A Sparsity-Driven Lightweight Generative Adversarial Network via Structural Channel Pruning for Hyperspectral Anomaly Detection
Abstract
Recently, deep learning has become a dominant paradigm for hyperspectral anomaly detection (HAD), providing enhanced feature representation capability and improved detection accuracy. However, these performance gains are often accompanied by larger parameter scales and higher computational costs. This trend poses substantial challenges for deployment on resource-constrained edge platforms. To address this problem, this article proposes a lightweight generative adversarial network (GAN) for efficient HAD. The proposed model learns representative background distributions from hyperspectral images and reconstructs a pure background through a generative adversarial framework. Anomalous targets are then detected by analyzing the residual differences between the original hyperspectral image and the generated background. To obtain a lightweight detector, L1 regularization and soft-threshold shrinkage (STS) are applied to batch normalization (BN) scaling factors to promote channel sparsity. Channels with scaling factors below the predefined threshold are structurally pruned, thereby suppressing redundant feature responses and further reducing the model complexity. Experimental results demonstrate that the proposed method reduces both the parameter count and computational cost by more than 90% while maintaining nearly unchanged detection accuracy.