Entropy-Based Locally Adaptive Thresholding for Image Segmentation

We presented a local thresholding method that is adaptively computed to segment images. The construction of the locally adaptive threshold mimics how human visual system interprets images, that is, identifying homogeneous regions in images in the presence of noise, clutters, and low contrast. Using a sliding window to raster scan an image, the method uses local statistics to measure homogeneity of pixel distribution in the window. The locally adaptive threshold is determined by evaluating a pixel against the local mean and an offset term, which is built by the local standard deviation and entropy. In implementation, except for the size of the sliding window, the method involves no parameters and, thus, requires no user input. For this reason, the method has a consistent and robust performance across users. Test results on different types of images demonstrated the robust performance of the method in separating fine details about an image and suppressing false positives. In practice, the method is extendable to parallel processing as the calculations are locally determined.

Keywords: segmentation, entropy, adaptive thresholding, standard deviation