Drainage Pipeline Defect Detection Based on Improved YOLOX

Aiming at the problems of low automation and long detection cycle of urban drainage pipeline defect detection tasksan improved YOLOX-s object detection algorithm was proposed. Based on YOLOX-s, the MobileNetV3 network was used to replace the CSPDarknet53 feature extraction network to achieve lightweight improvement. Introduced SPPF(Spatial Pyramid Pooling Fast) structure to fuse local and global features from the outputs of three different layers of the MobileNetV3 feature extraction network, enhanced the model's discriminative ability for distinguishing objects from similar backgrounds. The CIoU(Complete-IoU) loss function was used to solve the lossless gradient backpropagation problem that existed when the prediction box and the ground truth box did not overlap during the training phase. It sped up the model convergence speed and optimized the fitting effect of the prediction box and the ground truth box. Experiments are conducted on the urban drainage pipe defect detection dataset. The results show that the improved YOLOX-s model reduces the calculation amount by 37.09%, the parameter amount by 39.13%, increases the detection speed by 29.12%, and increases the mAP(mean average precision) by 0.88%. While ensuring high detection accuracy, it significantly enhances the speed of model detection.