Handling Unbalanced Data on Binary Image Segmentation Modeling Using Kolmogorov Forward Theory and External Force

A binary image with imbalanced data refers to an image in which pixel classes are not equally represented. In other words, one class is much more present than the other. In the case of the binary image, the concentration of black pixels on a white background is not balanced. The image could have a large majority of white pixels with some sporadically scattered black areas or conversely a small number of white pixels could be surrounded by a large area of black pixels. In imbalanced images, where one class is significantly overrepresented, some segmentation algorithms may have difficulty identifying objects or boundaries. This can lead to biased results. If the segmentation is performed using a supervised learning model, a severe imbalance can lead to biased learning where the model favors the majority class. This can give unsatisfactory results because the model does not properly learn to segment the pixels of the minority class. The aim of this article is to design a method to use the cost function based on the resampling method to balance the image data, thereby exposing the two black and white classes to a more representative proportion, which can improve the detection and segmentation of target objects via the Kolmogorov forward equation which is derived from the Fokker-Planck equation.