Research On Infrared Thermal Radiation Artifact Detection Algorithm Based On Deep Learning

Thermal infrared imaging technology is a passive infrared technology that determines the thermal radiation energy of objects.It is widely used in urban security monitoring and other field because of the limitation of human vision.However,infrared thermal radiation artifact seriously interferes with the subsequent detection and recognition processing of infrared images.Therefore,in order to solve the interference of infrared thermal radiation artifact,make full use of artifact information and address the speed and computing ability of the mobile embedded platform,the purpose of this paper is to detect the thermal radiation reflection artifacts of infrared images and propose a lightweight detection algorithm to meet the needs of industrialization.Based on the traits of thermal radiation reflected heat reflection artifacts often exist near a heat source,similar to visible light image reflection phenomenon,therefore this paper,respectively,puts forward the two-step pedestrian artifacts detection algorithm and general thermal radiation reflection artifacts object detection algorithm.The main work and innovations of this paper are as follows:(1)A two-step pedestrian artifact detection algorithm is proposed.Pedestrians with a large amount of thermal radiation are common sources of radiation emission,and the direct detection effect of infrared pedestrian artifacts is poor because of the low contrast and blurry edges of infrared images.Our algorithm is devided into two parts: "pedestrian-pedestrian joint artifact" object detection and fine location of artifacts.Firstly,data enhancement and feature fusion network are improved in the part of target detection,with Ranerasing＿Mosaic data enhancement method designed to expand the data set,which can alleviate the interference source occlusion problem in infrared images,and BiYOLO network constructed,in which Bi-FEM module can enhance infrared weak and small features and edge features well.Secondly,fine location algorithm designs on anchor boxes to suppress useless background information.The experiment verifies the favorable detection performance of the two-step method for pedestrian artifacts,and the accuracy reaches 96.39%,which is 6.22% higher than that of the direct artifact detection,while the average detection speed is 19.61 FPS,and the weight of the training model is 189.56 MB.(2)A lightweight general object artifact detection method is proposed to deploy in the edge computing system.Because of smooth and single characteristic information in infrared images,general object artifact detection tend to false detection and inaccurate positioning,with the requirement of higher computing resources and memory.An exhanced feature-based lightweight algorithm is designed to extend the detection of thermal radiation reflection artifacts from pedestrains to general objects.Firstly,a lightweight detection network named LBi-YOLO is constructed to enhance infrared features,with backbone network,feature fusion network and suppressing algorithm of bounding boxes improved in order to reduce the complexity of network.Secondly,a fine localization of general object based on Gaussian is proposed,which takes artifact constraint information as parameter and uses Gaussian modeling to adjust the rough estimation artifact region.Results show that the proposed algorithm has highly efficacious thermal radiation artifact detection results,the accuracy reaches 91.52%,while the average detection speed is 86.22 FPS,and the weight of the training model is 22 MB,which is suitable for deployment in mobile terminal.