The Research And Application Of Deep Learning In Image Recognition

Precise recognition for image Has very important research significance, imagerecognition technology is widely used in Medicine, space military,industry and agriculture.As now most method of image recognition Used artificial feature extraction which Not onlylaborious, but also difficult to extract. Deep Learning is a kind of unsupervised learning, Inthe learning process we need not know the values of samples, The whole process can alsoextracted good characteristics without human participation. in recent years, The deeplearning used in image recognition become the hot research topic in the field of imagerecognition, Has achieved good effect, and have a broad space for research.In this paper, we based on the study on the theory of image recognition analyzes thedeep learning the basic models and methods,then do experiment on some image data sets.Given deep learning more for large sample set, we improved a algorithm proposed to use itinto small sample set, the work can be described as follows:(1) Analysis basic principles of the convolution neural networks (CNNs), introducethe training process and model structure of it. The convolution layer can make the originalsignal enhancement, and reduce noise as well as improve signal-to-noise ratio byconvolution operation, use the model into the handwriting data set MNIST, compared toother classical algorithms, analyze their advantages and disadvantages about time andrecognition rate.(2) aim at the inadequacies of the convolution neural network, Analysis the basicprinciples, training process and model structure of deep belief networks (DBNs). Thestratified training mechanism of DBNs greatly reduces the difficulty and reduces the trainingtime of it. We use Softmax classification system as the classifier, the use this model doexperiment on MNIST datasets, compared to convolution neural network, mainly in therecognition rate and time-consuming,which can be proved that DBNs has the sameidentifying rate of CNNs, but the elapsed time is greatly reduced, and then analyze thereasons; addition, use the model into the CIFAR-10databases, compared to the otheralgorithms.(3) aim at the deep belief networks algorithm is only applicable to large data sets, raisea improved algorithm of deep belief network aim at small sample set. Before the pre-training, down-sampling of samples, after the training, in the parameter fine-tuning phase, use the dropout ideas. Down-sampling and dropout can effectively prevent of overfitting,the improved system is applied on a MNIST subset and ORL datasets, experiments showthat this system indeed prevent over-fitting, the algorithm has good improvements in therecognition rate and time-consuming both.