Research On Image Classification Method Based On Deep Learning

With the rapid development of Internet technology,people have entered the era of big data information,and images have become a very important form of information carrier in people's lives.As image databases become larger and larger,how to quickly extract visual information by classifying image data has gradually become an important research direction in the field of image processing and computer vision.Image classification has important theoretical significance and practical application value in the fields of image retrieval,image recognition,intelligent security,medical diagnosis and so on.The traditional image classification method requires manual design to extract image features,the calculation amount is large,and the classification accuracy rate is low,which has become unable to meet the needs of the times.Deep learning has been widely active on the stage of image classification in recent years,it has strong character representation ability and generalization ability,and it can learn image data independently and extract features automatically.Therefore,this paper proposes an image classification method based on deep learning,and further studies the application of convolutional neural networks in image classification in deep learning models.In this paper,two design schemes are proposed to improve the classification accuracy of deep learning model,reduce the size of network and prevent model over-fitting.The first is an image classification algorithm based on improved convolutional neural networks.A 7-layer convolutional neural network is designed to improve and optimize the classical convolutional neural network AlexNet algorithm from both network architecture and internal structure,by adjusting activation function and pooling mode,and learning the feasibility of using extreme learning machine as convolution neural network classifier.The CIFAR-10 and CIFAR-100 two compleximage datasets were used to train,test and optimize the network parameters of the improved convolutional neural network.The experimental results show that the improved model accuracy is improved by 12.87% and 6.85% respectively.Secondly,the deep convolutional neural network is studied and a natural scene image classification method based on Mobilnet V2 networks is proposed.The ResNet50 and Mobilnet V2 networks are pre-trained with large-scale natural data set ImageNet.The trained model is transferred to the new network,and then a self-built natural scene data set with 9 categories is retrained and fine-tuned with the new network.The experiments show that the size of Mobilenet V2 network model is 11 times smaller than ResNet50 network,the speed is increased 3 times,and the classification accuracy can reach 90.01%,the validity of the model size and speed in image classification is verified.At the end of this thesis,the test of the image outside the data set shows that the network can effectively avoid over-fitting and strong generalization ability.