Towards camouflaged object detection via global guidance and cascading refinement

Abstract

Camouflaged object detection is characterized by targets with fuzzy boundaries, diverse sizes, and backgrounds similar to the target objects. Due to these characteristics, existing methods tend to improve detection performance by building very complex models while ignoring computational efficiency. In addition, the difficulty in capturing the target makes it difficult to localize, and the huge background noise of the target leads to the loss of detailed features in the process. To address the above problems, we propose an efficient Global Guidance and Cascading Refinement Network (GCNet) with a streamlined structure. Firstly, considering the model size, we employ a lightweight SMT as the backbone. Secondly, we design a Rough Position Module (RPM) to coarsely localize the target by collecting global semantic information and guiding global features to anchor near the target location with high quality. Finally, we introduce a Feature Refinement Module (FRM), which employs a reverse attention mechanism to enhance feature discrimination and helps to highlight the camouflaged regions by refining features through an efficient cascading manner. Extensive experimental results show that GCNet outperforms 20 current methods on four benchmark datasets. Importantly, GCNet boasts a low number of parameters, a low computational complexity, and a very competitive inference speed, successfully balancing the incompatibility between model size and recognition accuracy. The codes are released at https://github.com/wd61419/GCNet.