Infrared Target Detection And Recognition Based On Deep Learning

In recent years,with the development of artificial intelligence,target detection,as one of the important tasks of computer vision,has been widely used in military and civil fields.At present,the target detection algorithm based on deep learning has become mature and has achieved remarkable results.However,most of the researches are based on visible images and lack of researches based on infrared conditions.Infrared imaging technology has the characteristics of long operating distance,strong anti-interference ability,and capable of all-weather monitoring.It is widely used in fields such as automatic driving and security.The challenge of infrared target detection is that,compared with visible images,infrared images have the characteristics of low signal-to-noise ratio,fewer target texture features,and poor resolution,resulting in poor performance when using current mainstream target detection networks for infrared target detection.In view of this,this paper focuses on the following problems: 1)Infrared image features such as blurry imaging and low contrast make network detection accuracy drop significantly;2)At present,open source infrared data is scarce,and data enhancement is needed to improve The generalization of the network;3)When the target scale span is large in the image,the performance of the detection algorithm will be affected;4)The target detection network has the problem of inaccurate positioning.5)Special tasks require the algorithm to be transplanted to the embedded platform,there are problems such as computing power and storage.The main research contents are as follows:In terms of infrared image preprocessing,aiming at the problems of low signal-to-noise ratio and low resolution of infrared images,which seriously affect the performance of target detection,according to the characteristics of infrared images,the current mainstream image enhancement methods are analyzed,and a data enhancement algorithm based on infrared image channel expansion is proposed.This algorithm takes advantage of the infrared image as a single channel,generates the infrared image suitable for target detection through the information selection and enhancement algorithm,realizes the infrared image channel expansion and data enhancement,effectively increases the amount of information in the original image,and achieves the purpose of improving the detection accuracy of the network.In terms of target detection,considering the perspective of detection accuracy and speed,the YOLOv3 algorithm with speed advantage is selected as the basic network for infrared target detection.As the feature map used for detecting large and medium-sized targets in the YOLOv3 network is not large enough,the accuracy of the YOLOv3 network is improved on small targets,but the performance of the YOLOv3 network is deteriorated on medium and large targets.Therefore,this paper proposes to add an SPP module after the feature extraction network to integrate local features and global features to improve the expression ability of network features,increase the receptive field of the network,and further improve the localization ability of the network by modifying the loss function.The experimental results show that the algorithm can significantly improve the detection accuracy of the large target in the image,adapt to the change of target scale and locate the target more accurately.In order to meet the needs of actual projects,it is necessary to transplant infrared target detection algorithms to embedded platforms,but the current mainstream detection networks are relatively complicated and difficult to deploy to embedded platforms.Therefore,this paper compares the current mainstream embedded networks,chooses the high-precision YOLOv4-tiny network,uses the weight file after training on the visible UVA data set as the initial weight of the network,accelerates the convergence of the network,and data enhancement algorithm based on infrared image channel extension is used to enhance the data.By using the self-built infrared UVA data set for testing,the experimental results show that this method can further improve the detection accuracy and transplanting the trained algorithm to the TX2 platform can basically achieve real-time detection.In conclusion,this article in view of the infrared target detection in low SNR and low resolution,the target is not easy to identify,easy to fault detection and other issues.Carried out the research of infrared image data enhancement preprocessing,image data generation channel expansion and data enhancement.Based on the enhanced data,a detection algorithm that adapts to multi-scale infrared targets in complex scenes is proposed,and it is implemented in real-time on the embedded platform.Experiments show that the algorithm can effectively realize the real-time detection of infrared targets on the embedded platform.