Research And Application On Image Recognition Based On Sparse Coding

With the development of compression perception theory,a large number of algorithms based on sparse representation of image denoising,image fusion and image super-resolution are proposed and widely used.However,the research of the application of sparse representation classification algorithm to all aspects of image recognition field still remain to be perfected.The Sparse representation classification algorithm can be simply divided into three stages: feature selection and extraction,over-complete dictionary construction,sparse representation and classification.In this paper,the three stages of sparse representation classification algorithm are analyzed in depth and the improved algorithm is put forward.For the selection and extraction of features,this paper proposes a discriminant Gabor multi-component cooperative representation algorithm.For the construction of over-complete dictionary,this paper combine the IFDDL algorithm and the F-FDDL algorithm,then proposes a fast incoherent fisher discrimination dictionary learning algorithm.For the sparse representation of the objective function itself,inspired by the ProCRC algorithm and presents a robust sparse representation classification algorithm based on probability analysis.In addition to the theoretical research and innovation in depth,this paper analyzes the problems existing in the current cloth printing and dyeing detection process,and applies the sparse representation image recognition algorithm to the automatic detection of printing and dyeing cloth.The research and application of sparse representation image recognition are as follows:(1)Based on the in-depth study of sparse representation classification algorithm,inspired by ProCRC algorithm,an algorithm for explaining sparse representation is proposed by using probability analysis,and an improved robust sparse representation classifier is derived.The robust sparse representation classification algorithm based on probabilistic analysis is to analyze the feasibility of probability derivation from the perspective of probability subspace.The optimization result of final probability analysis is exactly the same as sparse representation algorithm,perfects the sparse representation Classification algorithm theory.(2)Because of the different discriminating ability of Gabor features,this paper analyzes the discriminating ability between different components,and proposes the method of using interclass variance and interclass variance ratio to judge the discriminating ability.In this paper,we use some strategies to eliminate the characteristic components with poor discriminating ability,and use the remaining feature components to classify and identify them.In the case of voting to determine the final classification attribution,this paper calculates the weight of the feature component for classification by calculating the neighbors of each k sample,and the final result is determined by weighted voting method,which makes the classification more accurate.Due to the particularity of face recognition in image recognition,this paper uses the cooperative representation classifier to replace the sparse representation classifier,which is more accurate and robust in the case of uncertainties such as illumination,expression,attitude and occlusion.Indicating that the classification algorithm avoids the solution of the 1L norm optimization,so the computational efficiency is greatly improved.(3)In order to solve the Fisher discrimination dictionary learning stage,the coherence between dictionary atoms and the time consuming problem in sparse coding stage are not taken into account.This paper combine the IFDDL and the E-FDDL,propose the IF-FDDL algorithm.In the dictionary update phase,by introducing irrelevant terms,the dictionary atoms are more independent,and the representation of the target sample is stronger.In sparse coding phase,the upper bound of the objective function is solved and the NAGDA algorithm is used to optimize the solution to achieve the purpose of reducing the computational complexity.Experiments show that the fast incoherent fisher discrimination dictionary learning algorithm(IF-FDDL)proposed in this paper is not only superior to other supervisory dictionary learning algorithms in terms of ability and recognition rate,but also improves the efficiency and convergence speed.(4)The sparse representation classification algorithm is applied to the automatic detection of cloth printing and dyeing,turning the cloth printing and dyeing detection problem into image classification problem.In this paper,we design a printing and dyeing cloth automatic detection system,rotating the original normal image to obtain more abnormal samples to detect the situation of printing and dyeing cloth.When the image to be detected is divided into the non-state category,indicating that the cloth is crooked during the printing and dyeing process,need to be adjusted in time.Several improved algorithms proposed in this paper are applied to the detection of cloth printing and dyeing.The experiment proves the rationality and the strong detection ability of the system design.