Boosting the Performance of Image Restoration Models Through Training With Deep-Feature Auxiliary Guidance

Many neural network architectures have been proposed for image restoration to improve the accuracy of restored images while maintaining reasonable computational costs. However, most previous studies have primarily focused on designing new architectures, with relatively less attention on optimizing training strategies. In this paper, we introduce deep-feature auxiliary guidance (DFAG), a novel training method that enhances the accuracy of image restoration models without increasing inference time or modifying their structures. DFAG adds auxiliary losses to the feature maps at each multi-scale level during training, which stabilizes deep feature learning and improves feature quality. All modules used for DFAG are only active during training and are removed for inference, resulting in no additional computational cost.We validate DFAG on various multi-scale encoder-decoder-based image restoration models. Our experiments demonstrate consistent performance improvements across multiple tasks, including real image denoising, Gaussian denoising, motion deblurring, JPEG artifact reduction, and super-resolution. Our results highlight that DFAG is an effective strategy to boost restoration performance without any architectural modifications. The source code and pre-trained models will be released.