Research On Target Recognition Methods In Remote Sensing Images Based On Knowledge Inference And Visual Mechanism

With the development of remote sensing technologies, spectral, spatial and temporal resolutions in remote sensing continually elevate, and thus the amounts of image data, which are acquired by remote sensor and then transferred to ground, drastically increase. How to properly recognize specific ground targets from huge-amount remote sensing image data becomes a quite challenging and urgently demanding important research topic. By simulating human visual functionality, studying the recognition process and physiological mechanisms of human to recognize targets is a critical research direction. The reason for human's capability of rapidly and properly recognizing specific target is that human have necessary knowledge and can acquire new knowledge via knowledge inference. The basis of the ability of employing knowledge is the perfect visual perceptual system which is formed via long-period evolution. Therefore, researches on how to apply knowledge inference and visual mechanisms to recognition of ground targets in remote sensing images are important.In this dissertation, the human's ability of knowledge inference and partial visual mechanisms of human are simulated, and the application of knowledge inference and visual mechanisms to ground target recognition in remote sensing images is researched. Research content include remote sensing ground targets recognition method based on knowledge inference, remote sensing textural target recognition method based on visual feed-forward model, background texture suppression for remote sensing targets based on surround suppression mechanism.First, for target recognition in remote sensing images, a knowledge-based framework and an architecture design scheme of a general knowledge base were proposed. The knowledge-based target recognition framework centers on knowledge and can better achieve sharing and re-using of knowledge during target recognition process. The framework accomplishes target recognition tasks mainly through knowledge inference. According to the proposed framework, airport target recogntion based on knowledge inference was obtained. The general knowledge base take ontology as knowledge representation form, and it can better achieve managing and employing knowledge, and also it can make knowledge inference be a more important role. Besides, in this dissertation, primary researches on the key step in the target recognition framework, i.e., semantic mapping method, was presented.Secondly, based on visual feed-forward model and using Standard Model Features (SMF), recogniton for the typical textural target, suburb inhabitant, in remote sensing image is addressd. When recognizing textural targets using SMF alone, the pixels near target region boundary are often mis-recognized. The influence of mis-recognition is reduced via combination of textural region partition results and SMF recognition results, and thus effective recognition for suburb inhabitant target is achieved. In addition, irrelevant information was excluded in advance via circumstance discrimination between suburb regions and urban regions, textural region partitions, which favor the elevation of final proper recognition. Experiment results demonstrate the effectiveness of our method.Thirdly, surround suppression mechanism was applied to the suppression of background textures surrounding a target, and the goal of effectively reducing background textures and giving prominence to target contour is achieved, which provides beneficial condition for target recognition. Surround supperession mechanism is the visual mechanism of human visual system, which can suppress textures surrounding target contour and maintain contour itself. We combine surround suppression mechanism with multiscale properties of images, and through strong suppression to textural edges as well as maintenance of contour, background textures are remarkably suppressed and nice contour maintenance is acquired, therefore, the computational amount for properly recognizing target reduces, which greatly increased recognition rate. Experiment results of airport target demonstrated the practicality of the presented method.Finally, whole conclusions are made and further research directions are addressed.