DS-PRNet: lightweight camouflaged object detection with dynamic sparse attention and feature-, attention-, and output-level distillation

Abstract

Existing camouflaged object detection (COD) models are computationally demanding, which limits their deployment on resource-constrained edge devices. To address this challenge, we propose DS-PRNet, a lightweight COD framework that integrates a Dynamic Sparse Attention (DSA) module with a feature-, attention-, and output-level distillation strategy. The DSA module employs a learnable head-selection mechanism that dynamically selects informative attention heads while suppressing redundant attention responses, preserving essential global contextual information for distinguishing subtle foreground-background differences in camouflage scenes. To compensate for the performance degradation caused by model compression, the proposed distillation framework transfers complementary knowledge from a teacher model at three levels: feature representations, attention maps, and prediction outputs. Specifically, we construct a teacher feature memory bank to preserve representative feature representations and employ differentiable attention map transformation to provide sparse spatial guidance. This design enhances feature representation consistency and alleviates feature confusion during the distillation process, enabling the lightweight student network to retain discriminative information. Extensive experiments on four benchmark datasets demonstrate that DS-PRNet substantially reduces model parameters and computational cost while maintaining competitive detection accuracy, with only a modest performance trade-off compared with the teacher model. The proposed network also achieves faster inference speed and better boundary-sensitive feature representation. Further validation on embedded hardware platforms is left as future work to confirm its practical deployability.