In this paper, a local activity measurement of the clipped and normalized variance or standard deviation is proposed to drive anisotropic diffusion and relative total variation (RTV) to work better for structural preservation. Firstly, two novel edge-stop functions are introduced for our local activity-driven anisotropic diffusion (LAD-AD) to efficiently remove severe artifacts and preserve the fine geometry structures in HEVC-compressed depth images. Secondly, we propose a simple yet effective local activity-driven RTV (LAD-RTV) with the way of the product between gradient and the local activity measurement for image smoothing and scale representation. Meanwhile, both color-sharing information and each-channel discriminative information are considered, which are significant to color image edge-preserving but not included in the RTV model. Besides, LAD-RTV leverages the form of the division of gradient and the local activity measurement to resolve the problem of general image de-noising by regarding the noises as the duplicate texture elements. Experimental results have validated that the proposed LAD-AD can greatly improve the precision of the HEVC-compressed depth image and the quality of its synthesized image. Additionally, large numbers of results have shown our LAD-RTV is superior to several existing methods.
Image filtering is an effective way to improve the performance of many applications, such as edge detection and image editing [1, 2, 3, 4, 5, 6, 7, 8, 9]. Since different types of images have different characteristics and different applications have different requirements, image filtering algorithms should be designed for each case properly. For example, depth images having smooth regions divided by sharp boundaries represent scene’s geometry structures. The high-quality boundaries should be preserved, because they will strongly affect 3D video coding’s efficiency and the quality of view synthesis with depth image-based rendering (DIBR). Therefore, the quality of the virtual-view images should be enhanced after filtering contaminated depth images. Meanwhile, the precision of depth image should be kept at least or even be greatly improved. For natural images, when we want to remove image noises, we need to preserve both image’s structures and textural details at the same time. If we want to apply image smoothing, we should remove texture details but keep major structures. Although there are a large number of works for image filtering [5, 6, 7, 8, 9, 10, 11, 12, 13, 14], most of these algorithms tend to be computationally complex, which are not well suitable for practical applications. Meanwhile, their algorithms are specifically designed for one model, which lose the sight of generalization, so we need to re-design a new algorithm for each new model. Based on the above observations, our motivation is whether a robust statistic measurement can be easily inserted into some models to adaptively control model’s trade-off parameter between data term and regularization term. Meanwhile, this statistic measurement should not significantly increase computational complexity. Besides, this measurement can be easily put into most of low-level image processing model without complicated expert’s design. It is generally known that standard deviation is a good measurement on the degree of dispersion for a set of data. Because each image patch’s standard deviation can be quickly computed through matrix operations, it will not significantly increase the complexity of the filtering. Consequently, we introduce a local activity measurement of variance or standard deviation to drive different models for better solutions.