Deep Learning Architecture with Adaptive Attention and Multi-Scale Fusion for Infrared Spectrum Target Recognition

Abstract

With growing demands for accurate infrared spectrum analysis in industrial, military, and medical applications, traditional methods typically cannot meet the requirements due to limited feature extraction and recognition. This article proposes a novel deep learning model featuring an adaptive attention module, a multi-scale feature fusion module, and a classification decision module, designed to enhance performance. The model is trained using a cross-entropy loss function and learns with backpropagation, employing an exponential decay learning rate policy, over more than 100 training epochs. Experiments are run on three test datasets: NATO RTO SET-103, Thermal IR Benchmark, and FLIR Thermal. The model achieved an average feature extraction accuracy of 90.8% and a target recognition accuracy of 89.7%, which significantly surpassed those of traditional models, such as DenseNet, ResNet, VGGNet, and Basic CNN. The performance was robust in the face of changing data distributions, demonstrating high generalizability and robustness. The result substantiates the model's capability of accurately extracting important infrared features and recognizing targets with high accuracy. This work presents an effective solution to real-world problems in infrared spectrum analysis.