Research On Image Representation And Recognition Method Based On Basal Structure

The scale of image data on the network is getting larger and larger,and the accompanying requirements of image,such as storage,transmission and recognition,are also facing greater challenges.The research on image recognition and image representation has become the key point that needs to be solved to cope with new challenges.At present,structural features,texture features,subspace,machine learning and other methods used in image recognition.Linear class methods such as matrix decomposition,sparse coding,linear discriminant analysis and nonlinear class methods such as kernel learning,manifold learning and machine learning and other methods used in image representation.All these methods have some problems,such as complicated calculation and analysis,process abstraction,difficulty in optimization,and high requirements on training data.In this thesis,a research method in the field of signal processing is used.Similar to the idea of Fourier series,the method represents a complex signal as a linear combination of a series of simple signals.The method of image representation and recognition based on basal structures obtained by using unspecific data sets is studied.The main research contents are as follows:Firstly,In view of the image representation method,this thesis proposes an image basal structure decomposition method based on feature extraction and uses the decomposed basal structure to represent the image.The algorithm first randomly takes several image blocks of the same size as training samples from the image data,then uses feature extraction algorithm to decompose the basal structure,and then constructs the projection and reconstruction model of the image in the basal structure space,so that the decomposed basal structure can represent the image.Experiments show that the original image can be represented well by using a certain number of basal structures.Secondly,for the decomposition method of the basal structure,a quantitative analysis method based on the PSNR is proposed.By calculating the PSNR between the original image and the reconstructed image using the basal structure,the properties of the decomposed basal structure are quantitatively analyzed,including the contentindependence,the influence of size change,and the one-to-one correspondence.Experiments show that the decomposed basal structure can be independent of the specific type of image data and can represent any type of image.Thirdly,combining with neural network,the image recognition method using basal structure is studied.The experiment uses the public image set,such as MNIST character set,etc.The algorithm projects the data in the basal structure space for dimensionality reduction and uses the feature matrix obtained by projection to input the neural network for image recognition.The experiment shows that the training process of neural network can be optimized by reducing the dimension using the basal structure.Fourthly,combining with convolutional neural network,the algorithm uses the basal structure to initialize the convolutional layer of the neural network.The experiment first compares several common convolutional layer initialization methods,Then analyzes the influence of initializing different convolutional layers in the same network using the basal structure initializing method,and then analyzes the influence of reducing the depth of convolution kernel on different initialization methods.The experiment shows that compared with the common initialization methods,the basal structure initialization method has better effect,faster convergence speed,and less influence from the change of depth value of convolution kernel in the initial iteration.This method can simplify the network structure.The image representation method proposed in this thesis can represent any type of image.The decomposed basal structure does not depend on specific data set and the content is not abstract,which has physical significance.The initial iteration accuracy of the image recognition method combined with neural network is higher and the convergence is faster.At the same time,initializing the convolutional layer with the basal structure is helpful to deepen the understanding of image convolution operation.