Research On Representative Target Recognition Of Complex Ground Scene Based On Biological Perception Principle

Ground scene image take more and more important role in the field of militaryreconnaissance and precision attack. Consequently, researches on feature extraction andtarget recognition of the complex ground scene image are more and more significant.This dissertation focus on the problems of array target recognition, and use the basicprinciples of biology theory, and propose a series of methods of sub-object featureextraction and array recognition. The main work and achievements of this paper are asfollows:1. According to the difficulties of saliency detection in high resolution groundscene image, research on region extraction of ground scene image. First, introduce thescale effect of scale transformation and adopt several values to evaluate it. Then theappropriate scale of ground scene image is calculated by local variance method, andapplies spectral residual saliency detection to the appropriate scale image to extractsalient region feature. Last the experiment result show the effectiveness in quick regionextraction in ground scene image.2. According to the difficulties of contour extraction in ground scene image withtraditional edge detection algorithms, proposed a contour extraction algorithm based ongradient saliency map. First, studied the ill-posed problem in traditional gradient mapbased edge detection algorithms. Then applied spectral residual saliency detectionapproach to the gradient map, which output is the gradient saliency map. Last, thesegmentation based on gradient saliency map is proved to increase the performance ofground scene image contour feature extraction.3. First, the sub-object of array target is extract by the combination of regionfeature and contour feature obtained above, experiment results show the precision ofsub-object extraction and reduce the information of non-target. Then, according to thedifficulties of array recognition, research on the spectral graph theory and proposed theself tuning spectral clustering algorithm, which could automatically estimate theclustering numbers and obtain all the clustering vectors by one calculation. Last, theexperiment results of different ground scene image prove the effectiveness of theproposed algorithm.