Study On Object Recognition Based On Contour Feature

Object recognition is always one of the hot topics in the field of image processing, pattern recognition and computer vision, and it is widely used in every area of the daily lives, industrial applications and military activities. Contour is one of the apporaches to describe the object and plays an important role in object matching and recognition. From the angle of human being’s cognitive things, the description and evaluation of object features are two key parts for object recognition. Therefore, it is very important to obtain the optimal object matching that how to effectively describe contour features and properly evaluate the similarity between different features.This dissertation focuses on the extraction, description and evaluation of contour features, and studies the effect of contour feature applied to the following aspects:the smoothing of noisy contour, the alignment of the contour starting-point and the recognition of object contour. As for the object contour, current contour smoothing algorithms cannot achieve a better tradeoff between feature preservation and noise suppression. And, existed starting-point alignment algorithms cannot be applied to the affine-transformed objects or are of high computational complexity. Moreover, state of art of the affine object recognition algorithms cannot meet the requirement of real-time applications. Furthermore, many occluded object recognition algorithms did not consider the influence on object recognition of the relationship between local contour features and the whole object. The main content and researching results of this dissertation are listed as follows:1. Constructing a kind of framework of describing contour features. Within this framework, current contour feature description methods are summarized and classified into three levels:global features, local features and primitive features. Further, those methods are analyzed in three aspects:reliability, computational complexity and the application range. And it is beneficial for choosing the optimal manner of describing contour feature in the cases of various applications.2. Proposing the affined object recognition algorithm based on segmented contour-smoothing. To suppress the noise and obtain the accurate feature description, the noisy contour is smoothed in advance. The contour is divided by curvature into two categories:feature zone and non-feature zone. By smoothing contour parts located in those zones by the Gaussian filters with different variance, the tradeoff between noise suppression and feature preservation is well achieved. After the contour-smoothing processing, the affined contour can be recognized by utilizing the affine invariant moment and minimum distance classifier. Simulation results testify the improvement by the segmented contour smoothing on recognition results.3. Proposing the fast affined object recognition algorithm based on the alignment of contour starting-point. To decrease the high computational complexity of existed starting-point alignment algorithms, the algorithm based on the joint affine invariant arc-length is proposed and aligns the starting-point of contour successfully and efficiently. Then, the cascaded affine invariant function with translation invariance is constructed by using the multi-level approximation coefficients of discrete wavelet transform. Compared with similar recognition algorithms, the proposed algorithm greatly reduces the overall computational complexity. Simulation results show that the proposed alignment algorithm and recognition algorithm are robust to the noise and meet the requirement of real-time application.4. Proposing the partially occluded object recognition algorithm based on contour features. Starting with the analysis of contour feature, the contour-splitting scheme based on the distribution of local curvature is proposed and builds the contour fragment database, which describe integrally local contour features. Then, the partially occluded object recognition algorithm based on features description integrity is proposed. The shape context distance is introduced to measure the similarity between contour features. And, the matching degree between occluded objects is measured by the combination of shape context distance and the reliability of contour fragments Simulations show that the proposed recognition algorithm needs less training samples and achieves the goal efficiently and accurately. Further, to strengthen the stability of local feature description under various occluded cases, contour fragments obtained by the contour-splitting scheme are classified into different categories with the help of object structure. Then the multi-level description model of local contour features is built. Quoted from the philosophical issue that "the relation between the wholeness and the part", this dissertation explores in depth the relation between the whole contour and contour fragments. And two parameters, importance and partiality, are presented to evaluate contour fragments. On that basis, the partially occluded object recognition algorithm based on multi-level description and evaluation of contour features is proposed. To measure the matching degree, the two evaluation parameters and the similarity between local contour features are combined to introduce the weighted partial similarity. Compared with current similarity measurements, the weighted partial similarity reflects properly the relation of the similarity between the local contour features and that between the corresponding whole contours. So, the partially occluded object recognition algorithm based on multi-level description and evaluation of local contour features is proposed. Simulations show that the proposed recognition algorithm achieves the stable recognition results under various occluded cases.