Research On Image Local Invariant Feature And Its Application

Image feature extraction is an important research content in the fields of imageprocessing, pattern recognition and computer vision. Local invariant feature, due to itsinvariance for image translation, rotation, scale, illumination and viewpoint transformation,has now developed into one of the most important research areas in computer vision andbeen a research hotspot both at home and abroad. This paper analyses the scale space theory,studies some of the existing extraction and matching methods of the local invariant feature,and accordingly improves the existing problems of these methods, and the new improvedmethod applies image matching, image registration and image recognition, and obtains abetter result.Firstly, this paper studies the development history and mathematics theory basis ofscale space, analyzes the property of Gauss scale space, and reveals the origin of localinvariant feature invariance. Through the analysis of scale detection mechanism ofone-dimensional signal in the scale space, this paper gains the general process of scaleselection, and validates the advantage of using scale space as the data representation oflocal invariant feature, makes clear the role and significance of scale normalization of scalespace, analyzes the properties of the partial derivative matrix of scale space and Hessian,and points out its important role in the estimation of local image properties. This paperintroduces the basic operator of spot detection, which provides a theoretical basis for thesubsequent chapters.Then this paper studies the scale invariant feature transform algorithm. To address theproblems of the complex process of SIFT detector and the high dimension of the descriptor,this paper puts forward a fast matching measure method based on SIFT with optimizedsimilarity measure.In the feature vector matching process, this paper uses the linearcombination of the chessboard distance and block distance instead of Euclidean distance forimage matching, aiming to reduce the time complexity of the similarity measure.Theexperimental results show that, there is no change in feature points and feature pairsobtained through the new similarity measure, but the matching time has been reduced.Therefore, under the condition of ensuring the unchanging of the robustness of the SIFTalgorithm, the matching time efficiency has been effectively improves.Then this paper studies the SURF algorithm. The SURF algorithm uses the BBF feature point to search in the feature matching process, which has the problems of takinglong time and damage of precision.To solve these problems, this paper proposes to userandom K-D tree algorithm to search in the feature points matching process, and thismethod guarantees the search accuracy, improves the search speed, eliminates the errormatching points through the RANSAC algorithm, and finally completes the registration byestimating the spatial geometry transform parameter of the two images according to theremaining matching points and pairs.Experiments show that this algorithm is a fast robustimage registration method.Finally this paper studies the variation of Gabor algorithm. This paper proposes aface recognition based on Gabor wavelet transform. Based on excellent spatial locality,frequency and direction selectivity of Gabor wavelet transform, it can capture the localfeatures of the face images at different frequencies and directions. The Gabor dimension istoo high, not only needing a large storage space, but also consuming too much time for therecognition process, this paper presents the method of block used in the2DPCA, which notonly achieves the dimension reduction of Gabor feature, also retains the two-dimensionalinformation of image features. Finally the experiments on ORL and JAFFE face imagedatabase prove that this method can obtain better result of face recognition.