The core challenge of camouflaged object detection is the difficulty in visually distinguishing targets that are highly similar to the background. However, existing methods mostly rely on edge detection and full-image segmentation, and the detection effectiveness decreases dramatically when the target is highly similar to the background. To address this problem, we propose the Localisation-Region Generalised Attention Network (L-RGANet), which combines a Transformer feature encoder, a cyclic feedback localisation mechanism and a generalised attention module. First, the Transformer backbone network is used to extract multi-scale feature information and enhance semantic and spatial information through top-down feature fusion. Subsequently, the Loop Feedback Localisation Module calibrates the position of the camouflaged target and reduces the background interference through continuous iterative localisation. The generalised attention module simulates biological vision, dynamically adjusts the attention with the target in mind, and blurs the background outside the target region to improve the accuracy of detection. Experimental results show that LRGANet performs well on multiple public camouflage target detection datasets and significantly outperforms other existing camouflage target detection methods. In both quantitative and qualitative analyses, L-RGANet demonstrates higher accuracy and robustness in detecting camouflaged targets, especially in identifying targets better in complex backgrounds.
Localisation Region Based Generalised Attention Camouflage Object Detection Network
2024-10-23
942066 byte
Conference paper
Electronic Resource
English
Engineering Index Backfile | 1944
|NTIS | 1945
|