Generalized Deep Learning Model for Restoration of Degraded Images under Multiple Degradations
Jiseon Moon, Siwon Hwang, Jitae Shin · 2024
In the real world, images are degraded by a variety of degradations, most of which belong to adverse weather conditions. Multiple degradation restoration studies remove degradation through complex networks consisting of multiple encoders and decoders. To this end, we propose a single deep learning model for multiple degradation removal in images. First, we pro-pose a novel classifier to improve the quality of images corrupted by multiple degradations. Given an input image with unknown degradation, contrastive learning is performed to identify the type of degradation. Images with the same degradation are defined as positive pairs and are trained to get closer to each other, while images with different degradation are defined as negative pairs. Regularization on the contrast loss prevents overfitting and specific weight values from becoming large. Second, we propose a restoration network to recover from degradation. For restoration, we apply a deformable convolutional layer and a spatial feature transformation layer and perform feature attention to learn meaningful information from each output. In addition, a mixup operation is applied to preserve the features of each layer. The proposed method is trained on a total of four degradation phenomena: image dehazing, image desnowing, image de raining, and image deblurring. With this contribution, we propose a multi degradation removal model for images based on deep learning. Extensive experimental results demonstrate the image restoration performance both quantitatively and qualitatively. Experimental results shows up to 1.60dB improvement in PSNR compared to the existing work.