Research On Infrared Multi-target Detection Method Based On Deep Learning

Infrared thermal imaging has the advantages of passive imaging,and the working condition is independent of light visibility and weather conditions.Therefore,target detection systems based on infrared thermal imaging have broad application prospects,especially playing an irreplaceable role in the military and civilian fields.This paper conducted related research on infrared target detection based on the deep learning method.The main research contents are as follows:1.Research of infrared target detection methods based on machine learning was conducted.The FLIR ADAS multi-target infrared dataset was used to verify the infrared target detection method based on histogram of oriented gradient and support vector machine.The experiments proved that,due to the limitation of expression ability of the hand-crafted feature and sliding window mechanism,the above method has drawbacks such as low detection accuracy and low speed in multi-target infrared detection.2.Aiming at the problems of expression ability of hand-crafted feature and other problems,deep learning was introduced into infrared target detection,and research on infrared target detection methods based on deep learning was conducted.Experiments show that compared with the method based on machine learning,the two-stage model Faster R-CNN and one stage model SSD has higher detection accuracy and faster detection speed.Meanwhile,the accuracy performance of Faster R-CNN is better than that of SSD,and the detection speed of SSD is faster than that of Faster R-CNN.3.Aiming at the problems of low detection accuracy of the one stage model and slow detection speed of the two stage model,a pseudo-two-stage model using the characteristics of infrared images was studied.The FLIR ADAS multi-target infrared data set was used to verify the infrared target detection method based on the pseudo-two-stage target detection model.Experiments show that the pseudo-two-stage target detection model can retain the advantages of faster detection speed of SSD and obtain higher detection accuracy than Faster R-CNN.Meanwhile,the effectiveness of the anchor pre-refinement module,dual-pass fusion module,adaptive channel-wise enhance module and focal loss added to the one stage model was verified.