Research on No-Reference Image Aesthetic Quality Assessment Methods

In the digital age, image aesthetic quality assessment is essential in computer vision. We enhance the efficiency of no-reference image aesthetic quality assessment using the No-reference Perceptual Quality Prediction Network (SAP-Net). By incorporating mixed-precision training and a multi-model fusion strategy, we reduce the need for extensive training data and shorten the training cycle. We also introduce the Super-resolution Reconstruction (BSR) architecture in the WBER module to enhance the reconstructed image quality. The loss function structure is optimized, and constraints are added for stability. Spatial attention is improved by increasing the complexity of the residual module and introducing the Multi-scale Spatial Attention mechanism (MSSA) to capture key information at multiple levels. Post-processing of attention maps is added to focus on practical features, reducing redundant information and enhancing evaluation accuracy. The improved SAP- Net algorithm boosts training efficiency by 10% in the IQA-ODI dataset and yields a 166% average improvement in the VAL_PSNR index for reconstructed image performance. Comparative experiments demonstrate that the enhanced SAP-Net algorithm enhances model performance, stability, and prediction accuracy, offering new perspectives and technical support for automated image aesthetic assessment.