Kang Ruan

Camouflaged Object Detection (COD), which aims to accurately segment objects that perfectly blend into the surrounding environment, is a complex task. Although many deep learning-based COD methods have emerged, they are still unable to effectively segment camouflaged objects from the background in a complete and fine-grained manner. Based on this, we will propose a novel network DCEGNet (Deformable Convolutional and Edge-Guided Networks) for camouflaged object detection using deformable convolution as well as edge-attached semantic features. Specifically, in order to address the situation that complex objects are difficult to be segmented effectively, we design a deformable convolution based on a self-attention mechanism to change the shape of the convolution kernel, on top of which we also propose an Edge Detection Module (EDM) to explore the additional edge semantic information to guide the feature learning of the COD, which enables the model to utilize the boundary information to accurately locate and segment the object. Extensive experiments on three COD benchmark datasets show that our DCEGNet significantly outperforms 20 existing methods on four widely used evaluation metrics.