Loss Function for Training Models of Segmentation of Document Images

Abstract

This work is devoted to improving the quality of segmentation of images of various scientific papers and legal acts by neural network models by training them using modified loss functions that take into account special features of images of the appropriate subject domain. The analysis of existing loss functions is carried out, and new functions are proposed that work both with the coordinates of bounding boxes and use information about the pixels of the input image. To assess the quality, a neural network segmentation model with modified loss functions is trained, and a theoretical assessment is carried out using a simulation experiment showing the convergence rate and segmentation error. As a result of the study, rapidly converging loss functions are created that improve the quality of document image segmentation using additional information about the input data.