LEAN: Local-Enhanced Attention Network for Bad Weather Image Restoration

Bad weather image restoration is an important pre-processing task for many vision applications. Seeking a remarkable performance, existing works select to place their focus on one specific degradation, lacking the ability to cope with various weather conditions simultaneously. In this paper, we propose a local-enhanced attention network (LEAN) for restoring images with multiple weather degradations. First, we initiatively bring in cross-level transformer blocks to the encoder-decoder structure for a more effective utilization of the multi-scale features. Second, we design a local-enhanced self-attention (LESA) mechanism using a residual convolutional projection to reserve richer original information for subsequent feature reconstruction. Third, we bond smooth-L1 loss, frequency loss and perceptual loss together to accelerate the training. Experiment results show that LEAN has surpassed all the existing peering works in PSNR, increasing 2.06% on OutdoorRain-Test1, 3.04% on Snow100K-L and 3.36% on Raindrop-TestA, respectively and simultaneously.