Research on Steel Surface Defect Detection Based on Efficientnet-YOLOX

In order to ensure that the complexity of the model and the physical memory occupied by the model are reduced under the condition of high detection accuracy, an EfficientNetYOLOX object detection algorithm under the framework of deep learning was proposed, and the structure of EfficientNet+FPN was used to realize enhanced feature extraction and feature fusion. In addition, the unique YoloHead of YOLOX network is retained for prediction, which not only improves the accuracy of small target defect detection, but also makes the network model more lightweight. The experimental results show that the parameters of the EfficientNet-YOLOX network model are about 19.4M, the parameters of the YOLOXL network model are about 54.2M, and the Mean Average Precision (mAP) of EfficientNet-YOLOX is 88.24%, which is 2.75% higher than that of YOLOX-L, which meets the accuracy requirements of steel surface defect detection. And it greatly reduces the complexity of the network model and the number of parameters.