CoNet: A Consistency-Oriented Network for Camouflaged Object Segmentation

Camouflaged object segmentation (COS) is a recently emerging task due to its broad application prospect. The coloration and texture similarities between the objects and their surroundings makes it a challenging task. Motivated by this, we propose a consistency-oriented network (CoNet) to address these challenges by looking into the visual consistencies between object and background. Specifically, we design a primary detection module (PDM) to firstly locate the object by fusing the backbone features. A filter is introduced to better focus on the object’s foreground feature based on its primary location. To obtain the visual consistency between the object and background, the foreground feature is then fed into the consistency evaluation module (CEM) to interact with the global feature. Both features are simultaneously processed by a shared discriminator and then fused together to attain the consistency attention map. The final feature refinement is conducted in the detail refinement module (DRM) by merging the consistency attention map with the global features via hierarchical feature fusion. Extensive experiments on benchmark COS datasets show that the proposed CoNet outperforms the state-of-the-art (SOTA) models in most cases. Ablation experiments verify the effectiveness of different backbones, designed modules and upsampling methods. Furthermore, extra studies on the labelling techniques and interdisciplinary applications demonstrate the great potential of the proposed CoNet.