Research Of Image Classification Algorithms Based On Convolutional Neural Network

Convolutional neural network,as a kind of neural network in deep learning,has excellent performance in the field of image processing.Its advantage is that it can directly act on the image pixels,convolve the pixels,and extract the feature information of the image..On the other hand,the weight sharing and pooling layer of the convolutional neural network can greatly reduce the number of parameters of the network,making the network training faster and more efficient.For deep learning,the amount of data often determines the performance of the model,otherwise the model will often have serious over-fitting after full training.Based on this researcher,the concept of migration learning is proposed,but the migration learning needs to use the network structure of the migrated model,and the different network structures have different representation capabilities.Moreover,for different problems,the optimal model structure It is not the same,so you can't just migrate the model parameters that have been trained.The main research content of this thesis is how to make the algorithm model achieve good classification accuracy while taking into account the time performance in the case of insufficient data.Based on this paper,the following attempts are made:(1)Optimize the classic Vgg-9 model structure,and delete the number of convolution kernels in the original model structure and the nodes in the fully connected layer,and change the last largest pooling layer to the global maximum.The pooling layer is designed to construct the convolutional neural network model structure.The experimental results show that the optimized model structure is higher in time performance and prediction accuracy than the original model structure.(2)A noise reduction self-encoder module is added to the convolutional neural network,and the noise problem of the data may be effectively prevented by the noise reduction self-encoder.The experiment proves that the model resists image noise after adding the noise reduction self-encoder.The ability has been further improved,effectively mitigating the impact of data noise on model performance.(3)Using cifar-100 as the initialization data set of the model parameters,using it to initialize the model training,in addition to the data enhancement operation of the experimental data,the effective data volume of the expanded experiment,the experimental results prove the parameters Both the initialization and data-enhanced models have a certain improvement in both convergence speed and prediction accuracy.In addition,this article also applied a lot of mature optimization measures in the model,and achieved good experimental results in the experiment.Finally,the accuracy of the model in this paper reached 90.38%,and the single sample prediction time was 0.6ms.The requirements of the image classification algorithm for accuracy and time performance are basically satisfied.