Efficient Camouflaged Object Detection via Knowledge Distillation and Lightweight Network Design

Camouflaged object detection （COD） aims to accurately identify objects concealed within complex backgrounds and is a critical factor affecting the performance of computer vision systems. It has significant application value in intelligent security surveillance， environmental perception， and smart monitoring. With the rapid development of deep learning， deep neural network （DNN）-based COD algorithms have achieved remarkable performance improvements. However， existing DNN-based COD methods typically rely on models with a large number of parameters， leading to low inference efficiency and high computational cost. To address these challenges， this study proposes EfficientCOD， a highly efficient and lightweight COD framework. The core lightweight modules of EfficientCOD include： （1） the cross-scale asymmetric depthwise convolutional shuffle-selection module （CS-ADCS²M）， which enables efficient cross-scale feature interaction with a reduced number of parameters， thereby enhancing the model’s capability to represent camouflaged objects at varying scales； and （2） the hierarchical multi-head asymmetric depthwise convolutional shuffle-selection module （HMH-ADCS²M）， which facilitates hierarchical context encoding and semantic fusion during decoding， strengthening boundary discrimination and fine-grained texture perception between camouflaged objects and their surroundings. Moreover， because lightweight models generally have weaker feature representation capability than large-capacity models， we further design a structure-aware and semantic-consistent distillation framework for COD （SaScD-COD）. Using the high-capacity MambaCOD as the teacher network， SaScD-COD performs multi-stage knowledge transfer from fine-grained structural cues to high-level semantic representations. This enables EfficientCOD to inherit the strong representation ability of MambaCOD without increasing inference cost， thereby achieving a superior balance between detection accuracy and model efficiency. Extensive experiments on the CHAMELEON， CAMO， COD10K， and NC4K benchmarks demonstrate that EfficientCOD achieves comparable performance to the 47.8M-parameter teacher model MambaCOD while using only 31.4% （15.01M） of its parameters， delivering strong detection accuracy and computational efficiency. These results establish EfficientCOD as a new paradigm for efficient COD.