| Compared with visible light imaging,infrared imaging system can work all day,not susceptible to bad weather,has strong concealment and other advantages,so it has a very wide range of applications in civil,military,security and other fields.However,due to the influence of imaging mechanism,infrared images are usually characterized by low contrast,fuzzy imaging and noise interference.In addition,small and dim targets have a small pixel proportion,low resolution and obscure feature details in infrared images,so it is difficult to extract features from small targets in complex scenes,which greatly increases the difficulty of target detection.Compared with traditional detection methods,the deep learning algorithm has the advantages of flexible network structure and powerful ability of automatic feature extraction.As a target detection method with strong robustness,versatility and high accuracy,it is more suitable for dim-small target detection in infrared images under complex background.The existing classical target detection network has a good detection effect on medium and large targets with no overlapping occlusion phenomenon in general scenes,but the average detection accuracy of dim-small targets in infrared images under complex background is much lower than that of medium and large targets.Therefore,there is still room for improvement in the detection of dimsmall targets in infrared images.In view of the above problems,this paper focuses on the following three aspects:(1)The characteristics of infrared image and dim-small target are introduced,and the influence of background characteristics and noise characteristics on infrared dimsmall target detection is analyzed.The performance evaluation indexes of dim-small target detection are introduced,and the difficulties of infrared dim-small target detection are summarized.In addition,the traditional object detection method and deep learning method are analyzed and tested respectively,which serves as the theoretical basis for the research work of this paper.(2)A detection method based on YOLOv5 combined with attention mechanism is proposed to solve the problems of difficult target feature extraction and low detection accuracy due to the low percentage of dim-small target pixels and insignificant feature details in single frame infrared images.This method is based on YOLOv5 network.By designing Sim AMC3 attention mechanism module,designing target detection head and improving prediction box screening mode,the average accuracy of dim-small target detection in single frame infrared images can be improved.The experimental results show that compared with some mainstream target detection algorithms,the detection effect of the proposed algorithm is greatly improved.Compared to the baseline algorithm,the proposed algorithm increases m AP by 4.8% and 7.1% in the two infrared dim-small target data sets.(3)A detection method of YOLOv5 combined with interframe information is proposed to solve the problem of false alarm in dim-small target detection caused by noise and complex background in sequential infrared images.In this method,after decoupling the YOLOv5 detection head,the feature information screening module and interframe information link module were combined to detect the real target and eliminate false alarms.The experimental results show that compared with some mainstream target detection algorithms,the detection accuracy of the proposed algorithm is greatly improved.Compared to the baseline algorithm,the average accuracy of the proposed algorithm is improved 4.1%,false alarm rate decreased by3.97%,can effectively combine the interframe information for infrared sequence image target detection.The proposed single-frame algorithm and multi-frame algorithm are tested on relevant data sets respectively,and can effectively detect infrared dim-small targets in different complex backgrounds,showing good robustness and adaptability,and can be effectively applied to infrared dim-small targets in complex backgrounds. |
PDF Full Text Request