Anisotropic Dynamic Diffusion Model for Texture Preserving De-noising of Tomato Images

Abstract:

Due to interference of external environments and monitoring systems, the acquired images of agricultural products are degraded by noises. The noises affect quality testing of agricultural products. This paper researched the denoising of tomato images based on anisotropic diffusion model. First, by analyzing anisotropic diffusion process of Perona-Malik (P-M) model, a new method of calculating gradient threshold was proposed. It introduced local variances of images to the 2norm method. As a result, the new method distinguished texture details and achieved dynamic selection of gradient thresholds. Second, structural similarity image measurement (SSIM) was selected as the stopping criterion, which made selection of diffusion iterations adaptive. These two steps together formed an anisotropic dynamic diffusion model for texture preserving denoising of tomato images. Finally, two groups of comparison tests were taken under different noise standard deviations of 5, 10, 15, 20, 25, and 30. The first group of comparison test was performed among the SSIM criterion, minimum mean squared error criterion，SNR criterion and decorrelation criterion. Results of the first group showed that using SSIM as iterative stopping criteria was effective and stable. The second group of comparison test was performed among the proposed model, the conventional P-M model, and the 2norm model. From visual effect, images denoised by the proposed model had more and clearer texture details. And objective evaluation of the denoised image quality was achieved by using the peak signal to noise ratio (PSNR) and gradient magnitude similarity deviation (GMSD). Compared with P-M model and 2norm model, average PSNR of images denoised by the proposed model was the highest and average GMSD of images denoised by the proposed model was reduced by 15.5% and 19.1% respectively. It demonstrated images denoised by the proposed model had lower residual noises and greater similarity to original images. In conclusion, the proposed model can remove noises while maintaining texture details, which can contribute to subsequent quality testing of agricultural products.