Camouflaged Object Detection Based on Feature Aggregation and Global Semantic Learning

In camouflage object detection (COD), a large amount of contextual information is usually required because the object is very similar to its surrounding environment. However, due to the current problems of insufficient feature aggregation in the backbone network and insufficient effective interaction between high and low semantics, it is difficult to fully mine contextual information. This situation hampers the differentiation between background and targets, leading to potential misjudgments. To address these issues, we propose a feature aggregation and global semantic learning network (FAGSL-Net). We design a Feature Aggregation Enhancement Module (FAEM) to aggregate features from different levels, which progressively aggregates feature maps from high to low levels in a cascaded manner to obtain robust fused feature representations, further generating preliminary forecast chart. Then, leveraging the designed Global Semantic Fusion Module (GSFM), we perform high and low semantic interaction learning between the preliminary forecast chart containing multi-level features and low-level feature maps to comprehensively utilize semantic information from different levels to enhance camouflage object detection performance. Finally, we utilize a multi-scale strategy to mine information from global features. We fully test and evaluate on three widely used large-scale benchmark datasets, and the results show that our model outperforms existing COD models.