An Adaptive PM Model Based on Difference Eigenvalue for Image Restoration

Classic Perona-Malik (PM) model is usually used to smooth noise for the degraded image. However, to our knowledge, a well-known defect of the PM model is prone to cause ‘staircase’ effect and blur image fine feature because of the lack of correct diffusion strength in diffusion process. To tackle the problem, we will improve the PM model by developing a new structure descriptor to adjust anisotropic diffusion strength for smoothing noise and preserving image feature adaptively. The new structure descriptor is based on image local geometry-Hessian matrix, called difference eigenvalue, which can adaptively track edge feature and high degree of homogeneity in an image, even when the observed image is blurred. Experiments on both nature and medical images show that the improved PM method can achieve a superior performance than the traditional methods in terms of visual inspection and quantitative measurements.