DenoisingNet: An Efficient Convolutional Neural Network for Image Denoising
Yang Li, Zhuang Miao, Rui Zhang, Jiabao Wang · 2019
Since convolutional neural networks (CNNs) were used in image denoising field, they have showed state-of-the art denoising quality beyond traditional methods. However, the existing CNN denoising models have a high computational cost and require high memory capacity, which is impractical for embedded applications or mobile platforms. Inspired by the efficiency of the depthwise separable convolutions introduced in MobileNet, this paper proposes a novel network structure, named DenoisingNet, which designed specifically to minimize model size while maintaining image denoising performance. The resulting DenoisingNet model possesses a model size of 369 KB, which is an order of magnitude lighter than other state-of-the-art networks while maintaining the denoising quality. Extensive experimental results indicate that very small deep neural network architectures can be designed for real-time image denoising that are well suited for embedded scenarios. To the best of our knowledge, DenoisingNet is the first lightweight network architecture for image denoising.