COS-former: hierarchical fusion transformer for camouflaged object segmentation

Camouflaged object segmentation aims to accurately delineate objects that are seamlessly integrated into their surrounding environment. The intrinsic similarity between foreground objects and the background environment poses a challenge for current deep models in extracting distinctive features. To address this issue, we have introduced a Transformer framework for camouflaged object segmentation, known as COS-Former, which is designed to extract and integrate multilevel features. Our approach utilizes a simple multilevel feature extraction encoder based on a general Transformer block to capture global context information learned by different levels of Transformer blocks. Leveraging the self-attention mechanism of Transformers enables the establishment of global perception while retaining local feature information. To decode features from the Transformer, we have developed a feature conversion decoder comprising two modules: hierarchical feature fusion (HFF) and object and edge fusion (OEF). HFF focuses on integrating multilevel features from each transformer layer, whereas OEF takes into account both object features and edge information, incorporating camouflage object edge supervision to enhance features for improved localization of camouflage object edges. The experimental results demonstrate that our model achieves competitive performance on the four benchmark datasets. Compared with other SOTA methods, the average maximum improvement of S-measure, E-measure, weighted F-measure, and mean absolute error is up to 9.8%, 9.6%, 16.6%, and 2.9%, respectively.