Techniques for Improving Color Segmentation in the Task of Identifying Objects on Aerial Images

摘要

Automatic identification of objects on aerospace images allows to increase the efficiency of making the necessary management decisions in important areas of human activity. A promising approach is the object-oriented image analysis, in which image segmentation is performed and the resulting image regions are classified into target object categories. The main problem constraining the effectiveness of this approach is the lack of segmentation accuracy. To solve this problem, the paper proposes techniques aimed at obtaining more relevant color segments of the image: partition of the image into frames with subsequent merging of color areas lying on the borders of adjacent frames, splitting color regions in relatively narrow places, as well as adaptive approximation of the edges of color areas. An experimental study of improving the quality of identification of objects as a result of the application of the developed techniques is carried out. The experiments were conducted on high resolution aerial images from a publicly available dataset. It is shown that the proposed techniques make a significant contribution to improving the efficiency of the logical approach to the identification of objects based on shape features.