Research On Image Classification Based On Complex Network And Convolutional Neural Network

As an important method for recording effective information,image has high preservation significance and research value.In the medical field,medical images are one of the most important tool for obtaining vital information from patients.Image description is a method which use a set of numbers or symbols to represent an image.The obtained numbers or symbols are called feature vectors.The feature vectors can be used in image classification,image clustering,image retrieval or other works as a criterion to judge the relationship between images.In traditional image classification,the features used are all designed manually,so it is a key step to select an appropriate image description method.In recent years,more and more people pay attention to convolutional neural network.Because it combines neural network technology and deep learning theory,compared with traditional image classification methods,convolutional neural network has better generalization ability and better learning and understanding ability for high-dimensional features.At present,it has been widely used in many fields.The work and innovation of this paper are as follows:(1)For the poor stability of traditional image description methods,this paper proposes an image description method based on complex network model.According to the static statistics of complex networks,this paper establishes the degree matrix of images under different thresholds,and completes the texture description of images by calculating the degree distributions of network nodes in each state.By regarding the Harris corner in the image as the node of the complex network,the initial complete complex network model is established.A series of sub-networks are generated by the dynamic evolution process of the complex network.The shape feature extraction of the image is completed by a series of statistic descriptions,like the degree,joint degree,shortest path length,average path length,clustering coefficient of each sub-network.Combining the texture and shape features of the image,the whole image is described as a whole feature.Image classification experiments show that this method has good robustness,stability and high accuracy comparing to traditional image description methods.(2)To improve the accuracy of convolution network model for image classification without increasing the computational complexity,a method of image depth convolution classification based on complex network model description is proposed.First,the image is described by complex network,computing the degree of difference between each pixel and other pixels,setting a series of thresholds,and considering that the two pixels whose encounter degree is lower than the threshold are related,that is,the two nodes in the complex network model are interconnected,this provide degree matrices from image RGB components in complex network model under different thresholds.Then,based on the description of image degree matrices,the feature vectors are obtained by introducing the degree matrices into a set of deep convolution networks connected in parallel.Finally,the images are classified according to the obtained feature vectors.Classification experiments show that compared with the original convolution neural network model,this method can increase the computational complexity and has a higher accuracy.