| dc.description.abstract | In the domain of image denoising, there has been significant and rapid devel opment recently. Many prior noise modeling based and deep learning based al gorithms have shown outstanding results in denoising. However, the networks used by the state of the art methods are very deep and complex. We propose a imple yet effective Deep Multi patch Hierarchical Network that uses less mem ory and has fewer network parameters. In this network, multiple features of noisyimage patches from different spatial regions are combined using a fine to coarse hierarchical representation to get a clean image. While deep neural networks have had great success in denoising images using additive white Gaussian noise (AWGN), their performance on real world noisy images is weak. This is because their trained models are likely to overfit the simplified AWGN model, which differs significantly from the complex real world noise model. Hence we trained our model by real world noisy data to generalize the ability of the denoise network. In the encoder section of the network, we also added a non local module to extract dependencies between longdistant pixels in the image, which enhanced PSNR to 0.25 dB. Additionaly, The parallel connection of channel attention (CA) and pixelattention (PA) is added into the decoder to further enhance the performance. Different channels and pixels contain different levels of important information, and attention can give more weight to relevant information so that the network can learn more useful information. This resulted in a PSNR increment of 0.17 dB. When compared to most deep learning approaches, our architecture shows com petitive results while using fewer network parameters. | |
