Deep Learning Methods Of Infrared Target Detection In Complex Background

A computer vision technique called infrared target detection varied scenarios to determine the category and orientation of interest targets.Infrared target identification with complicated backgrounds now confronts two issues: the first is the scarcity of pertinent data sets,and the second is the model’s sluggish convergence speed and poor detection accuracy.In order to perform in-depth research on infrared target identification technology in complicated backgrounds issues,this project employs deep learning,the major work is as follows.(1)A network model for infrared target detection was constructed.The detection network is based on the YOLO series model as the framework,CSPDark Net-53 as the feature extraction network,and Feature Pyramid Network as the feature fusion network.The efficient long-range attention network module is added,and the Coordinate attention method is implemented,to help the deep network train efficiently.The constructed model effectively improves the accuracy of infrared target detection.The experimental results show that compared with the benchmark network model,the mean Average Precision increases from 34.9% to 38.1%.(2)A pre-training method for instance features in complex background is proposed.To address comprehend scenes slow convergence of the target detection network,the method first constructs an image stitching template to reconstruct the input image into an image with a complex background.The feature extraction network is then pre-trained using the rebuilt picture.Lastly,transfer learning technology is used to load the pre-trained model into the detection network.And introduced the Mosaic data augmentation technique in data processing to further solve the problem of low infrared image data.The experimental results show that the number of iterations required for the network model to converge is reduced to half of the original number,and the convergence speed is effectively improved.(3)The ablation experiment consists of two parts: parameter ablation and module ablation.The optimal network model parameters are determined by parameter ablation;the ablation training is performed for each module in the module ablation experiment to determine the effect of modules on network performance.In the comparison experiment,the model in this thesis infrared target detection performance is compared and assessed with the current mainstream approaches.It improved the results show that the existing detection accuracy of the model in this thesis for all targets,with the most improvement for the pedestrian AP,from 77.2% to 88.1%.