Research on Recognition Method of Infrared Dim Small Flying Targets Based on Multi-layer LBP Features

This paper proposes a method that combines the pre-trained VGG19 model with the local texture description method of Local Binary Pattern (LBP) to improve the feature description performance of infrared dim small targets. Specifically, LBP texture features are extracted from the outputs of 16 convolutional layers of the VGG19 network, using Mean LBP to enhance robustness to image brightness changes caused by the impact of atmospheric environmental. For recognition, a framework based on sparse representation theory is employed. This includes dictionary learning, sparse coding, classifier training, and target classification. The Orthogonal Matching Pursuit (OMP) method is used for sparse coding, and Support Vector Machine (SVM) is adopted as the classification model. The proposed method is evaluated using a dataset simulating four types of aerial targets under four typical meteorological conditions. A total of 1143 images are generated through simulation, and the five-fold cross-validation method is used to assess performance. Finally, the classifier achieves an average precision of 81.34%.