Research And Application Of Lightweight CNN Model For Mobile Terminal

The convolutional neural network does not need to be designed manually during the entire feature extraction process,and the machine will automatically complete it.However,traditional neural networks generally have phenomena such as high computational complexity and large model sizes.Therefore,researchers have proposed lightweight convolutional neural network models directly for mobile terminals.These network models use a more efficient convolution method.It is suitable for mobile terminal devices with poor performance such as computing power and storage space,such as mobile phones.Considering that the existing lightweight convolutional neural network model still has a large number of parameters,there is still room for optimization of the operation speed.In summary,this paper focuses on the lightweight CNN model for mobile terminals.This paper analyzes several common lightweight models.Based on the MobileNetV2 model,the ShuffleNet model's channel shuffling method is integrated to establish the M-MobileNet model.The main construction method is to use ShuffleNet channel shuffling to replace part of the 1 × 1 convolution.At the same time,in order to avoid loss of model performance,the penultimate layer of the network structure uses a deep convolution layer to replace the original average pooling layer.According to the experimental results,the model implemented in this paper has a TOP-1 accuracy rate slightly lower than that of MobileNetV2 on the CIFAR-10 data set,and the TOP-5 accuracy rate is comparable to MobileNetV2,but it is worth noting the parameter amount and calculation complexity Compared with MobileNetV2,it has decreased by 35%and 38%respectively.M-MobileNet compared to the advanced MobileNetV2 on the ImageNet data set,the operating speed has been increased by nearly 40%,and the same level of accuracy is guaranteed.Based on the M-MobileNet model,this article conducted experiments on the application of two actual scenarios:plant disease identification and real-time target detection.In the Plant Village test set,the model adopted in this paper is very close to the recognition accuracy of MobileNetV2;the average recognition time M-MobileNet is better than MobileNetV2,which is improved by about 23ms.Mobile phone real-time target detection also maintains a high detection accuracy.The detection speed on Huawei MATE20 can reach 50 frames per second.Overall,good results have been achieved,and it can be considered for other fields in the future.