Improvement Of Convolutional Neural Networks For Image Classification

Image classification refers to the image processing technique of mining and acquiring different features of image information and distinguishing different types of images.The core content of image classification lies in the extraction of image features.Extracting high-level semantic features can correctly understand the content or essence of images,thus improving the accuracy.With the advent of the era of big data and the rapid development of Internet technology,deep learning methods have become a research hotspot in image classification tasks.As one of the most widely used methods in deep learning technology,convolutional neural networks have strong feature extraction ability and nonlinear fitting ability.They are the most effective algorithms in image classification task nowadays.Many researchers have carried out a lot of research and made excellent progress,but convolution neural networks still have some shortcomings,such as complex network models,long training time,large amount of computational resources,slow convergence speed,over-fitting and classification accuracy to be improved.Therefore,it has become one of the important research directions to improve the convolution neural network and design better performance convolution neural network method for image classification.(1)Aiming at the problems of deep network structure,large number of parameters,long training time,over-fitting,weak generalization ability and recognition accuracy,the paper proposes a deep network transfer method for flower image classification.We choose the classic network structures: VGG16,VGG19,InceptionV3 and ResNet50.Firstly,pre-training models are performed on the ImageNet dataset,so that the deep network parameters have the feature extraction performance of the image in the natural scene and the optimization parameters of the network are obtained.Then the pre-training model is input into the target dataset for transfer training,and further network training and Parameter optimization,obtaining rich high-level semantic information of flower data,and comparing with random initialization network,the experimental results show that the recognition performance of ResNet50 transfer method is the best,and the classification accuracy in Oxford-102 dataset reaches 96.57%,and The classification accuracy on the Oxford-17 dataset also reached 95.29%,and in the Guangxi flower dataset,the models have excellent recognition accuracy and robustness.(2)In order to further improve the recognition accuracy of Oxford standard flower dataset,reduce the network training time,and effectively improve the robustness and generalization ability of the model,the deep network secondary transfer technology is proposed based on the transfer method,which is obtained through the training of ImageNet dataset.The pre-training model,moved to the Guangxi Flowers Dataset to further train the network and optimize the parameters,and the resulting flowers transfer training model was again transferred to the Oxford flowers dataset.The experimental results show that the secondary transfer method is compared to the transfer learning method in the model.The recognition accuracy and generalization ability are significantly improved.The classification accuracy of the ResNet50 secondary transfer method in the Oxford-102 data set reaches 97%.(3)Aiming at the network degradation problem caused by the excessive number of parameters in the convolutional neural network,the network is over-fitting and the number of layers is too large,and the parameters are too large,which leads to the accuracy degradation,an improved shallow dense convolutional neural network method is proposed.In the original dense network,the improved fire module was introduced,and the model feature extraction ability caused by compression parameters was reduced.A two-dimensional convolution kernel separation technique was proposed to obtain multi-scale features of target diversity,thus improving the characteristics of the model extraction ability.In this paper,the proposed model is compared with the original dense network,and compared with the latest classical deep convolution network,and the model simplification experiment is carried out in the proposed model.The effectiveness of the proposed method is demonstrated from multiple angles.