MSLEM-YOLOv7: Multi-Scale Lightweight Efficient Model for Pipeline Defect Detection Based on YOLOv7

Based on the collected real-world dataset, this paper proposes an improved YOLOv7 network model aimed at enhancing defect detection in urban sewer pipes. The model integrates a defect detection layer, a lightweight convolution operation, and the CBAM attention mechanism to perform multi-scale feature extraction and fusion, all while minimizing the number of parameters. Testing on an urban sewer pipe defect dataset shows that the model achieves an average precision (mAP@0.5) of 97.29% and an average prediction time of 69.38 ms. Compared to the original YOLOv7, the model reduces the parameter count by 11.21M and the computational cost by 28.71G. Experimental results show that the performance of the Pipe-YOLOv7 model reaches the current state-of-the-art level.