Research On The Design Method Of Convolution Neural Network For Mobile Device Video Stream Object Recognition

With the development of technology,the application of object recognition in video stream based on mobile devices has a broad prospect.Traditional image matching algorithms and machine learning algorithms cannot gain a good recognition performance.Compared with the deep convolutional neural network,which has achieved a significant progress in object recognition,a noteworthy trend is to apply the deep convolution neural network technology on mobile devices to realize a high accuracy object recognition.However,the deep convolution neural network is a computationally application,at the same time considering the resource-constrained of mobile devices.Directly deploying the deep neural network on mobile devices will cause a series of problems,such as high time delay caused by insufficient computing capacity,the insufficient cruising ability caused by excessive energy consumption,the running unavailability caused by excessive memory consumption,and so on.These problems directly hinder the development of object recognition on mobile devices.The existing solutions mainly include the cloud computing and the network compressed.For the cloud computing,the mobile device only serves as the data collection,sending and receiving platforms.But using the cloud computing will cause a series of problems,like the privacy leakage.Gaining more concern is the network compressed.By compressing the convolution neural network size,enabling the network can be deployed on mobile devices.However,these existing works only focus on the network structure.In particular,when facing complex identification tasks(with many kinds of objects to be identified,such as the object recognition in real-life scenes),there are still many problems,such as very complex network,high time delay and high energy consumption.For this challenge,this paper deeply analyzes the intrinsic connectedness between the object recognition and the context.Proposing a new compressed convolution neural network architecture design solution.The main design idea of this paper is to divide the whole object recognition task into two sub-tasks: the context recognition and object recognition within this context.The task decomposition will reduce the network complexity,and fundamentally guarantees network the low time delay and low resource.Furthermore,this paper builds the energy consumption prediction model,then establishes the direct connection between energy consumption and floating-point computing amount.And putting forward the network structure design and parameter setting scheme according to the floating-point computing amount.Ensuring that the designed network can achieve better recognition performance within limited resources.In addition,this paper proposes a series of performance enhancement and optimization techniques for the network.In particular,the proposed convolutional decomposition technique cannot only enhance network performance,but also make network design scalable.In order to test the compressed neural network architecture design performance,a cognitive aid prototype system is implemented based on these design ideas,which has been successfully deployed on smart phones,and extensive experiments are conducted.Finally,the experimental results show that the prototype system can achieve good recognition performance.And at the same time,the system can be deployed on mobile devices,with low time delay and low energy consumption.In particular,this paper conducts detailed experiments on network performance enhancement and optimization techniques,and the experimental results show that these techniques can effectively optimize and improve network performance.In addition,this paper tests the energy consumption prediction model in detail.The results show that the model can accurately predict the convolutional neural network energy consumption on mobile devices.