Research On Multispectral Target Detection Algorithm Based On Improved YOLOv4

Object detection is an important research topic in the field of computer vision,and it is also the theoretical basis for some advanced vision tasks.Most of the existing object detection algorithms are based on RGB images of visible light.These methods depend on lighting conditions.In the case of insufficient lighting,the missed detection rate will increase rapidly;thermal images can obtain stable target features under poor lighting conditions.Therefore,this paper,based on the target detection algorithm based on YOLOv4(You Only Look Once v4),which optimizes the YOLOv4 algorithm by combining visible light images and thermal images for the problem of poor detection results under insufficient light during the day and no light at night,thereby improving the accuracy of thermal target detection under poor lighting conditions.The specific work of this paper is as follows:(1)Aiming at the problems of modal imbalance and precision speed imbalance in multispectral target detection,A YOLOv4 based on Discrete Cosine Transform(DCT)conditioning network is proposed.This method uses the difference idea to mix the features of visible RGB image and thermal image to achieve the effect of data enhancement.Specifically: add the TMDF data enhancement algorithm at the input of YOLOv4,use the improved CSPMobilenetv3 in the backbone network to replace the CSPdarknet53 structure of YOLOv4,and add a depthwise separable convolution at the network neck to further reduce the number of parameters,and finally ensure the accuracy.On the premise,the detection speed of the network is improved and the complexity of the network is reduced.(2)Aiming at the problem that the missed detection rate of thermal image pedestrian detection in the KASIT dataset is too high,a YOLOv4-based Discrete Cosine Transform(DCT)adjustment network is proposed.This method uses DCT to obtain the frequency domain information of the image,and divides it into the backbone network of YOLOv4.An adjustment layer is created,and the intensity of the lighting conditions is used as the final output of the adjustment layer,and the output head is used to reversely adjust the output head of YOLOv4 to obtain the final network output,which achieves the purpose of reducing the missed detection rate.In this paper,the method(1)is verified with the KASIT data set and the selfmade data set.Finally,on the premise that the m AP value is basically unchanged,the detection speed is improved by 27% and 24% in the KASIT data set and the selfmade data set.In this paper,the method(2)is verified with the KASIT dataset,and the missed detection rate is reduced by 13.01% under the premise that the detection speed is only reduced by 5%.