Research On Power Optimization Of Image Classification Tasks For Mobile Devices

In recent years,deep learning has developed rapidly in the field of image recognition,and various improved convolutional neural networks for different application requirements have emerged one after another.As mobile system performance rapidly increases,mobile users expect image classification techniques with lower response delay.Due to the instability of the network state and the limitation of network bandwidth,simply deploying the convolutional neural network model in the cloud computing mode cannot guarantee the mobile user experience expectation,and at the same time,limited by the limited resources of the mobile terminal,the mobile device cannot meet the requirements.The model's need for computing,storage,and power resources.Therefore,it is necessary to design a new calculation mode,so that the image classification task based on the convolutional neural network model can satisfy the user's expectation of fast response,low energy consumption and high accuracy.Therefore,based on the energy efficiency and classification characteristics of the convolutional neural network model,this paper studies the task scheduling of mobile image classification.The main work is as follows:1.Aiming at the problem that the inference of the deep learning model is too long for the unstable network state.Based on the classification characteristics of different convolutional neural network models for the scenes of image classification tasks performed by intelligent mobile terminals,this paper proposes an adaptive depth model selection strategy.We firstly analyze the classification results of different models,and inputs the greedy algorithm to construct the image classification model based on its energy efficiency.Then,the pre-classification algorithm is arranged and combined to form a variety of alternative classification schemes.Finally,each image classification model in the sequence is used to constructing a pre-classification model corresponding and selecting the pre-classification scheme with the best performance.This strategy selects the most suitable image classification model for different images.The experimental results show that compared with the use of Inception_v4 alone,the adaptive model selection strategy in this paper reduces the inference time by 15.8%,the energy consumption by 71%,and the accuracy by 7.6%.2.Aiming at the problem of high delay and high energy consumption for performing image classification tasks locally on the mobile end.We combine with the idea of edge computing,an edge scheduling strategy based on image pre-classification is proposed.Firstly,based on the inference time and energy consumption of images on different convolutional neural network models and different devices,lightweight models and high-performance models are deployed on the mobile and edge servers respectively.Then,based on the feature images and the inference results of the models,we use different classification algorithms to train the pre-classification model,and select the optimal pre-classification model.The strategy is scheduled by the classification result of the pre-classification model,so that the mobile device and the edge device work together.The experimental results show that compared to relying on the intelligent mobile terminal,the image inference time is reduced by 91.6%,the energy consumption is reduced by 92.5%,and the classification accuracy rate is increased by 3.1%.