Research On Image Recognition Algorithms Based On Structure And Dictionary Learning

Image recognition is one of the key research highlights in the field of artificial intelligence.The core algorithm of image recognition aims at how to make the computer extract effective visual feature information from image data set and quickly recognize its specific categories,which is also the most urgent,difficult and important problem in the application of image recognition.Based on the theory of dictionary learning and structured representation learning,this paper aims to solve the problem of small sample size of data,enhance the discrimination of recognition model and boost the learning efficiency.In this paper,a series of learning models are proposed and applied to many image recognition tasks.A large number of experiments show that the three algorithms can improve the recognition accuracy and speed to a certain extent,and promote the performance of the image recognition system.Specifically,the main work of this paper is as follows.In view of the difference between space-frequency domain and traditional dictionary learning methods,a joint dictionary learning method based on space-frequency domain cooperative representation is proposed.The proposed method considers both the inherent space domain and the Fourier-transformed frequency domain of the data set,and provides two key complementary strategies of data domain complementary and classification algorithm complementary.In this method,both the original data set and the Fourier transformed data set are applied to perform dictionary learning,which makes the data complementary in the spatial domain and frequency domain,and the dictionary learning and collaborative representation are combined to classify.The proposed method is complementary to each other in two aspects,which improves the discrimination ability of dictionary learning.In order to solve the problem of image preprocessing with geometric symmetry and the problem of small samples size of data,a composite dictionary learning method based on virtual samples is proposed.According to the symmetry of image geometry,this method generates new virtual image samples automatically and efficiently to expand the data set and provide more information for recognition.The proposed method utilizes the adaptive weighted fusion strategy to combine the original dictionary with the virtual dictionary and construct the classification model based on the composite dictionary learning.The adaptive weight can well offset the randomness of pixel exchange in the process of virtual sample generation.Experiments on several different open face datasets show the feasibility and effectiveness of this method.In light of the lack of discrimination of traditional least square regression algorithm,this paper proposes a classification method of least square error with adaptive graph structure.An adaptive probabilistic graph structure is constructed to enhance the compactness and discrimination of the data and to ensure that the internal structure of the samples is fully mined in the mapping subspace.In addition,this paper designs an effective optimization algorithm,and explains the optimization problem from the theoretical and experimental perspectives.This method is evaluated on six datasets,including face recognition,character classification,object recognition and scene classification tasks.In comparison with various classification based algorithms,it was verified that the proposed method has better recognition performance.