A Comparison Study of Different Deep Neural Networks on Aerial Target Intention Recognition

Deep learning-based methods for aerial target intention recognition have emerged, yet few have systematically compared the performance of fundamental network structures in this context. This study explores the application and performance of various basic deep learning network structures for aerial target intention recognition. Findings indicate that the fully connected network, relying solely on linear layers, performs poorly. Among recurrent neural network methods, gated recurrent unit outperforms long short-term memory. Additionally, one-dimensional convolutional neural network demonstrates superior recognition capabilities. Experiments further reveal that integrating convolutional and recurrent network structures significantly enhances intention recognition performance. Notably, the model utilizing the Transformer Encoder achieves the best results across all tests, highlighting the substantial potential and utility of the self-attention mechanism in addressing aerial target intent recognition challenges. Overall, this research provides insightful contributions to the field of aerial target intention recognition and offers vital references for the design and optimization of future deep learning models.