BANet: Camouflaged Object Detection Based on Boundary Guidance and Multiple Attention Mechanisms
Siyu Du, Cuili Yao, Yuqiu Kong, Yicheng Yang · 2023
With a wide spectrum of real-world applications, Camouflaged object detection(COD) seeks to segment objects that flawlessly fit into their surroundings. However, the inherent similarity between foreground and background makes it difficult for existing deep-learning methods to identify camouflaged objects accurately. So, we’d like to suggest a BANet for the COD task in this work. To fully integrate spatial and semantic information, we propose an attention-based spatial-semantic interaction module (SSIM) that integrates information from SwinB-Transformer and Res2Net-50. After that, a boundary generation module (BGM) based on a reverse attention mechanism is proposed to guide network learning. Finally, multiple improved channel attention (CA) and spatial attention (SA) are combined to generate prediction from the boundary and extracted features. Using four commonly used benchmark datasets, we perform extensive experiments and evaluate our strategy against other state-of-the-art (SOTA) models. The outcomes reveal that the suggested BANet works well and makes a significant improvement.