Research On Object Detection Based On Spike Neural Network

With the advent of the era of big data and the rapid increase in computing power,deep neural networks in Artificial Neural Networks can be trained to obtain high accuracy.However,with the increasing complexity of task scenarios,the amount of model parameters and calculations have also increased sharply,resulting in higher storage,calculation,and power consumption requirements.The Spiking Neuron Networks based on neuromorphic computing has the advantages of small calculation amount,low power consumption,and fast information transmission speed,and it provides a good solution to the above problems.However,spiking neural networks are difficult to train,which makes it difficult to be widely used in complex machine vision applications.Therefore,the exploration of new model structures and training strategies is of great significance to the research and engineering applications of spiking neural networks.This paper first selects the LIF neuron model with a fixed firing threshold to construct the spiking neural network model,and realizes the target recognition and detection through the method based on the coding of the LIF neuron population.Compared with the H-H model that requires 1200 floating-point operations for cumulative operations,this model only requires 5times,which greatly improves the computational efficiency.In addition,this article applies group coding to the selected LIF neuron model to achieve information transmission.This solution uses neuron populations to encode the input stimulus signals,and solves the shortcomings that the previously used encoding methods are not accurate enough to represent the analog quantity.This method achieves an effect closer to the expression of information by biological neurons,and makes the process of network processing information more bionic.Finally,the accuracy rates of 98.73%(MNIST)and 86.43%(CIFAR-10)are obtained on the image classification data set,and the m AP of Tiny YOLOv2 model on the VOC2007 target detection data set is 20.Due to the limited effect of the impulse neural network encoding the LIF neuron population,this paper further explores the scheme based on the parameter transfer conversion network.This method is based on the principle that the frequency of the pulse sequence excited by the pulsed neuron matches the activation value of the simulated neuron through the activation function.And adopting the strategy of parameter migration and conversion,the YOLO model is used as the research object to realize the conversion of its parameters from artificial neural network to impulse neural network,including the corresponding realization methods for residual and FPN structure.Aiming at the problem of insufficient activation of neurons in the later layers of the network,this paper not only uses ordinary normalization,but also proposes a channel normalization scheme based on the normalization ratio.After experimental verification,the activation rate of the later layer neurons has been significantly improved.In the end,the accuracy rates of 99.28% and 86.82% are obtained on the MNIST and CIFAR-10 data sets,which reached a result comparable to the original network,and the YOLOv3 model is used to obtain a m AP of 41.23 on the VOC2007 data set.Since the information transmission speed of the spiking neural network is fast,and the resource consumption is small,and the power consumption of the SNN in this article is only 1/5 of the original network,this research scheme has great applicability in the scene with high real-time requirements but limited computing resources,which provides ideas and reference value for the research of SNN application in more complex machine vision tasks.