A Diffusion Equation for Improving the Robustness of Deep Learning Speckle Removal Model

Speckle removal aims to smooth noise while preserving image boundaries and texture information. In recent years, speckle removal models based on deep learning methods have attracted a lot of attention. However, it was found that these models are less robust to adversarial attacks. The adversarial attack makes the image recovery of deep learning methods significantly less effective when the speckle noise distribution is almost unchanged. In purpose of addressing the above problem, we propose a diffusion equation-based speckle removal model that can improve the robustness of deep learning algorithms in this paper. The model utilizes a deep learning image prior and an image grayscale detection operator together to construct the coefficient function of the diffusion equation. Among them, there is a high possibility that the deep learning image prior is inaccurate or even incorrect, but it will not affect the performance and the properties of the proposed diffusion equation model for noise removal. Moreover, we analyze the robustness of the proposed diffusion equation model in terms of theoretical and numerical properties. Experiments show that our proposed diffusion equation speckle removal model is not affected by adversarial attacks in any way and has stronger robustness.