Research And Application On Image Recognition Based On Machine Learning

In recent years,image data has shown the explosive growth of network with the rapid development of Internet and Mobile Internet.Image data is widely used as the carrier of information exchange,because image data is simple and intuitive and contains abundant information.Image recognition based on the image content can start from the image itself,first extract the salient features from the image,and then identify the image according to the difference of the features,which has a good recognition effect.Machine learning has a very important position in artificial intelligence.Machine learning learns with data through algorithms and then makes prediction and decision,which is the key to realize artificial intelligence.Image recognition with machine learning is based on the image content.Machine learning is based on image data itself,extract the bottom of the image features from the image data,and then establish the relationship between underlying characteristics and high-level image semantics,which can achieve the image recognition.Image recognition is the basis of machine vision.The image content is varied,so the machine vision is different.For computers,the underlying features of images are not directly related to the high-level image semantics,so solving the "semantic gap”is the important and difficult problem of image recognition.Machine learning has two stages of the shallow and the deep learning,and the experts and scholars put forward many models of the algorithm,which has made much progress in the aspect of image recognition,speech recognition and artificial intelligence.Firstly,this paper introduces the support vector machine model which is based on statistical theory.The support vector machine model is used to train classifiers according to the characteristics of human extraction and achieve the goal of image recognition by classifiers.For the extraction of the underlying features,this paper introduces the extraction methods of color,shape and texture.For the principle of the support vector machine,this paper introduces the kernel function and penalty factor.For classifier construction,this paper introduces the shortcomings and improvement methods of binary tree classifiers.For the image feature,this paper introduces the shortcoming of single feature and the fusion method of multiple features.Through the comparison experiment,the improved method of binary tree classifier and the method of the fusion feature can improve the accuracy of image recognition.This paper also introduces artificial neural network model which is based on the simulation of biological neural network.Artificial neural network can extract the features from the underlying pixels to train network model and realize the goal of image recognition through the output layer.For the principle of artificial neural network,this paper introduces the training of the network model through the reverse propagation algorithm.For the structure of the network model,this paper has studied the design of the node number and layer number.For the initialized of weights,this paper introduces the shortcomings of the initialization method and the improvement method.Through the comparison experiment,the new design model structure and weight initialization can improve the accuracy of image recognition.Finally,this paper introduces the convolutional neural networks which is based on the simulation of biological vision system.The convolutional neural network is able to handle the two-dimensional image data,so it is very suitable for image recognition.For the principle of convolutional neural network,this paper introduces the connection algorithm of the convolution and pooling layer.For network model structure,this paper introduces the local connections and the sharing weights of the convolution and pooling layer,the Dropout method of the full connection layer,the using of support vector machine classifier model as the output layer,and the using of the new designed weights initialization.For the selection of the activation functions,this paper introduces the shortcomings of the commonly used activation functions and the fusion function based on the S function and R function.Through the comparison experiment,the improved network model and the improved activation function can improve the accuracy of image recognition.