CSFIN: A lightweight network for camouflaged object detection via cross-stage feature interaction

Camouflaged object detection (COD) aims to identify the objects that are hidden in their surroundings, which is a very challenging task due to factors like complex contours and high similarity to the background. Existing COD methods often face the issue of neglecting the impact of feature transformation after the backbone, which leads to accuracy constraints in the subsequent decoding step. Additionally, most COD methods typically demand a greater number of parameters and higher computational complexity to achieve notable performance. To tackle the above issues, a lightweight cross-stage feature interaction network (CSFIN) is proposed in this paper. In the CSFIN, a cross-stage feature perception neck (CSFPN) is designed to achieve effective feature transformation, which improves each stage backbone feature with its adjacent stage information, thus obtaining more comprehensive feature representations for the final decoding. In CSFPN, a bi-directional feature interaction module (BFIM) is introduced between two adjacent stage features to deeply aggregate informative camouflaged cues. As the key component of the BFIM, a multi-scale cross-attention modulation block is constructed to fully investigate the multi-scale long-range dependency between two input features, which can effectively incorporate external guidance. Extensive experimental results illustrate that the CSFIN outperforms other COD models with fewer parameters and lower computational complexity.