Fast camouflaged object detection via multi-scale feature-enhanced network

The aim of camouflaged object detection (COD) is to identify objects that are hidden or camouflaged in the visual scene. Since camouflaged objects have fuzzy boundaries and are very similar to their surroundings, the task of COD, especially multi-scale COD, is still challenging. Based on ERRNet, we proposed a multi-scale feature-enhanced network (MSFENet). Specifically, we have developed a multi-scale feature enhancement module (MFEM), which adopts a coarse-to-fine manner to improve the ability of a single layer to represent multi-scale information. This module can extract more complete large-scale target feature information and retain much more small-scale target feature information and less regional background information. The experimental results on publicly available datasets show that our proposed MSFENet outperforms 10 mainstream methods. The ablation studies show that the proposed module is effective in improving the detection performance of multi-scale camouflaging objects and improving the overall performance. Compared with ERRNet, the average \(S_\alpha \) $S_{\alpha }$ , \(E_\phi \) $E_{\varphi }$ and \(F_\beta ^W \) $F_{\beta }^W$ scores of the MSFENet are 1.2%, 0.6% and 2.5% higher for the multi-scale COD task. In addition, the proposed MSFENet can be directly used for real-time detection due to its fast inference capability (i.e., 75.3 frames per second).