Optimization And Application Of Target Detection Model Based On FPGA Heterogeneous Platform

With the development of artificial intelligence technology and the increase of social demand,target detection technology is used in more and more scenarios.The target detection technology use machine instead of human to retrieve and identify the target in the complex scene which increases the accuracy of retrieval and recognition and improved the efficiency of work.However,at present,most target detection applications are built on traditional PC machines or heterogeneous machines equipped with GPU,such target detection applications have disadvantages such as high cost,high energy consumption and poor reconfiguration.Because FPGA has advantages of high parallelism,flexible configuration and low power consumption,this paper use yolov3-tiny algorithm to implement high-performance target detection application under FPGA platform and optimized the algorithm model to solve the problems of the model has low detection accuracy and the shortage of resource storage on FPGA chip.The optimization of the model mainly includes the following aspects: we proposed a method which improved the model structure to improve the ability of feature extraction,we used channel pruning method to reduce the volume of the model,then we quantified the model to improve the speed of model detection and adapt to FPGA.This paper verified the effect of model optimization through comparative experiments.The experimental results show that the improvement of model structure effectively improves the accuracy of model detection;The pruning of the model reduced the size of the model by 40 percent with almost no damage to the accuracy of the model;The quantization of the model greatly improved the detection speed of the model on FPGA platform,which is almost the same as that on GPU platform.Finally,according to the actual project requirements and different from other products that use GPU as the algorithm accelerator,we used a heterogeneous platform with an optimized model to design and implement a real-time automatic function module in the prison management system,which proves the effectiveness of the algorithm model optimization in this paper.