Abstract: The surface defects of strip steel in industrial production have the characteristics of changeable defect types, cluttered background and low contrast. The existing defect detection methods cannot detect the complete defect objects. In order to solve the above problems, a strip defect detection algorithm based on U-shaped residual encoder-decoder network is proposed. In the encoding stage, the algorithm uses a fully convolutional neural network to extract rich multi-scale defect features, and combines an attention mechanism to accelerate model convergence. In the decoding stage, the saliency information encoded in the encoding stage is recovered using the proposed U-shaped residual decoding network. Furthermore, a residual refinement network is designed to further optimize the coarse saliency map. The experimental results show that the proposed algorithm has strong robustness.