Research On Feature Analysis And Image Matching Method For Complex Scenes

Image matching is a technique for evaluating the similarity and establishing the corresponding relations between two or more images acquired under different conditions.This technology has been widely used in many fields,such as national defense,military,aerospace,medical diagnosis and so on.In the field of weapon equipment and navigation,the reliability of image matching is usually required to be high.Before the real-time matching,it is necessary to determine the comprehensive performance metrics of local features about their stability,invariance and discriminative ability under various scenes.It is also necessary to analyze specific scenes and determine the performance parameters,such as correct rate and false alarm rate for the key-point feature matching in advance to accurately locate and match the corresponding key points in different scenes and improve the image matching speed in the real-time online stage and reduce the incorrect matching rate.However,in complex scenes,the changes in imaging conditions such as illumination,scale and other environmental factors such as camouflage,similar background interference seriously pose challenges to both the stability and discriminative ability of feature and the image matching performance.The existing image matching methods still face great challenges in terms of matching precision,recall rate and efficiency.To improve the matching performance of local features in complex scenes with repetitive structures,this dissertation broadens the scope of analysis from both local and global directions based on the research of feature stability and saliency.By studying the method of feature saliency analysis and the method of comprehensive feature performance evaluation,it overcomes the problems of high incorrect matching rate,low recall rate and low utilization rate of similar features from repetitive structure in complex scenes.The main innovations of this dissertation are as follows:(1)The feature analysis method under complex scenes are studied and an efficient feature matching method based on it is proposed.To solve the problem of high incorrect matching rate caused by repetitive structures in complex scenes,a multi-region semi-random sampling consistency matching method based on local feature analysis is proposed.In the process of feature analysis,the influence of scene changes on the stability of local features is studied,and the stable features can be selected.An evaluation criterion is defined for stable features to describe their discriminative ability.According to such criterion,the stable features can be classified into two categories,including salient features and non-salient features(repetitive features).On the basis,the salient features with high discriminative ability are selected from the stable feature matching set for matching.To establish the homography model and improve the performance of the algorithm,the reliable matching samples with reasonable spatially distribution are extracted from the salient feature matching set,which has high proportion of inlier rate,by using the proposed multi-regional semi-random sampling strategy.The experimental results show that the proposed multi-regional semi-random sampling matching method can effectively improve the matching accuracy and efficiency in complex scenes.(2)The feature regional saliency analysis method and the feature matching method based on region devision are studied.In order to improve the utilization rate of repetitive features,a feature matching method based on saliency region analysis is proposed in this dissertation.The method of region devision is used to reduce the interaction between the repetitive features and enhance the feature saliency.Following the principle of feature saliency,the region salient features with strong discriminative ability can be selected for matching and the reliable seed pairs can be found.A progressive strategy of Delaunay triangulation is used to suppress the matching ambiguities caused by similar features.By the way of gradually enhancing the feature saliency,the utilization rate of repetitive features can be improved.The recall rate of feature matching can be improved without reducing the matching accuracy.The experimental results show that the reliable matching set of salient features can be selected as the seed point pairs for the precise devision of regions and the reduction of matching ambiguity based on the regional saliency analysis of features.The matching utilization ratio of similar features can be improved.The high matching accuracy and recall rate can also be obtained in the complex scenes with repetitive features.(3)The strategy of local feature analysis and comprehensive performance evaluation is studied.In order to determine the matching performance of local features under various scenes before real-time matching,the invariance,stability and saliency of features and the compatibility between feature matching points and projection models are comprehensively analyzed and evaluated.The problem of image matching is formulated as structured output prediction.The feature analysis and evaluation is performed by combining the feature saliency analysis and the structured weight training.A single evaluation score is trained to describe the comprehensive performance of features.The experimental results show that the training feature scoring criterion has the ability to comprehensively evaluate the feature performance by using the feature analysis method based on structure prediction.It can also be used for feature selection,model consistency verification and the performance improvement of the matching algorithm effectively.