Chen Geng ; Liu Si-Jie ; Sun Yu-Jia ; Ji Ge-Peng ; Wu Ya-Feng ; Zhou Tao

Abstract

Camouflaged object detection (COD) aims to identify the objects that concealthemselves in natural scenes. Accurate COD suffers from a number of challengesassociated with low boundary contrast and the large variation of objectappearances, e.g., object size and shape. To address these challenges, wepropose a novel Context-aware Cross-level Fusion Network (C2F-Net), which fusescontext-aware cross-level features for accurately identifying camouflagedobjects. Specifically, we compute informative attention coefficients frommulti-level features with our Attention-induced Cross-level Fusion Module(ACFM), which further integrates the features under the guidance of attentioncoefficients. We then propose a Dual-branch Global Context Module (DGCM) torefine the fused features for informative feature representations by exploitingrich global context information. Multiple ACFMs and DGCMs are integrated in acascaded manner for generating a coarse prediction from high-level features.The coarse prediction acts as an attention map to refine the low-level featuresbefore passing them to our Camouflage Inference Module (CIM) to generate thefinal prediction. We perform extensive experiments on three widely usedbenchmark datasets and compare C2F-Net with state-of-the-art (SOTA) models. Theresults show that C2F-Net is an effective COD model and outperforms SOTA modelsremarkably. Further, an evaluation on polyp segmentation datasets demonstratesthe promising potentials of our C2F-Net in COD downstream applications. Ourcode is publicly available at: https://github.com/Ben57882/C2FNet-TSCVT.