Multi-scale Camouflaged Object Detection Method Based on Regional Feature Enhancement

Abstract

Camouflaged Object Detection(COD) can detect camouflaged objects with high similarity to the background in complex environments and is particularly important in military investigation and industrial detection. To address the low utilization of region-level feature information presented by existing COD methods, a multiscale COD method based on a Regional Feature Enhancement Network(RFE-Net) is proposed, which can accurately detect camouflaged objects under visible-light conditions. The RFE-Net includes a weak-semantic-feature enhancement module, spatial-information interaction module, and context-information aggregation module. First, the weak-semantic-feature enhancement module introduces strip pooling and asymmetric convolution to dynamically adjust the search regions by optimizing the receptive field of the network, thereby strengthening the connection between long-range weak semantic features. Subsequently, the cascading U-shaped block structure is combined into a spatial-information interaction module, which eliminates the interference of erroneous prediction samples. Finally, a context-information aggregation module is designed, which significantly improves the prediction accuracy by fully fusing deep semantic information and shallow fine-grained information to refine the object edge details. Experimental results show that the proposed method can strengthen weak semantic associations within an object and improve the distinction between camouflaged objects and the background. On the largest test set, NC4K, the proposed method achieves optimal values for four metrics: structural measure, enhanced alignment measure, weighted F1 value, and mean absolute error. In particular, the structural measure and mean absolute error are 1.1% and 7.7% higher than those of another method, respectively.