Hybrid-Recursive-Refinement Network for Camouflaged Object Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Camouflaged object detection (COD) is challenging due to subtle textures and semantic ambiguity.
• Existing Convolutional Neural Network (CNN) or Transformer models struggle with incomplete feature representation and boundary detail loss.

Purpose of the Study:

• To develop an innovative hybrid architecture for improved camouflaged object detection.
• To enhance feature representation and boundary detail accuracy in detecting concealed objects.

Main Methods:

• Proposed a hybrid architecture integrating CNNs and Transformers.
• Devised a Hybrid Feature Fusion Module (HFFM) to harmonize hierarchical features.
• Designed a Combined Recursive Decoder (CRD) for adaptive feature aggregation and multi-scale detail capture.
• Introduced a Foreground-Background Selection (FBS) module for progressive contour refinement and background suppression.

Main Results:

• The hybrid architecture achieved state-of-the-art performance on four public COD datasets (CHAMELEON, CAMO, COD10K, NC4K).
• The proposed modules (HFFM, CRD, FBS) significantly boosted representational quality and structural detail capture.
• The method demonstrated superior performance in precisely detecting and delineating camouflaged objects.

Conclusions:

• The proposed hybrid CNN-Transformer model offers a significant advancement in camouflaged object detection.
• The synergistic integration of different network strengths overcomes limitations of individual architectures.
• This approach provides a robust solution for accurately identifying objects in complex, ambiguous backgrounds.