Research On Performance Evaluation For Small Target Detection And Recognition Technology

The detection and recognition of small target is an important component and research direction of automatic target recognition. The problem of detection and tracking of dim target comes from application of long-range surveillance system. In recent years, researchers pay more and more attention on the detection work of dim target in infrared (IR) image sequence. It is an increasingly impending requirement to evaluate the existing or developing algorithms and systems of small target detection and recognition, and to study and develop new algorithms for small target detection by way of performance evaluation. So the performance evaluation is crucial to study and develop new algorithms and systems of detecting and recognizing small targets, to look for the flexibility boundary and to improve the system's performance.The purpose of the paper is to present new theories, methods and supporting environment models of performance evaluation for small target detection algorithms and their systems. The work of the paper is to provide theory foundation and experience support for developing the algorithms and systems of the National Natural Science Foundation of PR China and the National defense key advanced research project of PR China. The work of the paper focuses on three aspects. First of them is the research work of the basic performance evaluation framework of algorithm systems of target detection and recongintion. The second is the research work of new performance evaluation methods for small target detection and recognition, including mechanism analysis method, performance evaluation based on experimental design methodology, performance estimation technology of the full processing flowchart, which are used to estimate the existion algorithms or improve and develop new algorithms. The last is the research work of the theoretic models of supporting algorithm development and performance evaluation of small target detection and recognition, and its implementation technologies are also discussed. The work of the paper provides technologic support of research and development of small target detection, recognition.By analyzing the evolution of performance evaluation of target recognition algorithm, the viewpoint is proposed that the process of the algorithm development and the process of performance estimation should be unified into one framework. The theories and methods of performance evaluation are discussed, the system identification model is introduced into the performance modeling of algorithm systems, and the multivariate data analysis approaches are developed. The organization and software supporting environment for the performance evaluation are discussed.The performance estimation model of the full processing flowchart for analyzing algorithm systems is made, which is applied to analyze the effects of image restoration on small target acquisition from the turbulence-degraded images is presented. The effect of the restoration operation on infared point target detection is analyzed. The broad basic problems during performance evaluation research are addressed: a) how to investigate the system as a whole, and b) how to evaluate the performance on realistic and varied data to determine and compare both the capabilities and limits of the existing systems. The evaluation experimental results demonstrate the validity of the proposed model and methods.The new mechanism analysis method and the evaluation technology based on scientific experimental design are presented, and the active effect of the mechanism analysis method on developing novel algorithms is demonstrated. A new model of IR dim small target image– a 1-rank adjacent space distribution model is made. The optimization problem of inverse problem is applied to the domain of small target detection, and the ill-pose problem of conventional clutter background prediction methods is analyzed. Based on this, a filtering framework using regularization technology is presented and a novel fast filtering method with'clutter-removal target-preserving'regularization is proposed. Detailed theoretical analyses and experimental results show that this new method provides good filtering results and robust adaptability of small IR target detection, moreover, its little computing complexity and simple computing structure are suitable to be implemented in real-time system.In order to reduce the complexity of evaluating small target detection and recognition, and to enhance the efficiency of algorithm development and estimation, a new supporting environment concept model for algorithm development and performance evaluation is presented. Based on expert knowledge in software programming domain and automatic target recognition domain, a new knowledge-based computing model using expert system is proposed, and its representation method and implementation technologies of the expert system are discussed. The practice shows that the development based on this new model can improve modeling of a novel algorithm prototype and reduce the developing process to get a higher efficiency; and it aso make the development and management more easier and efficient, which improve the productivity of algorithms and systems.