Adltformer Team-Training with Detr: Enhancing Cattle Detection in Non-Ideal Lighting Conditions Through Adaptive...

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Technology

Background:

• Cattle detection in real pastures is hindered by complex lighting (backlighting, low light, non-uniformity).
• Suboptimal lighting degrades image quality, causing detail loss, color distortion, and noise, reducing detection model accuracy.
• Existing methods struggle with the dynamic and unpredictable nature of natural farming environments.

Purpose of the Study:

• To develop an advanced image enhancement and detection technique for improving cattle identification accuracy under adverse lighting.
• To address the limitations of current methods in real-world farming scenarios.
• To enhance the compatibility of enhanced images with machine vision systems.

Main Methods:

• Proposed an Adaptive dynamic learning transformer (Adltformer) for image enhancement.
• Utilized a Detection transformer (Detr) for object detection.
• Implemented a team-training approach (AT-Detr) combining Adltformer and Detr.
• Employed a day-to-night image synthesis (DTN-Synthesis) algorithm for generating realistic low-light training data.

Main Results:

• The AT-Detr algorithm achieved a detection accuracy of 97.5% under challenging illumination conditions.
• Outperformed Detr alone and sequential enhancement-then-detection methods in accuracy.
• Demonstrated comparable runtime and model complexity to existing approaches.
• Enhanced images showed improved compatibility with machine vision requirements.

Conclusions:

• The AT-Detr approach offers a robust solution for cattle detection in complex, real-world farming environments.
• This method provides significant improvements in accuracy without compromising computational efficiency.
• The study validates the practical applicability and theoretical underpinnings of AI-driven image enhancement for agricultural monitoring.