An improved U²-Net segmentation model was proposed to address the issues of weak feature extraction ability and low segmentation accuracy in metal coating defect image segmentation. Firstly, an improved receptive field block light (RFB_l) module was embedded in the U-shaped residual block (RSU) to form a new feature extraction layer, which enhanced the ability to learn detailed features and solved the problem of low segmentation accuracy caused by limited receptive field in the network. Secondly, in the decoding stage of the U²-Net segmentation model, an effective contour enhanced attention (CEA) mechanism was introduced to suppress redundant features in the network, obtain feature attention maps with detailed position information, enhance the difference between boundary and background information, and achieve the more accurate segmentation results. The experimental results show that the average intersection and union ratio, accuracy, precision, recall, and F₁-measure of the model on two metal coating peeling and corrosion datasets are 80.36%, 96.29%, 87.43%, 84.61%, and 86.00%, respectively. Compared with commonly used SegNet, U-Net, and U²-Net segmentation networks, the performance of the model is significantly improved.