Target Recognition Of High Resolution Remote Sensing Image Based On Multi-Media Neural Cognitive Computing Model

The problem of target recognition in high-resolution remote sensing image is key technologies of civil systems, marine traffic monitoring, emergency rescue preparedness and Unmanned Autonomous Systems(UAS) etc. such as unmanned aerial vehicles, unmanned ground vehicles, unmanned underwater vehicles, unmanned surface vehicles and other autonomous robots. It is also core technologies of military Automatic Target Recognition(ATR) systems, such as military reconnaissance, precision-guided weapon, monitoring of the ocean conditions and so on. With the development of high resolution earth observation system,the applications require to extracts more valuable information of targets from high resolution image. However, the interpretation systems of moderate or low resolution of remote sensing image can not meet the high-performance demand of Target Classification and Recognition(TCR) system. Especially, the accuracy and real-time problems of target recognition are very remarkable to diversity target in the environment of complex background. Moreover, due to the application contains the military background, difficulties of data acquisition and technology control regime and other reason, the public reports of TCR core technology and method are relatively scarce at present. To address the technology bottleneck of target recognition, this paper focuses on TCR model building and algorithm design of high precision and high performance for high resolution remote sensing image, and provides state-of-the-art technical support to the TCR research.The research is conducted in the following steps. First, analyzing the technical bottleneck of TCR problem for the high resolution remote sensing images, the hierarchical Multimedia Neural Cognitive Computing(MNCC) model was proposed to solve TCR application problems of remote sensing image by simulation theory of neural and cognitive. The second step is that the algorithms of target detection and scene classification were design respectively based on processing and saliency computing of MNCC model. The third step is to design target recognition algorithm of high resolution remote sensing image based MNCC framework. The hierarchical ensemble classifier is constructed to improve the state-of-the-art performance of target recognition, and the data augmentation methods were designed to solve the problem of target recognition of small samples and complex objects. Finally, the parallel target recognition algorithms were studied to improve the recognition efficiency, andcorrelation algorithms were applied to research and development software of remote sensing image processing system.The main contributions of this dissertation as following subjects:(1)The hierarchical MNCC models were proposed to TCR application problems of remote sensing image. For TCR core problem of high resolution remote sensing images, the structure of white matter network and mechanism of neocortex information processing were analyzed deeply, and cognitive system of the hierarchical architecture of visual processes, as well as attention, emotions, memory and other functions were seriously studied. Then the brain-like and TCR-oriented hierarchical MNCC models were designed, and formal description of TCR algorithms based on MNCC model was given.(2)The algorithms of remote sensing target detection based on spiking neural network were proposed. In order to overcome the limitations of the natural image saliency algorithm,the remote sensing images visual saliency framework VSF-MNCC was designed based on spiking neural network and pulse coupled neural network under the guidance of the saliency theory of MNCC model. Then SD-SNN algorithm based on visual saliency for ship detection in remote sensing image was proposed, and its false alarm rate and missed detecting ratio was reached to 9.48% and 11.02% respectively in ship detect visible light data set HRSHTD and high resolution SAR ship images. The VSF-MNCC saliency map was more high resolution,and has good performance for blob target detection.(3)The algorithm of remote sensing scene classification based on MNCC model was proposed. As an important part of target recognition, the remote sensing scene classification algorithm SC-MNCC was designed based on the MNCC framework. These results make the good performance for the TCR application, and provide a priori context knowledge for improving accuracy. The algorithm’s average classification accuracy reached to 84.73% and88.26% respectively in the experiment of remote sensing scene data set HRSS and UCMLU,and better than the common scene classification algorithm, this initial verified the feasibility of MNCC model.(4)The state-of-the-art precision target recognition algorithm of high resolution remote sensing image based on MNCC models was proposed. To solve the problem of high precision SAR image tank target classification, the hierarchical target recognition mixed classifier based on Deep Spiking Convolution neural Network(DSCN) and Hierarchical Latent Dirichlet Allocation(HLDA) was designed, and target recognition algorithm based on hierarchicalensemble learning(TCR-EL-MNCC) was implemented. Then, TCR-EL-MNCC algorithm was validated on public MSTAR data set. The result show that the overall accuracy of the algorithm in the SAR tank classification is up to 99.82%, which is superior to the other traditional algorithms. Furthermore, to solve the problem of target recognition of small samples and complex objects in ship visible light images, the TCR incremental reinforcement learning algorithm based on object-oriented and multi-scale(TCR-IREL-OOMS) was proposed. Experiments on the HSTCR data set show that the average precision of TCR-IREL-OOMS algorithm is 97.00% for ship recognition, and it is close to the average precision of the SAR tank target recognition on the MSTAR data set. This research shows that using the incremental, reinforcement and ensemble learning mechanism to construct hierarchical computation model of object-oriented and multi-scale, which is consistent with the human cognitive characteristics of remote sensing images. Target recognition algorithms based on the MNCC model can effectively realize complex geographical target information extraction.(5)The remote sensing image information processing and display platform based on component was designed and implemented, and the parallel TCR solution and experiment were proposed. The functional requirements of remote sensing image information processing and display platform were further analyzed, and a software development scheme of remote sensing image processing based on component model was proposed, then the system parallel processing architecture was designed and implemented. The TCR algorithm based on MNCC model was applied in software research and development, and the final performance was analyzed and evaluated. In order to improve the efficiency and practicability of the algorithms,a hybrid heterogeneous parallel TCR algorithm based on MNCC model(PTCR-MP-MNCC)was designed by using parallel computing technology such as multi-machine, multi-core and GPU heterogeneous. PTCR-MP-MNCC algorithm was validated on HSTCR, MSTAR and MNIST data sets, and discussed the impact of the parameters to the algorithm efficiency.Experimental of HSTCR and MSTAR data sets show that the algorithm’s highest speed-up rate reaches to 39.49 and 73.28 respectively. The practice shows that the TCR algorithm based on MNCC model has important value for information extraction of remote sensing image.Above all, intelligent interpretation of remote sensing images involves complex cognitive process and professional knowledge, there are greater difficulties for sophisticated image semantic computing and understanding directly. Since targets of high-resolution remotesensing images has multi-scaled cognitive attributes, it is the first time that hierarchical MNCC model for remote sensing TCR-oriented application was proposed by neurocognitive mechanism. The object-oriented and multi-scaled target recognition algorithms of remote sensing image were designed, which can effectively enhance the accuracy requirement of target recognition. There is higher parallelism in high resolution remote sensing image processing, and effective parallel TCR algorithm can greatly improve the real-time for demands of UAS and ATR system. The TCR-oriented hierarchical MNCC model can effectively meet the task of target detection, scene classification and target recognition of high resolution remote sensing image, and it provides a useful reference in the realization of high resolution remote sensing image intelligent interpretation and complex semantic computing.