Extraction of low-contrast blurred objects from space images using convolutional neural networks
Valeria A. Gmyria, Andrey P. Treshchalin · Journal of Optical Technology · 2025
Subject of study . The study is focused on the efficiency of convolutional neural networks (CNNs) in extracting low-contrast blurred objects in images. Aim of study. The aim was to evaluate the accuracy and reliability of CNN models in extracting low-contrast blurred objects. In addition, we identified the networks that outperformed the traditional threshold-based algorithm among the set of networks investigated in this study. We also identified the signal-to-noise ratio (SNR) values of objects at which the best-performing networks outperformed the traditional threshold-based algorithm. Method. In this study, CNNs were used to obtain binary masks of starry-sky images. To suppress noise, pixel-wise multiplication of the binary mask with the original image was performed. The result of this multiplication was then fed to the input of a center-of-mass algorithm to calculate the centroids of the extracted objects. The accuracy of object extraction was evaluated using the segmentation quality and centroid error metrics, and the reliability of the extraction was assessed using the extraction coefficient. Main results. An object extraction algorithm based on CNNs was proposed. To generate training and test datasets, an algorithm simulating the operation of an onboard optoelectronic system of a spacecraft (OES SC) was implemented in the MATLAB programming environment. The results demonstrated that the U-Net and SegNet models outperformed the traditional threshold-based approach in extracting low-contrast blurred objects. The SNR ranges in which these architectures demonstrated the best performance in terms of object extraction efficiency were determined. Practical significance. The low-contrast object extraction performance of the proposed neural-network-based object extraction algorithm is higher than that of the traditional algorithm. Moreover, the error in centroid calculation is lower for the proposed model than for the traditional one. The results obtained in this study will serve as a basis for future investigations aimed at implementing the proposed object extraction algorithm in a prototype OES SC.