Research On Event-based Convolution Algorithm And Design Of Event-based Convolution Processor

Traditional vision sensors are based on the sampling mode of frame.With the increase of resolution and frame rate,the CMOS image sensor will generate a large amount of data and require a higher transmission power,which limits the processing speed of the visual system.Based on the study of biological vision,a bionic AER(Address Event Representation,AER)vision sensor is proposed,which can effectively reduce the data redundancy and has the characteristics of super high speed and low delay.The AER image sensors only output event trains with address information and properties,which makes the traditional methods are not suitable for AER image sensors.In order to deal with the output events effectively,we need to study adaptive processing module.In this paper,the event-based convolution algorithm is studied and a small size event-based convolution processor is proposed.Firstly,this paper briefly analyzes the basic principle of the AER vision system,including the AER protocol and the basic structures of AER vision sensors.Two kinds of spiking neuron model are introduced,based on which the event-based convolution algorithm is proposed.Then the basic structure of event-based convolution processor is investigated.Secondly,an event-based convolution simulator is established,based on which a hierarchical recognition system with two layers convolution operation is built.In the system,Gabor kernels are used to extract the features and spiking neural networks(SNN)are used to accomplish the classification.Finally,a small size event-based convolution processor is designed.A kernel RAM of32×32 2bits word is implemented to store the kernels.In each pixel unit,a 7bit counter is used to accomplish the accumulation instead of a traditional accumulator.In this paper,the event-based convolution algorithm is simulated on the Matlab platform.The simulation results show that the algorithm can realize the feature extraction based on the input events.The experiments on the Mixed National Institute of Standards and Technology(MNIST)image dataset demonstrate that the hierarchical recognition system can achieve a recognition rate 90.57%.In the Cadence platform,a 32×32 small size convolution processor is simulated.The experimental results show that the proposed convolution processor can achieve ideal convolution results.In the SMIC 0.18?m technology,each convolution unit occupies 37.5×40?m~2.The minimum latency between input and output event flows can be nearly 17ns.Input event throughput can reach 12.5Meps.It proves that the event-based convolution algorithm can efficiently realize the feature extraction,which is suitable for the applications of high speed AER vision system.