During the non-contact flaw detection of a rust-covered bronze mirror, X-ray imaging typically fails to reveal the extent of damage due to the thickness difference between the mirror edge and core. In this study, the X-ray signal from a bronze mirror was used as an input to construct a generative confrontation fusion network. An optimization strategy that enhances the bronze mirror X-ray information fusion was designed to address the reconstruction blur caused by the $L_{\mathrm{2}}$ loss and gradient operator, and the expression of multiscale feature details, such as textures and cracks. By utilizing the feature learning process of the $L_{\mathrm{2},\raisebox{1ex}{$\mathrm{1}$}\!\left/ \!\raisebox{-1ex}{$\mathrm{2}$}\right.}$ loss regularization generator, the smoothing of the data that was generated using the $L_{\mathrm{2}}$ loss was improved; moreover, the Laplacian $L_{\mathrm{t}\mathrm{e}\mathrm{x}}$ pattern loss was defined to strengthen the effect of training network on the extraction of decorations and diseases. Furthermore, a multiscale feature fusion module was added to the training network to improve the quality of the generated information. Thus, considering the experimental comparison involving seven fusion methods, the cross entropy value of the proposed algorithm in two of the five groups is poor. However, the values are optimal in the control data, including entropy, average gradient, spatial frequency, joint entropy, and non-reference feature mutual information. This can effectively reveal the detection information of the bronze mirror during X-ray flaw detection.