基于马尔科夫链的显著性区域检测算法研究

显著性检测算法常通过计算像素之间的差异来确定显著性，但是对像素的选取通常是固定的，容易忽略图像中物体的边界信息，导致最终检测结果中目标的边界比较模糊。借鉴生物视觉注意机制，提出了一种新的基于超像素和马尔科夫链的显著性区域检测算法，将图像分割成若干个超像素，使用Wasserstein距离衡量超像素之间颜色、方向和位置的差异来建立马尔科夫链，将显著性检测问题转换为马尔科夫链上的随机游走问题，使用它的平稳分布作为图像的显著度。实验结果表明，相对于两种经典算法，所提出的算法在主要目标及其边界的提取精度等方面取得了较为满意的效果。

Saliency detection algorithm based on Markov chain of superpixels

Current visual saliency detection algorithms mainly base on contrast of pixels, and most of the algorithms ignore the boundaries by selecting pixels in a rigid way. According to biological visual attention, this paper presents a new method for saliency detection based on Markov chain and superpixels. The paper extracts superpixels from the original image, and defines a Markov chain with the Wasserstein distance between superpixels, thus the problem of saliency detection is formulated as a Markov random walk model. Equilibrium distribution of the Markov chain is the result of saliency detection. Experimental results show that this saliency detection algorithm achieves a satisfactory result in extraction accuracy of main target and its boundaries compared with two representative algorithms.