A method of evaluation of the quality of segmentation image models, particularly of railway tracks and horizontal marking on roads

Abstract

The subject of the solution is a method for evaluating the quality of image segmentation models, in particular railway tracks and road markings, which converts image data (CHDS-in-1) in the input color format entering the first block (CHDS-B-1). The output of the first block (CHDS-B-1) is the image data (CHDS-B-1-out) of the image in the L*a*b* color format, whereby the colors are represented in such a way that the numerical distance between any two colors corresponds to the observer's perceived proximity of these colors. Next, the image data pixels (CHDS-B-1-out) are extracted based on the corresponding segmentation mask (CHDS-in-2), where the input to the second block (CHDS-B-2) is the image data (CHDS-B-1-out) of the image data of the tracks in the L*a*b* color format and the segmentation mask (CHDS-in-2) determining the regions of interest.Furthermore, in the third block (CHDS-B-3), color representatives (CHDS-B-3-out-1) of the pixels of the image data (CHDS-B-1-out) are generated and a model (CHDS-B-3-out-2) is introduced to them, mapping pixels to representatives of the image information. Furthermore, in the fourth block (CHDS-B-4), histograms (CHDS-B-4-out-1, CHDS-D-4-out-2, CHDS-D-4-out-3) of the pixels are generated against the color representatives (CHDS-B-3-out-1) of the pixels. Finally, the distance of the histograms (CHDS-B-4-out-1, CHDS-D-4-out-2, CHDS-D-4-out-3) of the color representation pixels entering the fifth block (CHDS-B-5) is determined, and the output from the fifth block (CHDS-B-5) is a score (CHDS-out) determining the quality of the performed segmentation, which determines the quality based on the ability of the segmentation to distinguish the object of interest from the background and at the same time distinguish the object of interest from its immediate surroundings.