KDTGAN：基于Transformer-GAN和知识蒸馏的高光谱目标检测

高光谱目标检测在地球观测中至关重要，被广泛应用于军事和民用领域。然而，由于高光谱图像的背景复杂性和目标样本的有限性，该任务面临较大的挑战。本文首先采用CEM（约束能量最小化）粗检测方法提取背景数据。随之，引入了一种新的知识蒸馏模型，即KDTGAN（通过Transformer-GAN实现）。教师模型的生成器采用了Transformer编码器的结构，并结合多尺度数据融合的方法，能够准确地学习背景分布，进而通过重构背景信息实现目标检测。为了克服GAN（生成对抗网络）训练不稳定的挑战，特别是纯背景数据的稀缺性，本文提出了一种新的损失算法，以减小可疑目标样本对模型性能的负面影响。为了降低模型的计算负担，本文引入知识蒸馏，并设计新的蒸馏损失对学生模型加以约束，使模型轻量化的同时提高学生模型检测精度。实验结果表明：KDTGAN相较于当前检测方法表现更优，具有更高的检测精度和鲁棒性。

KDTGAN： Knowledge Distillation via Transformer GAN for Hyperspectral Target Detection

Hyperspectral target detection is crucial in Earth observation for both military and civilian applications. However, complex backgrounds and the scarcity of target samples pose challenges in hyperspectral image analysis. In this paper, we first employ the CEM coarse detection method to extract background data. Subsequently, a novel knowledge distillation model, namely KDTGAN (implemented through Transformer-GAN), is introduced. The generator of this teacher model adopts the structure of a Transformer encoder and combines it with a multi-scale data fusion approach to accurately learn the background distribution, which in turn enables target detection by reconstructing the background information. To overcome the challenge of unstable GAN training, especially the scarcity of pure background data, we propose a new loss algorithm to reduce the negative impact of suspicious target samples on model performance. To reduce the computational burden of the model, we introduce knowledge distillation and design a new distillation loss to constrain the student model to lighten the model while improving the student model''s detection accuracy. The experimental results show that KDTGAN performs better than current detection methods with higher detection accuracy and robustness.