Realization Of Small Object Detection Algorithm Based On Convolutional Neural Network Acceleration Engine

There have been many researches on object detection already.Object detection algorithms emerge in endlessly,such as R-CNN,YOLO,SSD,etc.,but the detection effect of these algorithms on small objects is not satisfactory.At the same time,convolutional neural network has a high computational density,so it is particularly important to study how to use hardware to accelerate for real-time object detection.In this paper,the algorithm is transplanted to the target board and the hardware unit on the target board is used to accelerate the calculation.The optimization scheme is proposed to solve the problem of the algorithm migration accuracy decline.On this basis,a small object detection system based on acceleration engine is designed through Hisilicon MPP and SVP development platform.The main research contents are as follows:(i)Firstly,the object detection algorithm is studied.YOLO V3 is improved in detection performance and speed for the problem of poor detection performance of small objects.Then,the algorithm transplantation is implemented.Transplant the algorithm model to the Hi Silicon development board,and use the NNIE hardware module to accelerate the convolutional neural network through Hisilicon SVP platform.Finally,in order to solve the problem of reduced accuracy in algorithm transplantation,the optimization method of this paper is proposed: reasonable setting of conversion parameters and confidence thresholds,and image scaling in the form of equal proportions.(ii)A small object detection system based on the Hi Silicon SVP platform and MPP platform is designed.In the different stages of object detection in the system,different hardware on the development board is called for processing to detect objects quickly.Experimental results show that the confidence level caused by algorithm transplantation is reduced within 4% and the accuracy rate drops by about 2%,and the actual number of frames per second can be increased to 17 after the improvement of YOLOV3 algorithm.