Image Thresholding Based on Relevance Vector Machine

Image thresholding is a simple but effective segmentation approach which is widely applied in image analysis, pattern recognition and computer vision. To obtain the optimal threshold value or more threshold values, classical thresholding methods have to search the whole gray levels of a given image, based on specific optimization strategy. Recently, a novel multi-threshold technique based on support vector regression (SVR) is developed, where threshold values can be determined directly from support vectors (SV). However: 1.There are no reliable methods to select threshold values from abundant SVs; 2.The computational cost is generally high due to the cross-validation procedure to estimate the hyper-parameters incurred in SVR; 3.It does not well deal with the images with unimodal histogram. Motivated by relevance vector machine (RVM), a new method called RVM-based threholding algorithm is introduced. Under the proposed framework, the relevance vectors (RVs) are very sparse and thereby the thresholds can be easily found from a few candidate RVs. In addition, the unimodal case can also be handled well. That is, the proposed algorithm not only keeps the advantages of SVR-based thresholding method, but also solves its open problems. Experimental results validate the effectiveness, efficiency and adaptability.