Deep Learning And Target Detection For Weak Sample Infrared Images With Space-time Constraints

Infrared thermal imaging technology uses the infrared electromagnetic wave band actively emitted by the object to perceive the difference between different targets and the background according to the difference of the surface temperature of the object for imaging.Because of this feature,the infrared target detection system is more adaptable to complex environments and is suitable for long-distance observation in the military and aerospace fields.Infrared target detection,as the most important technical means for monitoring abnormal targets in the military and aerospace fields,is of great significance for maintaining the security and stability of national defense.Deep learning is widely used in target detection because it can extract more abstract deep features.The ground infrared images often show weak samples and low quality due to the particularity of the data source and the abnormal equipment during the imaging process.In addition,the infrared image data collection often shows sequentiality due to the motion properties of the infrared target.Therefore,the infrared sequence of image target detection has certain challenges.In this paper,the space-time constrained target detection algorithm is researched on infrared sequence image data.The main contents of the research are as follows:1?Researched the basic theories of target detection and deep learning,including the basic theories of space-time constraints,convolutional layers,pooling layers and activation functions of convolutional neural networks,and video target detection networks;2?Aiming at the problem of noise and low contrast of space-borne infrared images,we analyze the causes of image quality degradation to study and propose space-based infrared satellite image preprocessing methods,including image denoising based on guided filtering and adaptive median filtering And the contrast enhancement algorithm based on the equalization of the restricted contrast histogram.Aiming at the problem of the small sample size of infrared images,by studying the weak sample expansion methods commonly used in deep learning,an infrared image data expansion method with Gaussian noise,random occlusion and affine transformation is proposed;3?Aiming at the problems of motion blur and low accuracy in infrared sequence image detection,based on the idea of space-time constrained target detection,a HLPANet network detection algorithm based on Kalman filter constraints is proposed,which includes cross-stage feature combinations.Backbone network,attention module and feature fusion module.In the case of using CSFRes Net101,HLPA module and cascaded feature pyramid,the infrared test data set m AP in this article can reach 0.68;4?Aiming at the problem that the previously designed space-time constraint detection algorithm process is too complex,by analyzing the advantages and disadvantages of the video target detection network model,research and design a video target detection network that covers global and neighborhood information enhancement,and realizes infrared sequence The end-to-end detection of the graph,the m AP in the test set of this article can reach 0.708;...