Research On Infrared Targets Classification Technology Based On Deep Learning

Infrared imaging captures the heat radiated by the object itself due to temperature.Infrared images have become a research hotspot in image processing and deep learning in recent years.Compared with visible light,infrared images rely on its own energy,so it has stronger penetrating power and strong visualization under the starry night.It can obtain images of targets in complex combat environments.At present,the automatic recognition and classification of visible light images are widely used and developed relatively mature,such as automatic product classifiers,security monitoring,and remote sensing scene classification.Infrared images are widely used in military research because they are not affected by weather and light sources.However,because they involve national defense information and national security,there are not many articles on infrared image research,among which there are only a handful of articles on infrared target classification..The working conditions of the infrared imager are not affected by the light source,and the imaging will not be disturbed by the complex environment.However,infrared imaging is difficult to capture the detailed information of the target,the imaging quality is relatively fuzzy,and the resolution and signal-to-noise ratio are low.Therefore,in the research of infrared target classification,there will be difficulties such as unobvious details,blurred edges,and classification errors.This thesis combines the characteristics of actual infrared imaging and the background of deep learning to conduct research from image processing,classification model algorithm optimization,model generalization ability improvement,etc.The main research content and innovation points are as follows:(1)The characteristics of traditional infrared images have been studied and analyzed,and an RGB color infrared database has been constructed on this basis.The images taken by the infrared imager have more detailed information and color information,and intuitively show the infrared sensor working principle.(2)Through the analysis of the image taken by the infrared imager,several denoising algorithms for infrared image processing are studied,and the BM3 D denoising algorithm is improved,using wavelet transform to combine R,G,B component expression.It filters out the noise points in the image well and achieves the purpose of deep denoising.A new denoising algorithm is provided to improve the quality of infrared images.(3)A target classification algorithm suitable for three-channel infrared images is proposed.This thesis analyzes the history of the development of deep learning,as well as several existing target classification algorithms based deep learning and their advantages and disadvantages.Adding the scale fusion detection layer to YOLO-V3 and improve it.The proposed model improves the model's classification performance for small targets,fuzzy targets and occluded targets.(4)On the basis of the existing collected images,in order to expand the data set and increase the difference between the infrared images,the three-channel infrared image and the single-channel infrared image are randomly shuffled and re-experimented to improve the model.The training results showed the effectiveness of increasing the data set.