Trainable Diffusion Network Based on Morphological Laplacian
Gouki Okada, Makoto Nakashizuka · 2021 International Symposium on Intelligent Signal Processing and Communication Systems (ISPACS) · 2021
This paper presents a deep network for image processing based on diffusion equation and morphological filtering. The diffusion equation has been applied to image processing, and is unwrapped to the trainable network, of which parameters are trained by the back-propagation. In this paper, we introduce the morphological Laplacian that is defined by morphological filters to the diffusion network. Since the morphological filtering can be implemented without multiplications, almost computations of the network are min and max operations and addition. Due to this property, the proposed network can be implemented with low-precision unsigned integer arithmetic. We apply the proposed morphological diffusion network to the Gaussain denoising and image completion. Denosing examples show the proposed network obtains denosing results comparable to BM3D [1] with the number of parameters that is about 1/15 of full scale deep convolutional neural networks [2]. Image completion examples show that the result of the proposed method is superior to the convolutional neural network in terms of PSNR with less number of the parameters.