Polarization-driven camouflaged object segmentation via gated fusion.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Camouflaged object segmentation is a challenging task in computer vision.
• Existing polarization-based models face difficulties in effectively fusing polarization and light intensity features.
• A lack of public datasets for light intensity-polarization (I-P) camouflaged detection hinders research progress.

Purpose of the Study:

• To propose a novel multi-modal camouflaged object segmentation method using gated fusion.
• To effectively fuse polarization and light intensity features for improved segmentation accuracy.
• To address the data scarcity by building a new light I-P camouflaged detection dataset.

Main Methods:

• A spatial positioning module is employed for channel calibration and spatial attention alignment between polarization and light intensity features.
• A gated fusion module (GFM) selectively fuses object information, guided by semantic location features.
• Features from each layer are aggregated to generate the final segmentation map.

Main Results:

• The proposed gated fusion method demonstrates superior performance in camouflaged object segmentation.
• The newly constructed light I-P camouflaged detection dataset facilitates evaluation and training.
• Experimental results show the proposed method outperforms existing multi-modal segmentation techniques.

Conclusions:

• The developed gated fusion approach effectively integrates polarization and light intensity data for enhanced camouflaged object segmentation.
• The creation of the light I-P dataset provides a valuable resource for future research in this domain.
• This work advances the state-of-the-art in multi-modal camouflaged object detection.