Abstract
Camouflaged object segmentation (COS) is a challenging task in computer vision where the objective is to recognize and precisely separate objects that blend in with their environment. Traditional models, including the standard UNet architecture, struggle with this task due to ambiguous object boundaries, texture similarity between object and background, and over-segmentation or under-segmentation caused by redundant skip connections. CAMO-UNet addresses these issues by including residual blocks which improve feature learning by easing the gradient flow and enabling deeper architectures. The attention mechanism focuses on 'what' is important, 'where' important features are located in the spatial domain and captures long-range dependencies across the image. The Depth-aware triangular cyclic learning rate (CLR) dynamically adjusts learning rates at different network depths to enhance training efficiency. CAMO-UNet achieved 93.8% accuracy on benchmark datasets and outperformed state-of-the-art models like SINet, BGNet, PFNet, etc., in metrics including S-measure, F-measure, MAE, and accuracy.