Research On Mobile Image Recognition Based On Deep Learning

Today is the age of information,people's daily life is inseparable from information.Deep learning has made breakthroughs in the field of imaging,and the convolutional neural network is the leader of this wave of deep learning.Usually any image target in reality is not in a simple background and a single target environment.Taking flower images as an example,it is sometimes difficult to identify such images using CNN models.Furthermore,the increasing network depth and model size of convolutional neural networks pose great challenges for the deployment of deep learning on the mobile side.Such a large model can only be used under a limited platform,and cannot be transplanted to mobile terminal.Based on the theoretical guidance of pruning compression,target background separation and model transplantation,Select the flower set to start the experiment,and study the rules and methods applicable to more image recognition and model transplantation problems.In view of the current problem of excessive CNN model,the method of pruning compression was used.and MNIST data set was selected to train the model,then sorted the weight of each layer from small to large and deleted the connection with the least weight.However,this method brings sparse connections,but the GPU is not good at processing.Sparse matrices do not lead to efficient model compression.Then the flower data set was selected to train the model and introduced the idea of pruning based on L2 norm into the MobileNets network model.Instead of trimming the single connection,the whole convolution filter is trimmed and the size of the MobileNets model is further compressed based on the accuracy of the model recognition.For the image target in complex background,the foreground image target was detected by Boolean saliency,and then combined the GrabCut algorithm to separate the background of the image target,and finally recognized and expressed it by MobileNets convolutional neural network,compared with directly using MobileNets for recognition.Effectively reduce the impact of complex background on recognition accuracy.For the problem of mobile deployment,AndroidStudio development tools was used to transplant the models into different Android phones,and tested the CPU usage of the APP,start time-consuming,and identified time-consuming tests.The experimental results showed that The APP generated by the pruned model can satisfy the application of the mobile platform.