All-in-One Image Restoration via Frequency Domain Decomposition

In practical applications such as traffic accident detection, environmental factors can severely degrade image quality, making image restoration a critical factor for improving visual task performance. Since single image restoration methods struggle to address multiple degradation scenarios, researchers have attempted to use a unified model to simultaneously solve various degradation problems. However, existing methods face difficulties in addressing the coupling issues during the restoration process across different frequency domains of degraded images, which limits model performance improvements. Therefore, this paper designs a model that decomposes images into high-frequency and low-frequency spaces through Fast Fourier Transform (FFT). For the high-frequency components, a High-frequency Feature Adaptive Processing Module (HFPM) is employed for feature enhancement and cross-layer fusion, effectively extracting detailed textures while suppressing interference. For low-frequency components, a Low-frequency Feature Transformation and Enhancement Module (LTEM) captures global contextual information through cyclic convolution, improving the integrity of structural contours. To validate model effectiveness, tests were conducted on deraining, denoising and dehazing tasks. Experimental results demonstrate that this model maintains good restoration performance when simultaneously processing multiple degradation tasks.