Attention-Residual Convolutional Neural Network for Image Restoration Due to Bad Weather

Image quality degrades due to various reasons. In some circumstances, different weather conditions like fog, mist or rain have an impact on image visibility. Dust and pollution in the air can reduce the clarity of images taken outside. Thus, input images with poor visibility may reduce the effectiveness of computer vision related applications. The automated traffic monitoring systems, computer vision based smart systems used in different vehicles are few examples of such applications. Image restoration is essential of such applications for accurate implementation. In the proposed work, an end-to-end network is designed to restore images affected by rain, fog, dust and pollution. An integrated Convolutional Neural Network (CNN) with channel attention method is proposed for image restoration. In the proposed work a CNN is designed to reduce the loss between input degraded image and clear ground-truth image. Channel attention technique based on Style-based Recalibration Module (SRM) is applied on convolutional feature maps to improve the visibility of the restored image. The model is trained on a synthesized dataset and it is then evaluated on both synthesized and real-world out-door and traffic images. The experimental results demonstrate that the proposed method is more effective to several state-of-the-art methods both quantitatively and visually.