Fuzzy Scene Vehicle Re-Identification Using Spiking Neural Network

With the rapid development of artificial intelligence,the social requirements for the effects,types and power consumption of terminal artificial intelligence devices are increasing.Among the problems faced by the terminal equipment,the importance of the vehicle re-identification task to society is very obvious.Mature operation of vehicle recognition system,can be efficient management of vehicles in the society,social security,traffic stability has played an important role.However,under the influence of the epidemic on society,the problem of high power consumption of terminal artificial intelligence devices has become increasingly prominent.In the intelligent municipal terminal intelligent system,once the power consumption of the terminal device exceeds the threshold,the terminal device cannot run normally or the operation effect is not good.At the same time,in the process of data transmission,if the terminal equipment is affected by extreme weather or bad external conditions,and the terminal equipment has great loss,it needs to be customized again,resulting in a large amount of time,budget and labor waste.In this thesis,aiming at the problem of high overall power consumption and low quality data processing effect of vehicle reidentification task on edge equipment,A Fuzzy Scene Vehicle re-identification using Spiking Neural Network was proposed.The research and work are carried out in the following three parts:1)In this thesis,a fuzzy feature extraction model based on pulsed neural network(SNN)is designed for vehicle re-recognition in fuzzy situations.In the model,the activation layer is replaced by pulsed neurons,and the residual of input and output is calculated.Meanwhile,the channel spatial attention mechanism is added to complete the feature extraction task without increasing the overall power consumption of the algorithm.Experimental results show that the algorithm can complete the vehicle rerecognition task on fuzzy data sets with stable effect and can be applied to practical scenes.2)Aiming at the problem of high power consumption of the algorithm in the previous part of the thesis,a global feature extraction model based on pulse neural network is designed.In the model,the pulse Res Net50 network model is used,and the network structure is designed to avoid the gradient explosion and gradient disappearance of the pulse neural network.Experiments show that the pulse neural network has significant advantages in power consumption,and the terminal chip using the pulse neural network as the backbone network can be adapted to the terminal device,which has good stability and can effectively reduce the overall power consumption of the algorithm.3)In this thesis,the global feature extraction network model and fuzzy feature extraction network model designed in the first two parts of the research are combined,and other methods are used to optimize the whole,so as to obtain a low power algorithm for vehicle re-recognition task in fuzzy scenarios.After the fusion of the two parts of the algorithm,the training time is faster and the number of iterations is reduced,indicating that the algorithm has low power consumption and excellent performance in the vehicle re-recognition task,and can complete the vehicle rerecognition task in real scenes.