Parameter efficient image denoising by combining depthwise separable convolution and gating mechanism

抄録

Image denoising aims to remove noise that is inevitably generated in the image acquisition process, and is still a fundamental problem in image processing. In particular, there is a lot of interest in removing real image noise that occurs in actual devices, instead of synthetic noise such as Gaussian noise. Existing real image denoising methods that use convolutional neural networks (CNNs) or transformers tend to have extremely large parameter sizes and computational costs, which makes it impossible to use these methods on smartphones and embedded devices. Gated texture convolutional neural network (GTCNN), proposed by Imai et al., achieves a better trade-off between the number of parameters and performance on synthetic noise by using two independent CNNs for context extraction and denoising based on that, and combining the results using a gating mechanism. We propose a new variant of GTCNN that further improves the trade-off between the number of parameters and performance of the method, i.e., parameter efficiency, by replacing specific convolutional layers of GTCNN with depthwise separable convolution, and investigate its performance on real image noise. Since the selection of layers to which the replacement is applied is not obvious, we identified the optimal replacement points that do not impair performance through comprehensive experiments. The experimental results on existing real image noise datasets showed that the proposed method achieves significant improvements in parameter efficiency compared to existing image denoising methods, including lightweight methods such as GTCNN.