The Research On Shape Feature Extraction Based On Improved Chord Context

The feature extraction and corresponding invariability of similarity measurement algorithms are becoming the hotspot of intelligent vision technology research in recent years. It is considerable to detect objects' shape that is linked to objects' functional information and identity in pattern recognition and similarity retrieval. A large number of feature extraction and similarity measure methods using shape descriptors have been proposed in literature, moreover, various applications are likely to use shape features. The geometric invariance and deformation resistibility are important issues for feature extraction and similarity measure methods using shape descriptors.The product appearance detection system is the case of the application of machine vision technology in industrial production. The research of quickly seeking object in complicated environment and determining whether the appearance of the object meets the expected standards accurately is a challenging task for domestic and international engineering area with profound theoretical significance and widespread application value. The affine invariant features extraction algorithms in this thesis can be applied to product appearance detection system which can greatly expand the application field of the system. The application field of the algorithms in the thesis includes military, aerospace, environment, medical treatment, biotechnology, material, finance, traffic, family and so on.For a given shape, the chord context describes a frequency distribution of chord lengths with different orientations. The main advantages of chord context for retrieving images are summarized in the following two points:the flexibility and the accuracy. The experiments demonstrated that the chord context is invariant to several common images transforms such as scaling, rotation, boundary perturbations minor partial occultation and non-rigid deformation. The ability of the algorithm resistant to affine transform is insufficient. To enhance the resistance capability for affine transform and projection transform, two improved chord context methods are proposed in the thesis.1. We proposed a relative-chord context method based on chord context. We improve the feature extraction algorithm by using the relative-chord and normalize the parallel intervals with vertical direction-chord. The proposed method of relative-chord context is proved to be invariant under affine transformations. We use vertical direction-chord intervals instead of the fixed intervals in the chord context which improved the algorithm's performance further.2. In this thesis, a chord location matrix is proposed. The location information of chord distribution in different direction and the sorting information of different chord in the same location are considered as characteristic vectors. The chord location information represents the characteristics of the object shape in the proposed method. The features extracted by the proposed method in the thesis are invariant to affine transform.We analyze the properties of chord context firstly and propose the improved ideas for the algorithm. The algorithm normalizes the chord length only and doesn't consider intervals between parallel chords affected by affine transform. We first perfect its theoretical structure and excavate the inherent information of the image further and then we propose the relative-chord context and chord location matrix. Experimental results show that our algorithms can describe the features of different objects more accurately which have robustness for affine transform and they can play better performance in affine object recognition.