Minimum error thresholding for infrared image under constraint of cross entropy

Focusing on the infrared images which have similar statistical distributions between object and background, conventional thresholding methods only take variance or entropy by someway as criterions for thresholding selection and they don't consider actual characteristics of infrared images, so the segmentation results are unsatisfactory. In order to solve this problem, a novel method of cross entropy constrained minimum error thresholding of infrared image was proposed. Firstly, cross entropy was called to measure the similarity between object and background's statistical distributions. The smaller the cross entropy was, the more similar the distributions were. After that, classification error was minimized when the cross entropy was below a certain value. Constraint of cross entropy guaranteed the segmentation process fits actual characteristics of images. Meanwhile, minimum of classification error ensured the effectiveness of segmentation results. The principles of proposed method are clear and the parameter setting is simple. Experimental results on real images show that compared with several classic thresholding methods, the proposed method can improve thresholding segmentation accuracy of infrared image with similar statistical distributions between object and background effectively.