CL-CaGAN: Capsule differential adversarial continuous learning for cross-domain hyperspectral anomaly detection

TL;DR

This work tackles hyperspectral anomaly detection in open, cross-domain settings where prior information is limited and models suffer from catastrophic forgetting. It introduces CL-CaGAN, a continual-learning capsule-based GAN that reconstructs background via CaGAN while detecting anomalies through reconstruction error, augmented by a clustering-based replay buffer and a continual self-distillation loss to merge historical and new knowledge; differentiable data augmentation further stabilizes training. The framework combines a capsule-enhanced generator and discriminator, a coarse spectral-spatial background search, and a self-distillation objective, with $L_G$, $L_D$, $L_{recon}$, and $L_{CSD}$ guiding training. Experiments on five AVIRIS datasets show superior cross-domain HAD performance and robust forgetting control, indicating strong practical potential for deploying DL-based HAD in open-world, multi-scenario environments.

Abstract

Anomaly detection (AD) has attracted remarkable attention in hyperspectral image (HSI) processing fields, and most existing deep learning (DL)-based algorithms indicate dramatic potential for detecting anomaly samples through specific training process under current scenario. However, the limited prior information and the catastrophic forgetting problem indicate crucial challenges for existing DL structure in open scenarios cross-domain detection. In order to improve the detection performance, a novel continual learning-based capsule differential generative adversarial network (CL-CaGAN) is proposed to elevate the cross-scenario learning performance for facilitating the real application of DL-based structure in hyperspectral AD (HAD) task. First, a modified capsule structure with adversarial learning network is constructed to estimate the background distribution for surmounting the deficiency of prior information. To mitigate the catastrophic forgetting phenomenon, clustering-based sample replay strategy and a designed extra self-distillation regularization are integrated for merging the history and future knowledge in continual AD task, while the discriminative learning ability from previous detection scenario to current scenario is retained by the elaborately designed structure with continual learning (CL) strategy. In addition, the differentiable enhancement is enforced to augment the generation performance of the training data. This further stabilizes the training process with better convergence and efficiently consolidates the reconstruction ability of background samples. To verify the effectiveness of our proposed CL-CaGAN, we conduct experiments on several real HSIs, and the results indicate that the proposed CL-CaGAN demonstrates higher detection performance and continuous learning capacity for mitigating the catastrophic forgetting under cross-domain scenarios.

Figures (10)

• Figure 1: Comparison of different DL training model and the continuous learning method. (a) represents the traditional deep learning method, which obtains anomaly detection results by a set of independent well-trained parameters. (b) represents the joint learning method, which combines all data to train only one set of parameters for anomaly detection. (c) represents the fine-tuning method. The initialization of model parameters is based on the previous set of training parameters. (d) represents the proposed continuous learning method. The parameters of the model are continuously updated with the arrival of data, but the updated parameters will not forget the previously learned knowledge.
• Figure 2: The overview flowchart of the proposed CL-CaGAN for open scenario HAD. The entire CL-CaGAN for continuous anomaly detection process is mainly divided into three parts. 1. The cluster-based replay strategy: the representative pixels are retained in the task flow and in subsequent data to prevent forgetting phenomenon. 2. The proposed CaGAN framework: the specific structure for continuous learning AD task with cascaded generators and discriminators. 3. The proposed self-distillation loss function $L_{CSD}$ is designed to constrain the magnitude of parameter updates for preventing catastrophic forgetting.
• Figure 3: Overview of the proposed CaGAN for HAD. The CaGAN structure mainly contains three components.1.The coarse spectral-spatial background searching. 2.The Generator structure. 3.The Discriminator structure.
• Figure 4: The Generator structure of the proposed CaGAN includes a cascaded Encoder and Decoder, where the Encoder consists of a capsule network, including two layers: Primary-G and Capsule-G. $Z_G$ in Capsule-G represents the number of capsule groups in the Generator, and $K_G$ represents the number of capsules in each group of Generator.
• Figure 5: The discriminator structure of the proposed CaGAN. Including Primary-D and Capsule-D. Primary-D consists of convolution operations with different convolution kernel sizes. $Z_D$ in Capsule-D represents the number of groups of capsules in the Discriminator, and $K_D$ represents the number of capsules in each group of Discriminators.