Automatic threshold selection guided by maximizing Pearson correlation

Many images exhibit non-modal, unimodal, bimodal, or multimodal gray level distributions. Current thresholding methods often struggle with images whose gray level distributions do not conform to a bimodal or unimodal pattern. We propose a novel bi-level threshold selection technique guided by maximizing Pearson correlation, addressing these four distribution types within a unified framework. Our method entails a multiscale multiplicative transformation of the image to create a template, extracting contours from binary images at different thresholds, and using Pearson correlation to assess the similarity between these contours and the template. The threshold with the highest similarity is chosen as the final threshold. Tested against seven methods on 20 synthetic images and 50 real-world images with non-modal, unimodal, bimodal or multimodal distribution patterns, our method showed more flexible adaptability of threshold selection and lower misclassification error, although it did not exhibit an advantage in computational efficiency.