The Optimization And Application For Image Recognition Model Of Deep Learning

Deep learning(DL)represents the state-of-the-art technology in the field of artificial intelligence,and made a breakthrough in image recognition,speech recognition and natural language processing since 2006 in the background of big data and powerful computing hardware.However,its huge computational complexity and model memory footprint seriously hindered the universality and flexibility of artificial intelligence,so that many applications based on deep learning rely on cloud computing,at the same time,great power consumption is another question.For spiking neural network,the shortcoming of deep learning happens to be its advantages,this paper made a ser ies of research and applications to combine deep learning and spiking neural network.In this way,deep learning applications can be directly used in the embedded devices,which has a broad developing prospect in intelligent equipment and robotics,etc.Firstly,we analyzed the dynamic characteristics of several spiking neurons to lay the foundation for follow-up study in view of the neural network is make of a number of neurons.Secondly,we proposed some strategies to adjust super parameters of neural network training algorithm,such as gradient,momentum and so on.Then we changed the form of transfer information from rate-base to spike-base,and added the intrinsic plasticity mechanism to the network model,and verified its feasibility and effectiveness by handwritten digit recognition experiment.Finally,we optimized the convolutional neural network through adopting max-feature-map(MFM)neuron to replace rectified linear units(Re LU)and added network-in-network(NIN)structure to the model,and transplanted into the embedded device Raspberry Pi after realizing the real-time face recognition application.Handwritten digital experiment shows that it averagely spend about 5.8ms on identifying a single handwritten digital gray image and reduce to 2.5ms if adding intrinsic plasticity mechanism to the model,besides,the spiking deep belief network can adjust the intrinsic spiking rate adaptively.The real-time face recognition project shows that the time cost reduce to 67 ms with the accuracy of 98.13%,and it works stability after transplanting into Raspberry Pi.