SR-PAN-EffNet: A collaborative optimization approach for low-quality image classification and restoration

Low-quality images, due to degradation issues such as noise and compression artifacts, often lead to feature extraction distortion in classification models, thereby reducing classification accuracy. This issue is particularly prominent in practical computer vision applications. To address this, this paper proposes the SR-PAN-EffNet model, which integrates a semantic-guided SRGAN restoration module with a noise-aware PANet attention module, and employs end-to-end joint training to achieve the collaborative optimization of “restoration serving classification”. Experimental results show that the model achieves Top-1 accuracy of 68.5% and 57.6% on the ImageNet-1K and CIFAR-100 low-quality datasets, respectively, improving by 4.7–6.4 percentage points over NFNet-F4, with PSNR and SSIM also leading. Ablation experiments reveal that removing core modules results in a 2.9–6.8 percentage point decrease in accuracy. Future work will focus on improving the model’s real-time performance and robustness through lightweight design, extreme degradation adaptive optimization, and category-adaptive guidance, aiming to promote its application in scenarios such as surveillance and medical imaging.