Media.Player

Evolutionary segmentation (Technion)

总结

The world of image analysis is shifting from pixel-based to object-based algorithms. The first stage of an object based algorithm is finding objects, usually performed with a segmentation algorithm. Yet contemporary algorithms are not adequate. Each algorithm has many parameters and thresholds, and there is no way of knowing what good threshold values the user should set in advance. This significantly impacts the segmentation. The resulting segmentation is undefined, and so the user does not know how effective the segmentation is or how it can be improved. With an insufficient initial object-finding stage, the rest of the analysis will not be accurate.We present a segmentation method that outputs the best segmentation for a given segmentation evaluation function based on evolutionary algorithms. Instead of setting criteria and thresholds that produce an undefined segmentation, we propose a novel way of approaching image segmentation: First the properties that are useful in the specific area of research are defined, and then the algorithm finds the best segmentation with respect to these properties. The key elements of our method are: the recombination algorithm, which mixes two segmentation maps from the input to create 'offspring' containing a mixture of the initial 'parent' maps; and the mutation algorithm, which mutates the segmentation map. The results of the mutation are evaluated to determine if the stopping criteria for the segmentation have been met. If not, the process is repeated.

技术优势

User-defined image segmentation that finds the best segmentation with respect to a given segmentation evaluation function
Can be applied to any type of imaging
Research can now shift focus from investigating new segmentation algorithms to rather focus on segmentation evaluation

技术应用

Creating geographic information system (GIS) layers out of aerial and satellite images
Remote sensing
Medical imaging processing
Computer vision

ID号码

COM-1473

国家/地区

以色列