Lightweight Network Design And Hardware Implementation Of Infrared Small Target Detection Algorithm

With the advancement of materials science,the application of stealth aircraft poses significant challenges to active radar detection,which can be effectively addressed by infrared detection technology.However,due to the limited number of pixels and texture information occupied by small targets in infrared images,detecting them becomes difficult.Moreover,the associated detection algorithm models have complex networks and a large number of parameters,making them unsuitable for deployment on embedded devices.In light of this,this paper focuses on the research of lightweight network design and hardware implementation for unmanned aerial vehicle(UAV)target detection algorithms in infrared scenarios.The main contents include the following:1.This paper mainly studies the lightweight network design and hardware implementation of UAV target detection algorithm in infrared scene.Aiming at the problems of low accuracy,poor real-time performance and difficult detection of small targets in traditional target recognition algorithms under infrared scene,a lightweight model FS-YOLOv5 based on infrared scene is proposed.Firstly,YOLOv5s single-stage target detection network is used as the basic network,and a FS-ShuffleNetv2 network with SimAM non-parametric attention mechanism is proposed to replace the CSPDarkNet backbone network in the original network to extract feature maps.Secondly,based on the original network’s CIoU loss function,the Power transformation is introduced and replaced by α-CIoU to improve the network’s ability to detect small targets.Then,the K-means++clustering algorithm is applied to the homemade FR dataset to regenerate Anchors,and finally,the detection head of the network is further simplified based on the characteristics that the dataset is entirely composed of small targets.Through ablation experiments on the FR infrared small target dataset,it is verified that the FSYOLOv5 lightweight algorithm meets the detection tasks of UAV targets under infrared scene.Compared with the original network,under the premise of only reducing the average accuracy by 0.98%and the recall rate by 2.9%,the size of the FS-YOLOv5 model is reduced by 5.22M,and the detection speed is increased by 33FPS,and the network computational complexity is reduced by 13.4G,which meets the requirements for deployment on mobile devices.2.In order to meet the requirements of real-time performance and portability for UAV recognition systems,a hardware acceleration system design for UAV recognition algorithm was completed using the FACE-Z7-B development platform.In the hardware design,the hardware system was divided according to the characteristics of the FS-YOLOv5 network structure,and a fixed-weight pulsing array acceleration scheme was proposed.The design and simulation verification of the pulsing array module,pooling module,and activation function module were completed,and finally,the system was verified and performance evaluated based on the FACEZ7-B development platform.The evaluation results show that the hardware acceleration system designed in this paper only requires 172ms to process one image at a clock frequency of 150MHz,with a power consumption of only 5.461W.Compared with the inference speed of the Cortex-A9 processor,the accelerator achieved a speedup of about 70 times with an increase in power consumption of 3.926W.Meanwhile,the accelerator designed in this paper has the characteristics of low power consumption,which is about 1.7 times that of GPU and 20 times that of CPU.