VM-RTDETR: Advancing DETR with Vision State-Space Duality and Multi-Scale Fusion for Robust Pig Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Agricultural Technology

Background:

• Pig detection in intelligent livestock farming faces challenges with global context and multi-scale features in complex environments.
• Existing methods struggle to efficiently capture long-range dependencies and varied scale features crucial for accurate detection.

Purpose of the Study:

• To develop an advanced pig detection model for intelligent livestock farming.
• To enhance feature representation by integrating global context and multi-scale information.

Main Methods:

• Proposed VM-RTDETR model based on an enhanced RT-DETR architecture.
• Incorporated a Vision State-Space Duality (VSSD) backbone with Non-Causal State-Space Duality (NC-SSD) for long-range dependencies.
• Designed a Multi-Scale Efficient Hybrid Encoder (M-Encoder) using parallel convolutional kernels for multi-scale feature extraction.

Main Results:

• VM-RTDETR significantly outperformed existing mainstream detectors on a custom pig farm dataset.
• Achieved improvements in Average Precision (AP), AP50, and AP75 by up to 2.35%, 0.63%, and 2.76% over the R50-RTDETR baseline.
• Demonstrated enhanced detection robustness in complex farming scenarios.

Conclusions:

• The VM-RTDETR model offers a more comprehensive feature representation for pig detection.
• Provides an efficient and accurate solution for intelligent livestock farming, improving detection in challenging conditions.