Pattern Recognition and Image Analysis Компьютерное зрение и распознавание образов
Image Segmentation Based on Topological Decomposition and Pixel Resource
Аннотация
Abstract When solving practical problems, methods that can be adapted to a specific case without significant labor costs are generally used, since creating a proprietary solution will in any case require an evaluation of its effectiveness, i.e., a comparison with these very same universal methods. Therefore, adaptability and flexibility of methods can be called some of the main characteristics. In this paper, topological decomposition of images through the construction of connectivity components from the standpoint of adaptability in solving the problem of image segmentation has been considered. It has been shown that such decomposition consists of several stages: forming a queue of pixels, creating primitives, attaching a point to a primitive, and combining components. Each of them can be modified, thereby adapting the algorithm to the desired goal. In addition to the known parameters for component formation, two new modifications of the component merging stage, which are based on the sum of the brightness deviations of points within topological zones, have been presented. In the paper, the combination of different stages in solving the problem of segmentation of satellite images of the terrain has been compared. The first proposed modification determines the most significant component by the pixel resource, which is calculated through the sum of the deviations of its pixels. This has improved accuracy by 4–24% while still running fast, ranging from 200 to 486 ms depending on the steps selected for a 512 × 512-pixel image. In the second modification, the average deviation of the pixel brightness in the component, which significantly improves the segmentation accuracy with an accuracy of up to 93%, but requires more processing time, has been used.