Research Of Image Classification Methods Based On Deep Learning

Image classification is the important foundation of pattern recognition,machine learn-ing and artificial intelligence.Generally,image classification is composed of three impor-tant steps:region of interest selection,feature extraction and classifier modeling,where feature extraction is the foundation of fulfilling classification tasks.In most scenarios of pattern recognition,proper feature extraction is the key step,influencing the performance of the whole classification system directly.Among these feature extraction methods,deep learning is the most representative one,which learns complex feature representation from massive data directly.Although deep learning has achieved gratifying results,there are still some difficult problems in the previous work:1)How to simplify the structure and pa-rameters of network in deep learning methods on the premise of comparable classification performance?2)How to use deep learning when the dataset is small?In order to solve theses problems,we study the methodology of feature extraction based on the framework of deep learning.Also,the result is verified by image detection and classification tasks,such as medical image classification problem,facial expression recognition and so on.There three parts in this thesis:1)We propose a model called Constrained High Dispersal PCANet(CHDNet).We analyze the classification performance impacted by CHDNet's different components in de-tail.According to PCANet's limitation,we propose and design a nonlinear transformation layer and a multi-scale feature pooling layer aiming at improving the classification perfor-mance.We apply CHDNet to the medical image classification tasks,including automatic human physical function classification based on Kinect and computer-aided tongue diag-nosis,and the performance is satisfying.Moreover,in the case of seriously unbalanced data distribution between positive and negative samples,we prove that the CHDNet can learn more stable feature representation than other feature extraction methods by using weighted LIBLINEAR SVM.2)We propose Locally Linear Embedding Network(LLENet).Using image recon-struction set and our proposed intra-inter class discriminant matrix,we propose an im-proved method LLENet based on LLE algorithm,which is able to be embedded into the process of convolutional kernel learning and constrcution.The purpose of doing so is to increase the discrimination between different types of feature representation after convo-lutional layer.LLENet has a better ability of maintaining the manifold structure of the original image data,and according to the experimental results on facial expression dataset(JAFFE and CK+)and face recognition dataset(Extended Yale B),LLENet is proved to be effective.Feature representation learned by our proposed LLENet is not only better than classical handcrafted methods,but also superior to other counterpart deep learning methods such as CNN and PCANet.3)We study transfer learning and fully connected network(FCNet)in the case of small dataset.We analyze the strategy of transferring deep convolutional neural network model to small dataset.We also visualize feature representations of different classes by using heatmaps,and build a fully connected classifier FCNet.In the newly built classifi-cation framework,feature extraction and classification can be carried out in stages.Exper-imental results on the task of liver fibrosis classification for ultrasound images achieved 93.90%accuracy rate.In summary,this thesis mainly studied the neural network based on the local feature convolution kernel:CHDNet and LLENet,and the processing methods with small dataset.Experimental results verified the effectiveness and practical value of the proposed algo-rithms.