In noncontact nondestructive testing, an ancient copper mirror coated with rust cannot present complete disease information in X-ray imaging because of the different thicknesses between the edge and center of the mirror. However, to satisfy the requirements of cultural-heritage observations, X-ray images are fused, which contain numerous cracks, erosions, and other characteristics of various diseases. Scattered and small cracks are not only difficult to detect but are also susceptible to oversegmentation. To further improve the accuracy and image quality of the segmentation of ancient bronze mirror diseases, the root mean square error (RMSE) and peak signal-to-noise ratio (PSNR) indices of the images were improved while all cracks and erosions were segmented. The U²-Net was selected as the basic model and optimization strategies were considered to improve the quality and accuracy of the segmented images under the U-shaped nesting concept. Additionally, a spatial channel-attention mechanism was utilized to enhance disease detection and extraction. Subsequently, a UD-block module that utilizes dilated convolution was designed to improve the detection ability of scattered small cracks during network training. A two-level misalignment link mechanism was incorporated into the network structure to improve the PSNR and RMSE. In the fusion stage of saliency maps, a pyramid attention segmentation module was added such that the segmentation effect is more consistent with human visual perception. Finally, the result of experimental comparison and the analysis of current flaw-detection algorithms show that the Dice coefficient, Jaccard index, accuracy, Hansdorff dimension, RMSE, and PSNR obtained experimentally are the best, which can provide a reference for future studies pertaining to cultural-relic flaw-detection algorithms.