Research On Infrared Target Automatic Detection Technology Based On Lightweight Yolo

Infrared target automatic detection system has a wide range of deployment and application in the fields of anti missile,airspace early warning,sea surface monitoring and so on.As a hot research field,infrared small target automatic detection technology has ushered in rapid renewal and development in recent years.However,the automatic detection technology of infrared weak and small targets still fails to overcome the difficulties such as complex infrared target background,high signal-to-noise ratio,lack of texture information,small size and so on.Among many target detection technologies,the target detection technology based on deep learning algorithm stands out with epoch-making performance.Among them,YOLO series target detection algorithms have received great attention because of their complete network design and simple operation difficulty.To solve the above problems,this paper designs a lightweight infrared small target automatic detection technology based on YOLOv5 algorithm.The main work contents are summarized as follows:(1)Aiming at the fact that there is no high-quality,open data set directly facing the low,slow and small characteristics of infrared,a data set of infrared weak and small targets photographed by long-distance infrared camera with complex and diversified background and less morphological characteristics is established.The shooting subject is a small civil UAV,and the target frame annotation of the data set is completed at the same time.(2)Aiming at the scale problem of infrared weak and small targets,in order to ensure that the characteristic information of weak and small targets is not lost,a high-resolution backbone network is adopted.On the basis of YOLOv5 algorithm having three different scale detection heads,a smaller scale detection head is added to improve the detection accuracy of the algorithm.(3)In view of the problems of too many parameters and redundant feature map in the feature extraction process of ordinary convolution,the ordinary convolution in YOLOv5 benchmark network is replaced by ghost convolution proposed in Huawei Ghost Net,which not only reduces the parameters,but also improves the detection speed and reduces the size of network model.(4)In view of the fact that the target detection algorithm on GPU can only complete the task of training stage,GPU is combined with fmql on-chip operating system based on heterogeneous SOC to complete the inference stage of infrared weak and small target automatic detection algorithm.At the same time,ensure that the network performance after transplantation will not lose too much.The experimental results show that the automatic infrared small target detection technology has significantly improved the reasoning speed and model accuracy compared with YOLOv5 algorithm,and completed the task of transplantation to the mobile terminal.