ESINET: Enhancing Structural Integrity in Scribble-Supervised Camouflaged Object Detection

Scribble-supervised Camouflaged Object Detection (SCOD) aims to leverage sparse and coarse scribble annotations to detect objects that seamlessly blend into the background. Despite recent progress, current SCOD methods still face challenges with insufficient pixel-level supervision and poor structural integrity, leading to inaccurate segmentation. To address these issues, we propose an Enhanced Structural Integrity Network (ESINet), a one-stage method consisting of a Directional Context Attention (DCA) module and an Asymmetric Local Structure Consistency (ALSC) loss function. Specifically, DCA globally models horizontal and vertical dependencies through adaptive fusion to enhance structural representation, while ALSC propagates supervision from reliable pixels to uncertain pixels through confidence-aware directional constraints. Extensive experiments on multiple SCOD benchmarks demonstrate that ESINet consistently outperforms the state-of-the-art one-stage Weakly Supervised Camouflaged Object Detection (WSCOD) methods, validating its effectiveness and robustness.