HybridSR: A Hybrid Deep Learning-Based Image Super-Resolution Framework

Abstract

Single Image Super-Resolution (SISR) is a technol-ogy that reconstructs high-resolution images from relatively low-resolution ones. Therefore, SISR involves creating new image data from a single image, which makes it a structurally ill-defined process. This research involves HybridSR, an innovative hybrid deep learning approach that improves the perceived image details while ensuring accurate reconstruction and fidelity to the original image. HybridSR is designed to include hierarchical residual learning and channel attention mechanisms that are made up of several residual blocks that are used to capture multiscale feature dependencies from residual groups of images. A global skip connection allows for the retention of lowfrequency information in added detail and makes training more stable. The use of sub-pixel convolution (PixelShuffle) provides a more efficient method of upsampling image data in the feature domain, thereby providing HybridSR with an advantage. HybridSR utilises a composite loss function made up of L1 loss, perceptual loss calculated using a pre-trained VGG16 network, and structural similarity index (SSIM) loss. The purpose of the composite loss function is to analyse both the pixel-level accuracy and the realism of how an image is perceived by the end-user. Using test-time augmentation (TTA) was created in order to make HybridSR more robust when analysing images during inference and also to help reduce the variance of prediction results. HybridSR is trained and evaluated using the DIV2K dataset. The experimental results show that HybridSR is able to obtain a PSNR value of about 34 dB as well as an SSIM value of 0.92 which is better than many state-of-the-art techniques such as SRCNN, VDSR, EDSR, and RCAN. The proposed method has been shown to provide a superior reconstruction of high-frequency details, maintain structural consistency, and create sharper images with reduced artifacts, as evidenced by both quantitative and qualitative assessments.