DefocusSR2: An efficient depth-guided and distillation-based framework for defocus images super-resolution

Existing image super-resolution (SR) methods often lead to oversharpening, particularly in defocused images. However, we have observed that defocused regions and focused regions present different levels of recovery difficulty. This observation opens up opportunities for more efficient enhancements. In this paper, we introduce DefocusSR2, an efficient framework designed for super-resolution of defocused images. DefocusSR2 consists of two main modules: Depth-Guided Segmentation (DGS) and Defocus-Aware Classify Enhance (DCE). In the DGS module, we utilize MobileSAM, guided by depth information, to accurately segment the input image and generate defocus maps. These maps provide detailed information about the locations of defocused areas. In the DCE module, we crop the defocus map and classify the segments into defocused and focused patches based on a predefined threshold. Through knowledge distillation and the fusion of blur kernel matching, the network retains the fuzzy kernel to reduce computational load. Practically, the defocused patches are fed into the Efficient Blur Match SR Network (EBM-SR), where the blur kernel is preserved to alleviate computational demands. The focused patches, on the other hand, are processed using more computationally intensive operations. Thus, DefocusSR2 integrates defocus classification and super-resolution within a unified framework. Experiments demonstrate that DefocusSR2 can accelerate most SR methods, reducing the FLOPs of SR models by approximately 70% while maintaining state-of-the-art SR performance.