A Research On Infrared Small Target Detection Based On Prior Knowledge And Deep Learning

In infrared search and track systems,infrared small target detection in complex backgrounds plays a vital role.Nevertheless,problems like background clutter interference and the limited detection range of IR sensors result in low signal-to-noise ratios,very small target sizes,and a lack of feature information for infrared small target.Also,the environment’s complexity creates many difficulties and challenges for detection.Traditional detection methods have been developed for a long time,but still face the problems of noise and clutter interference sensitivity and high computational complexity.Deep learning methods do not perform well because small targets have fewer pixels and lack features that can be extracted in multiple layers.In this thesis,based on the traditional tensor low-rank sparse decomposition theory and data-driven deep learning target detection techniques,we study the high-dimensional tensor modeling and decomposition of infrared small target,and also investigate the deep learning infrared small target detection model and the fusion of traditional features with network models to complement the pure data-driven defects and improve the accuracy and robustness of infrared small target detection.The following are the key components of the research.The key theories of low-rank sparse tensor decomposition are investigated,including tensor decomposition theories and methods for optimization solution.At the same time,the basic theory of deep learning target detection and various basic structures of neural networks are organized to build the foundation for the next model design and algorithm research.Based on the characteristic analysis of infrared small target images,the traditional 3-D tensor model is extended to a 4-D spatial-temporal tensor model for the first time,adding temporal information while maintaining the spatial structure of the data.The Ntubal rank of the high-dimensional tensor is used as the rank approximation of the background tensor,fully utilizing the information between different modes of the tensor for better background estimation.The model is compared with 10 competing methods in five scenes.The comparison experiments demonstrate the best background suppression and target enhancing capabilities.Using residual feature coding and a priori loss,an infrared tiny target detection algorithm is created.To solve the feature scarcity issue of the data-driven approach,the local contrast prior is merged into the neural network by fusing the local residual module,and the feature pyramid module is employed to broaden the deep network’s perceptual area.In order to perform end-to-end segmentation model training with point label,the Gaussian label weighted loss is utilized in the model with the features of the infrared tiny target.The experimental results show that the model has outstanding detection capability and achieves optimal detection in all six scenarios.