A ternary encoding network fusing scale awareness and large kernel attention for camouflaged object detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing camouflaged object detection methods struggle with structural information loss and object occlusion in complex scenes.
• These limitations hinder accurate detection and segmentation of camouflaged objects.

Purpose of the Study:

• To propose a novel network, SALK-Net, that addresses information loss and occlusion in camouflaged object detection.
• To enhance the perception of global semantic information and minimize the loss of critical clues.
• To improve the prediction of challenging pixels, particularly at object boundaries.

Main Methods:

• The proposed SALK-Net integrates scale awareness and enhanced large kernel attention.
• It utilizes ternary images as input to mine multi-scale information.
• A shared feature encoder, enhanced large kernel attention for feature fusion, and a hybrid-scale decoder are employed.
• A dynamic weighting strategy for boundary and structure is incorporated into the loss function.

Main Results:

• SALK-Net was compared against 12 state-of-the-art methods on 4 public datasets.
• The method achieved high performance, with structural similarity reaching 0.861 (trained) and 0.872 (untrained).
• Enhanced alignment measures were 0.927 (trained) and 0.926 (untrained), demonstrating robustness.

Conclusions:

• SALK-Net significantly outperforms existing methods in camouflaged object detection.
• The integration of scale awareness and attention mechanisms effectively mitigates information loss.
• The proposed network shows strong generalization capabilities on both trained and untrained datasets.