Enhanced Object Detection in Thangka Images Using Gabor, Wavelet, and Color Feature Fusion.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Heritage

Background:

• Thangka image analysis presents challenges due to complex iconography, dense small-scale elements, and unique color-texture compositions.
• Existing object detection models often fail to capture both global structures and local details in such specialized imagery.
• Domain-specific visual priors are rarely utilized in current detection frameworks for cultural heritage artifacts.

Purpose of the Study:

• To develop an advanced object detection framework tailored for the specific challenges of Thangka images.
• To enhance the model's ability to detect both global structures and fine-grained local details.
• To incorporate domain-specific visual priors, including chromatic and directional cues, into the detection process.

Main Methods:

• Proposed a frequency- and prior-enhanced detection framework based on YOLOv11.
• Introduced a Learnable Lifting Wavelet Block (LLWB) for frequency decomposition and multi-scale fusion.
• Designed a Color-Gabor Block (CGBlock) and Color-Gabor Cross Gate (C2CG) for incorporating chromatic and directional features.
• Redesigned detection heads with decoupled branches, center-ness prediction, and an additional shallow head for small targets.

Main Results:

• Achieved 89.5% mAP@0.5 and 59.4% mAP@[0.5:0.95] on a curated Thangka dataset.
• Reached a recall of 84.7%, surpassing baseline detectors.
• Maintained a compact model size of 20.9 million parameters.

Conclusions:

• The proposed framework offers a robust and interpretable solution for fine-grained object detection in complex heritage images.
• Frequency decomposition and domain-specific prior integration significantly enhance detection performance.
• The model effectively addresses challenges posed by intricate details and small objects in Thangka art.