<p>Grain boundary characterization is fundamental to understanding the microstructure evolution laws and property modulation of steel materials. However, grain boundary recognition methods based on traditional image processing are often difficult to accurately extract the complete grain boundary structure in complex tissue backgrounds, while methods relying on expert experience are inefficient and subjective. The CycleResGAN model developed in this study introduces a residual linkage module in the generator based on the standard CycleGAN framework. Experimental results show that the model can quickly generate high-precision images of grain boundary features, and exhibits stable enhancement effects on metallographic images with different magnifications and different image qualities. The validation shows that the CycleResGAN model significantly outperforms the traditional CycleGAN method in the grain boundary reconstruction task, and the method can be extended and applied to the grain boundary analysis of various steel tissues such as pearlite.</p>