Learning Deformable Kernels Predict Network for Burst Images Denoising

Abstract

Mobile phones are becoming more common, people has higher requirement about the quality of the pictures that captured by handheld device, in order to reduce the noise in image, we studied the prediction network (KPN) denoising algorithm, to overcome the disadvantage of low efficiency of traditional convolution receptive field, a deformable kernel prediction network (DEF_KPN) algorithm for burst images denoising is proposed in this paper. The deformable convolution structure is used to preprocess the noise image, so that the kernel prediction network can generate a pixel-by-pixel filter core more suitable for the image structure, so as to make the image contour more clearer after denoised. We obtain bayer image by anti-ISP method, and then synthesizing the training data by jitter clipping and adding poissonian-gaussian noise. Applying Adam algorithm and annealing strategy to train neural network to convergence. The experiment results show that; the DEF_KPN algorithm proposed in this paper is superior to the KPN algorithm in both synthetic noise data and real noise data. In terms of synthetic noise data, the PSNR index is improved by about 0.5db. On the real noise data, PSNR index was increased by 0.5dB. The SSIM index has been improved by 0.005. Experiments on burst mobile phone bayer images illustrate that : The neural network trained by our method can successfully de-noise the multi-frame images taken by the phone.