Image Recognition Algorithm Based On Feature Coding And Deep Learning

Image recognition is an important research topic in computer vision and pattern recognition,and this technology is also an important tool to implement the intelligent society.The common image recognition model can be divided into four important steps: image feature extraction,image feature transformation,feature pooling and classifier training.Feature coding method is an important image recognition model and is widely used in image classification,object retrieval and video action recognition.The outstanding feature coding methods can effectively improve the performance of image recognition.The feature coding model is mainly used for image feature transformation or the final classifier in the above four steps.Although the feature coding image recognition model has demonstrated the excellent classification performance,in the fields of the discriminative shift-invariant feature coding,the tradeoffs between time and performance,and the end-to-end feature coding.There are still have some problems that are not perfect.Aiming at the shortcomings of the existing models,this paper starts from the two aspects of classifier and image feature transformation,and proposes novel and effective feature coding model to improve image recognition performance.The innovative results of this paper are as follows:1.For the traditional convolutional sparse coding model,which is unsupervised,it is not suitable for classification task.This paper presents a convolutional sparse coding classifier by supervised training.The proposed convolutional sparse coding classification model combines the convolutional sparse coding and the related classification strategy,and the convolutional filter with the invariant and class information is obtained.This paper presents the optimization problem of the convolutional sparse coding classification model,the corresponding optimization algorithm and the classification strategy.The convolutional dictionary obtained by supervised learning is more expressive than that in the sparse representation classifier.This paper gives the image recognition experiments in MNIST dataset and CIFAR10 dataset,the experimental results demonstrate that the proposed model improves 2-3% performance than the sparse representation classifier.2.The existing sparse representation and dictionary learning classification algorithms mostly need to solve the time-consuming l0 norm or the l1 norm minimization problem.In this paper,a dictionary pair learning method based on differentiable support vector function is proposed.This model can effectively reduce the training and testing time by using the projection dictionary to solve the coding coefficient.In the training stage,the proposed model jointly trains a synthesis dictionary,an analysis dictionary and a support vector discrimination term which can enhance the discrimination of the representation coefficients.In the test stage,the proposed model uses the reconstruction residual,the projective discrimination term and the support vector function to decide the label of the test sample.The image recognition experiments prove that the proposed method has higher image recognition rate and lower time complexity than the dictionary learning classification algorithm based on l0 norm or l1 norm.In the larger dataset,compared with the original dictionary pair learning model,the proposed model improves the recognition rate by 3%.3.For the sparse coding spatial pyramid matching(Sc SPM)model which can only learn the unsupervised dictionary,in this paper,a sparse coding network with spatial pyramid pooling layer is proposed.For end-to-end training the sparse coding model,the network will treat the sparse minimization problem as a recursive network layer,and the sparse coding network layer and spatial pyramid pooling layer and a deep convolutional neural network are trained together.Through such supervised training,the learned dictionary will contain the final label information.The sparse coding network can obtain more representative sparse codes by the convolutional neural network features.Compared with the Sc SPM model,the proposed model improves the recognition rate by 4-5%.4.In this paper,a localized and second order VLAD coding network is proposed.First of all,a localized and second order VLAD coding method is proposed.Next,we derive the back propagation functions of all the newly designed layers.at last,we extend this newly designed feature coding method to an end-to-end feature coding network,this new network can be jointly trained with a deep neural network for image recognition.Furthermore,this paper proposes a multi path feature coding network(M-LSO-VLADNet)for aggregating multi path feature codings.The experimental results demonstrate that the proposed network has the better performance than the Net VLAD and other convolutional neural networks.