Local Invariant Descriptor Applied In Image Recognition

This paper is a part of the research sponsored by the Foundation of National Excellent Doctoral Dissertation of China and the Scientific Research Foundation for the Returned Overseas Chinese Scholars,State Education Ministry.The research is also sponsored by the Shanghai Entry-Exit Inspection and Quarantine Bureau of P R China.The research mainly deals with the algorithms for automatic recognition of the nature cellulose fiber and the classification of shaped fiber.The automated fiber recognition system consists of image collection,image preprocessing, feature extraction and pattern recognition.It is designed to implement the automated fiber recognition.The process is executed without operator's interference.The input of the proposed algorithm in this paper is fiber image.The main works of this paper are pre-processing of fiber image,extracting of local invariant feature and analogy of features.The recognition model is developed.In pre-processing part of this paper,the history and basic method of wavelet transform are introduced.And then,the application of it in pre-processing is introduced in detail.In edge enhancement part,a set of expressions which aims at the high frequency coefficients are mentioned. Experiment shows they can improve the contrast problem of fiber edges and improve the recognition result of whole system. Then,the research background for fiber feature parameter calculation is reviewed briefly.Some related theory and algorithms are studied with comparison.The shortcomings of conventional algorithms are analyzed for the application of feature parameter calculation and fiber recognition.The algorithms are proposed to deal such shortcomings.Feature parameter calculation is an important pre-processing in pattern recognition.The recognition performance is largely depended on the proper selection of the feature parameters.Many local invariant feature extracting algorithms are mentioned in this paper.Furthermore, the scale invariant feature transform is studied,so is its advantage and disadvantage.SIFT is applied in the shaped fiber recognition system combining with the SVM algorithms which is used to analogy the features extracted.The result shows that they work together well.A pattern recognition system should be able to learn from the task and abstracting the decision rules.The theory of support vector machine (SVM) is an effective learning algorithm for small sample prediction. With kernel functions,SVM can constitute a high dimensional model with limited samples.It aims at the minimization rule for structural risk. Thereby it can achieve a good balance between empirical risk and classifier capacity.Therefore,SVM is applied for fiber recognition and the performance of the proposed algorithm is improved prominently.