Research On Image Classification Based On Deep Learning

Deep learning has attracted attention from the field of machine learning since it was proposed.Deep Learning is proposed to solve the feature extraction problem in machine learning.Comparing with the shallow learning models which extract features by hand,the deep learning models can extract the data characteristics layer by layer,making features more obvious and easier to classify.The most notable feature of deep learning is that it can split a complex problem into several simple problems.Firstly these simple problems are solved one by one,and then complex problems.Deep Belief Networks are widely-used deep learning models which have many hidden layers.Deep Belief Networks have better capability of learning deep features layer by layer.The training process of Deep Belief Networks is divided into two parts,pre-training and fine-tuning.The training process is supervised in pre-training process,while unsupervised in fine-tuning process.Deep Belief Networks can be used in image classification.We introduce the Gaussian-Bernoulli Deep Belief Networks to extend the value of input nodes from binary to Gaussian.And then we apply Deep Belief Networks,Gaussian-Bernoulli Deep Belief Networks and other classification algorithms to classify images to evaluate their performs.We propose a node selection method based on similarity to solve the selection problem of Deep Belief Networks.The node selection methods of Deep Belief Networks are similar with the traditional artificial neural networks methods.The number of nodes can not be changed in training process and not all nodes are fully utilized in original Deep Belief Networks.In this paper,the selection algorithm based on similarity is used to re-use the nodes,and it is able to make full use of the hidden layer nodes to improve the network accuracy under the condition that the model has been set up.Experiments show that the node selection algorithm proposed has better performance.We propose a new fine-tuning algorithm of Deep Belief Networks based on artificial bee colony algorithm.We use the global search ability of artificial bee colony to improve the back propagation performance.Through the alternant iterations of back propagation algorithm and the artificial bee colony algorithm,the model gets out of the local extreme point to search the global optimal point.We verify the performance through two databases of different scales.The experiment results show that the algorithm proposed in this paper can obtain higher classification accuracy for small-scale model and small-scale database.Finally,we introduce the local binary pattern to overcome the shortcomings of Deep Belief Networks.Deep Belief Networks have no ability to extract the local features of image.At the beginning,the local features of input data are extracted by the local binary pattern.Then we use the local features to train Deep Belief Networks.The preprocessing of features make the classification of Deep Belief Networks more accurate.