Curvilinear thresholding method for noisy images based on 2D histogram

Image thresholding is a useful method in many image processing and computer vision applications. Otsu method is one of popular thresholding methods and has been frequently used as a classical technique in various applications. A premise of this method is that the probability of edge region is supposed to be zero. However, the premise is not proper in some situations such as the images are corrupted by noises. To solve this problem, a curvilinear thresholding method (CTM) is proposed based on the traditional Otsu method. We give a recursive algorithm of line thresholding method (LTM), which is a particular case of CTM, to verify our ideas. In addition, Otsu based method has another weakness that is it gives satisfactory results only when the numbers of pixels in each class are close to each other. Otherwise the threshold will be biased. This paper introduces a two-dimensional (2D) histogram projection method to correct the Otsu threshold. A fast algorithm for searching the valley of one-dimensional (1D) projected histogram is also given based on wavelet transform. Experimental results show that the proposed method performs much better than the traditional Otsu method.