| Infrared small target detection plays an important role in many areas such as airborne guidance,tracking and surveillance.However,detection of small infrared targets remains a challenging task due to their small number of pixels,their tendency to be confused with background noise,and their lack of clear contour and texture information.As traditional IR small target detection methods rely on the designer’s a priori knowledge and manual tuning of parameters,they result in poor detection for changing scenes and complex backgrounds.Deep learning methods based on feature enhancement can automatically learn potential target features from a large amount of sample data and achieve better performance for IR small target detection.However,the current deep learning-based IR small target detection methods still suffer from insufficient feature extraction and loss of depth features,so the innovative research results in this thesis address the above problems as follows.(1)An infrared small target detection algorithm based on multi-scale feature fusion attention is proposed.A detection network using complementary fusion of multi-level shallow and deep features is proposed to address the problems of poor generalization and poor detection performance of general infrared small target detection methods.The method employs a convolutional attention module to enhance the target features from channel and spatial dimensions for the multi-level features obtained from the feature extraction network.For highlevel features the multi-scale perceptual field feature fusion module is designed to characterize the contextual information of the feature map as a way to expand the perceptual field for small targets.Finally,a bidirectional feature aggregation network is designed for the enhanced multiscale features to shorten the information exchange path between different layers and enhance the information reuse of the fused features.The experimental results show that the F1 and detection accuracy are 64% and 65.57%,respectively,on the open dataset of infrared small target detection,which are better than other target detection methods and suitable for infrared small target detection tasks.(2)A context-aware cross-level attention-based infrared small target detection algorithm is proposed.A network for feature extraction using long-range contextual information is constructed to address the problems of unsatisfactory small target feature extraction and insufficient contextual information.The method first proposes a self-attention-induced global context-aware module in the backbone network to obtain a multi-level attentional feature map by modeling long-range location relationships.The high-level feature maps with rich semantic information are then passed through the proposed multi-scale feature refinement module to recover target details and highlight salient features.The various level feature maps are compressed with redundant information and removed from background noise by the proposed channel and spatial filtering modules,and then used for cross-level feature fusion.In addition,we develop a multi-scene large-scale infrared small target detection dataset with high-quality annotation to alleviate the current problem of few publicly available datasets in the field of infrared small target detection.The experimental results show that the F-measures of 85.43%,65.86%,and 76.44% on the publicly available SIRST and MDFA and the self-built MIRST datasets,respectively,outperform the remaining 14 traditional detection methods and other advanced deep learning-based detection methods.The method effectively uses contextual information to solve the problem of insufficient feature extraction of small infrared targets and achieves a significant improvement in detection performance.(3)An infrared small target detection algorithm based on global interaction graph attention is proposed.A network capable of modeling the relationship between target pixels and the rest of the global position is constructed for the problem of sparse features and low contrast with the background of infrared small targets.The method first constructs a dimensional interaction graph attention layer in the shallow network to fully learn the global structure relationship of the small target,and three feature branches by cross-dimensionality to enhance the interdimensional dependence in the feature extraction phase.This is followed by learning the similarity relationships among feature voxels dynamically with graph attention.Finally,to complement the detailed features of the deep network,a multi-scale context fusion module is used for higher-order features,and the two-dimensional self-attentive module is used to model the global range of inter-pixel relationships and the null convolution to fuse the local perceptual fields at different scales.The experimental results show that the F-measurements of 84.94%and 61.7% on the publicly available SIRST and MDFA,respectively,outperform the remaining14 traditional detection methods and other advanced deep learning-based detection methods,effectively solving the problem of easy confusion between target and background. |
