Adaptive multi-threshold object selection in remote sensing images
Information and Control Systems
Detection, selection and analysis of objects of interest in digital images is a major problem for remote sensing and technical vision systems. The known methods of threshold detection and selection of objects avoid using the processing results, therefore not providing a low probability of false alarms, and not keeping the shape of the selected objects well enough. There are only few results from the studies about quantifying the quality of such algorithms on either model or real images.
... Studying the effectiveness of algorithms for detecting, selecting, and localizing objects of interest using their geometric characteristics, when the object properties and background are a priori uncertain, and the shape of the selected objects is kept unchanged. Results: We have obtained and studied the characteristics of algorithms for detecting and selecting objects of interest on test models of monochrome images. These software-implemented algorithms use multi-threshold processing, providing a set of binary slices. This makes it possible to perform morphological processing of objects on each slice in order to analyze their geometric characteristics and then select them according to geometric criteria, taking into account the percolation effect which causes changes in the area, and fragmentation of the objects. As a result of analyzing these changes, an adaptive detection threshold is set for each of the selected objects. The selection allows you to significantly reduce the number of false positives during the detection and to use lower thresholds, increasing the correct detection probability. We present the detection characteristics and the results of test model processing, as well as the results of object selection on a real television and radar image, confirming the effectiveness of the considered algorithms. Practical relevance: The proposed algorithms can more effectively select objects on images of various nature obtained in remote sensing, material research or medical diagnostics systems. Their microprocessor implementation is much simpler than the implementation of universal trainable neural network algorithms.