Research On Key Technologies Of Aerial Infrared Target Detection Based On Deep Learning

Infrared imaging has outstanding characteristics such as long detection distance,high concealment,and all-weather day and night work.It has an irreplaceable position in the field of space target detection.In the infrared detection imaging system,target detection plays a particularly important role.It can simultaneously realize image content recognition and target location,which is the basis of subsequent tracking tasks and also provides strong support for system decision making.In a typical application scenario,there are a large number of cloud undulations,ground objects and other interferences with strong infrared radiation characteristics.The traditional infrared target detection algorithm is not ideal in accuracy and robustness.In recent years,deep learning has achieved great results in many computer vision tasks,and its use in infrared target detection can help to make up for the shortcomings of existing algorithms.This paper focuses on the practical problems faced by flying infrared target detection in lowaltitude complex scenes,and combines image processing,computer vision,deep learning and other algorithms to conduct in-depth research.The main work and innovations are as follows:(1)An infrared target simulation algorithm based on C-DCGAN is proposed to realize deep migration learning based on data generation.Different from the traditional GAN network,the model organically combines the category labels and enhances the feature learning with convolutional networks,which improves the efficiency of generating multi-distributed infrared target images.It should be emphasized that the model is not a simple sampling or transformation of the existing image,but a new infrared target image generated after fully understanding the radiation characteristics of the infrared image.(2)Based on DCNN,an infrared small target detection framework is constructed to solve the problem that the background noise of the speckled target at a long distance is difficult to extract due to the undulating background clutter,and it is impossible to distinguish the highlight interference on the gradation feature.Firstly,the framework uses the regression DCNN for background component suppression and potential target enhancement,and then extracts the candidate target region by threshold segmentation,and sends them to the sub-type DCNN for target confirmation at last.The experimental results show that the proposed method is far superior to the traditional infrared small target detection method in terms of improving the signal-to-noise ratio of the image and distinguishing the false target.(3)Enhance the feature representation and positioning accuracy of the existing regression deep network detection framework to solve the problem of performance degradation caused by sparse infrared target features and low pixel ratio.After analyzing the reasons why the SSD model is not good at detecting small-area targets,a bidirectional fusion method of feature maps is proposed to enhance the expression ability of each size feature layer;and the semantic enhancement branch is introduced into the shallow high-resolution feature map.It has an abstract target representation capability and good positioning accuracy.In addition,lightweight MobileNet is used as the feature extraction network to achieve a balance between algorithm speed and detection performance.The enhanced model achieved 70.5% mAP / 23.7 FPS on the infrared target dataset,an 8.7% increase over the original SSD model,and no significant drop in speed.(4)The target detection of large field of view image faces the unfavorable factors such as surge in data volume,the more complicated background interference,and the decrease of the target pixel number in the whole picture.This paper proposes a multilevel joint detection framework.In the rough extraction stage,after the ROI is separated by local weighted entropy,the adaptive threshold of the constant false alarm rate is used to segment the candidate target regions of various sizes.In the target confirmation phase,the two high-performance target detections proposed before are integrated.The algorithm combines the two results to become the alarm information output.The model does not require prior knowledge of the target imaging morphology,has the ability to detect multi-size targets,and has advantages compared to traditional methods in overall performance,and has good practical value.