Super-Resolution Reconstruction of Single Image Combining Bionic Eagle-Eye and Multi-scale

The sufficient extraction of high and low frequency features is the key to solving the loss of reconstructed image details and ensuring the maximum restoration of useful information. This needs to fully exploit the effective features in the low resolution image and give play to the complementary advantages of different convolution kernels by fusing different features. Therefore, a multi-scale fusion single image super-resolution reconstruction method Be-MRN is proposed following the eagle vision system. Inspired by the eagle eyes to recognize objects from three directions, we design an eagle-eye feature extraction module using three ways in parallel to extract feature. It uses the principle of cross-field of left and right eyes to fuse information in three channels, and uses dilated convolutions to mine effective features. The feedback module receives effective features, uses the attention mechanism to capture cross channel interactive information and eliminates redundant information, and then completes deep feature extraction through four layers of iteration to ensure the complementary interaction and deep mining between different features. The multi-scale reconstruction module uses different convolutions to extract complementary information of deep features, and uses multi-branch in parallel to complete image reconstruction. Experiments show that, for Urban100 with scale factor × 4, when compared to MSICF, Be-MRN increases the PSNR and SSIM values by 1.8% and 2.1% respectively. For the four datasets with × 2, × 3 and × 4 scale factors, Be-MRN obtains the highest objective index, which can highlight the edge and texture information of the image.