Super Resolution Reconstruction of Bidirectional Feature Flow Images Based on SRGAN

Abstract

Over the years, in order to prevent the impact of leaf diseases on crop yields, deep learning technology has been introduced into the field of leaf disease recognition and has achieved good results. Deep learning techniques rely on high-quality images, and encountering low-quality images can cause a significant decline in model performance. To avoid performance degradation, low-quality images are often reconstructed using image super-resolution before being input into recognition models. However, existing super-resolution models often suffer from detail blurring and structural distortion due to their unidirectional feature transfer when dealing with complex textures and structures, which limits further performance improvement. To this end, this article innovatively introduces a bidirectional feature flow module, which achieves collaborative enhancement and more accurate reconstruction of multi-level image features through parallel processing and dynamic fusion of local details and global context. Compared to the original model, the PSNR and SSIM have improved by 0.47dB and 1.50% respectively.