Enhanced Image Restoration and Aesthetic Evaluation Using Modified YOLOv5 and Multi-Scale Residual Gated Convolutional Networks

Abstract

Cultural heritage represents humanity's invaluable treasure, yet existing image restoration methods suffer from inadequate edge feature extraction capabilities and aesthetic evaluation approaches that cannot be co-trained with emotional factors. Consequently, this research proposes an image restoration method based on an enhanced You Only Look Once version 5 small (YOLOv5s) and multi-scale residual fusion gated convolutions, integrating an aesthetic evaluation model that incorporates emotional elements. This approach employs the enhanced YOLOv5s for image extraction, utilizing the Canny operator as the edge detection algorithm. It incorporates multi-scale residual blocks and gated convolutional networks within the generative adversarial network, employing pixel reconstruction, perceptual, and style loss as the joint loss function. The research inserts emotion label extraction and emotion fusion modules into the residual network. Experiments demonstrate that the enhanced YOLOv5s achieves a maximum image extraction accuracy of 95.3%, surpassing both YOLOv5s and YOLOv8 by 12.5% and 0.3% respectively, whilst converging significantly faster. The image restoration model exhibits higher structural similarity indices, with restored images most closely approximating reality at a maximum value of 94.5%. Removing multi-scale residuals substantially impacts model performance. The aesthetic evaluation model achieves a maximum Spearman's correlation coefficient of 0.792 with the lowest computational complexity. Consequently, the proposed methodology effectively enhances image restoration capabilities and aesthetic evaluation quality, thereby facilitating the wider dissemination of cultural heritage.