Paper Title
Filter Flow-Based Lightweight Approach for Motion Deblurring

Abstract
Image deblurring is a widely researched topic in the field of Image Processing and Computer Vision. In the past two decades, many solutions have been proposed to solve the problem of image deblurring, but the necessity of faster and better performing deblurring algorithms for real-time applications still exists. One of the most popular applications of deblurring is removal of motion blur from images, which is created during data acquisition due to camera shake. In this paper, we propose a novel Convolutional Neural Network-based framework that achieves fast and accurate deblurring of images with both good visual performance and speed. The proposed framework leverages the concept of filter flow to reconstruct deblurred images from the visually blurry images. Given a blurred image, we predict the filter for the image using a two-stream light-weight CNN rather than using conventional computationally expensive methods. The predicted filter would then be applied to the blurred image along the RGB channels, eventually computing the reconstructed deblurred image. The proposed framework was trained on the BSD500, DIV2K and FLICK1024 datasets. The framework was evaluated on the testing set of moderate blur, large blur and compared with other state of the art methods. The proposed model was assessed using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) which are the generally used evaluation metrics for image quality assessment. The proposed framework offers interpretability which is not provided by the many CNN-based deblurring approaches. Our approach is also 10-100x faster with better results than the other filter flow-based methods. Keywords - Convolutional Neural Networks, Filter Flow, Image Deblurring