Research On Image Recognition Based On Deep Learning And Transplantation Of Android Platform

In recent years,deep learning have attracted widespread attentions in the many fields,especially,have achieved good results in the field of image process.At present,the models of image recognition based on deep learning are operated on the computers.Due to the limited resources of mobile devices,the model of image recognition with real time,small model and high accuracy has been proposed in this paper.Finally,an image recognition APP was developed after the model was transplanted to the Android system.The testing results show that it has a good recognition rate and a certain application value.The main works of the paper are as follows:(1)A model of the image recognition based on convolutional neural network is proposed.By analyzing the characteristics of LeNet-5model,AlexNet model and NIN model,the model with few parameters,short processing time and high recognition rate is designed to solve the problem of limited mobile device resources.Five convolutional layers are used to extract the image features,and the global average pooling layer is used to replace the fully connected layer,which reduces the number of the model parameter and processing time.Through the random transformation of the data set,the recognition rate is increased by 2%.Compared with the LeNet-5 network,the recognition rate of this model has increased by 11%.Compared with the AlexNet model,the recognition rate of this model has increased by 2.12%,the parameter is 1/13 of the AlexNet model,and the training time is 1/4 of the AlexNet model.Compared with the NIN network,although the recognition rate is 2.04%lower than the NIN network,the training time is 1/5 of the NIN network,and the parameter is 1/3 of the NIN network.(2)Developed a local real-time image recognition APP.Through the NDK tool to compile the optimal training model and generate the dynamic link library of.so.The Android application layer will obtain the picture from the camera and convert it to the bitmap format.The picture will be clipped to the size of 32×32 pixels and normalized,and then converted into a four dimension floating point array.By calling the JNI interface,image data is passed to the pre-training model.After the model running,an array is defined to receive the prediction results of the model and the results are displayed on the interface.The APP interface can display picture'previews,recognition results and processing time.The test results show that the objects can be identified at different angles.The average recognition rate can reach 92%.The average time of image recognition is 23ms.which meets the the requirements of accuracy and instantaneity.(3)The image recognition model proposed in this paper is optimized.The weight value of the training optimal model is fixed to a constant value,and the combination of the file of description model structure to generate a.pb file that an Android device can directly calling.Some nodes in the pre-training model can be removed in the application.By deleting these nodes,the model is reduced by 1K.The final optimized model is only 1.1M,which is 2.6M less than the NIN model and 13M less than AlexNet network.It can be transplanted into most of the Android devices.The test results show that the recognition rate of the optimized model is the same as that before optimization,and there is no obvious decrease.