Efficient Blind Image Deblurring Network Based on Frequency Decomposition

Many deep learning-based blind image deblurring solutions model high and low frequencies and reconstruct frequency representations to restore sharp images. However, these methods suffer from performance bottlenecks in real-world scenarios because they ignore averaging phenomena in blurring (e.g., frequency degradation when texture and color regions are mixed). To solve this problem, we consider exploring and reconstructing high frequency, low frequency, and medium frequency separately. We propose to decompose frequency information based on discrete cosine transform (DCT), and design intra-frequency interaction (Intra-FI) module to interact in a cross-dimensional manner within frequency-domain groups to reshape the weight representation of frequency channels. To further improve the quality of image restoration, we developed the inter-frequency attention (Inter-FA) module for frequency channel group fusion and interaction to better utilize effective information to guide restoration. We validate the built intra-frequency interaction and inter-frequency attention network (FIFANet) on the widely used benchmark dataset GoPro and the latest real-world blur dataset RSBlur and MC-Blur. Experimental results show that FIFANet achieves state-of-the-art results in terms of both performance and efficiency, and is also one of the best generalization models.