Feature-aware and iterative refinement network for camouflaged object detection

Yanliang Ge, Junchao Ren, Cong Zhang, Min He, Hongbo Bi, Qiao Zhang

Abstract: Camouflaged object detection (COD) is engineered to identify objects using visual camouflage techniques that seamlessly blend with the background. Although the existing methods have achieved good performance, it is still difficult to detect camouflaged objects that are extremely similar to the background. In this paper, we propose a new feature-aware and iterative refinement network (FIRNet) for exploring the integrity of hidden objects. Specifically, we design a multivariate feature perception module to capture multivariate context features better to locate the original region of the camouflaged object. Furthermore, we propose an iterative refinement module to investigate the correlation among distinct features, facilitating the iterative refinement of camouflaged objects. Rigorous experimentation across four challenging benchmark datasets demonstrates that FIRNet overcomes performance bottlenecks in different scenarios, yielding notable results in comparison with 21 state-of-the-art methods. It is worth noting that FIRNet achieves a score of 0.927 on the \(F_{\beta }^\mathrm{{max}}\) on the dataset COD10K, which is 1.0% higher than that of the suboptimal FPNet method. Concurrently, we examined the significance of FIRNet on two additional COD datasets, showcasing its adaptability for diverse downstream applications. The code and results of our method are available at https://github.com/RJC0608/FIRNet.