Flower Recognition System Of Mobile Based On The Deep Convolutional Neural Network

With the development of high performance computing chips,the deployment of deep learning models on mobile devices has become a hot topic.As a kind of deep learning technology,convolutional neural network(CNN)is very suitable for flower recognition,a fine-grained image classification category which is difficult to classify,due to its ability to learn adaptively and combine effective features.CNN usually has millions of computing nodes and parameters,which means that hardware facilities should have strong computing power and large storage resources.However,in the edge computing(such as mobile and embedded platforms),the device has low computing power and small memory,so it is difficult to deploy.This paper focuses on the deployment and optimization of mobile CNN,and explores a series of technologies to make the flower recognition model smaller in size,shorter in prediction time and higher in recognition accuracy.The main work of the paper includes:(1)In terms of model selection,this paper analyzed the evolution and network structure of CNN in detail,as well as various classic CNN architectures.Then,starting from the time complexity and space complexity of CNN,a lightweight CNN Mobilenet-V2 with good tradeoffs in scale,speed and precision was selected.(2)In terms of model training and optimization,an RMSProp optimization algorithm combining Momentum is proposed by analyzing the training process of CNN and the commonly used optimization algorithm.Compared with other optimization algorithms,this algorithm can decrease the loss function faster,increase the accuracy rate higher and approach the optimal solution faster within the same iteration rounds.In addition,this paper adopts dual-gpu parallel training of synchronous mode to shorten the training time,and adopts a series of training strategies such as transfer learning,L2 regularization and data augmentation to increase the accuracy of the model.(3)In terms of model compression,an efficient 8-bit integer operation neural network quantization scheme is used,which can convert CNN's 32-bit floating point operation into efficient 8-bit integer operation,reduce the model size and shorten the model prediction time,but the accuracy decline is very low.We use Mobilenet-V2 as the benchmark,compared with the original float32 model,the size of the transformed uint8 model was reduced by 52%,the prediction time was shortened by 48%,and the accuracy was reduced by 1%.(4)Model migration is implemented on android mobile terminal based on TensorFlow Lite open source framework.The APK size of the flower recognition system is 7.17 MB,the single frame prediction time on the mi 6 phone is 57 ms,and the Top-1 accuracy on the Oxford-102 flower public data set is 95.9%.Besides,the comparison with other flower recognition models shows that this system has good performance in scale,speed and precision.