is introduced to mitigate background interference,thereby enhancing the model's recognition accuracy for small and medium-sized targets.Furthermore,the anchor box dimensions are re-optimized based on the K-means clustering algorithm,and the Mosaic data augmentation technique is improved.The adoption of the DIOU_NMS algorithm increases the accuracy of filtering bounding boxes and reduces the rate of missed detections for small targets.Experimental results on the Tool-Data dataset show that the improved YOLOv5m achieves precision,recall and mean average precision(mAP)values of 90.5%,89.28%and 93.38%,respectively,on the tool detection dataset,which are 0.55,18.24 and 10.94 percentage points higher than those of the original YOLOv5m.The enhanced YOLOv5m lightweight model meets the requirements for high efficiency and precision in tool identification under complex conditions during the borrowing and returning process.