Abstract

Subject of the study. Methods for visible spectrum image augmentation in thermal image classification tasks are considered. Aim of the study. The study aims to improve the generalization ability of neural networks trained on visible spectrum images to recognize thermal images. Method. Existing sets of thermal images have limited size, and obtaining such data requires expensive equipment. At the same time, the classifiers trained on visible spectrum data show low classification accuracy on data of different optical spectra. Various methods for enriching thermal datasets and solving the problem of object recognition are available, including those using synthesized images. However, these approaches require the use of thermal images in specific forms, which restricts their possible applications. Meanwhile, artistic methods for modeling far-infrared scenes based on visible spectrum images exist. These methods allow visual similarity to be achieved, for example, by means of contrast correction and transformation of color channel values. We proposed and investigated a preliminary image transformation method to determine whether a classifying neural network was capable of extracting features from modified visible spectrum images that could be generalized to thermal data. Main results. With the developed method of augmentation and preparation of visible spectrum data, the level of classification errors was reduced from 17% to 6%. Practical significance. The proposed training method was able to improve the classification accuracy of thermal imaging data without using images of the appropriate spectrum in the training sample. This approach can be used as a method of data enrichment, for example, if the available resources for obtaining thermal imagery data are limited.

© 2024 Optica Publishing Group