Knowledge Distillation with Geometry-Consistent Feature Alignment for Robust Low-Light Apple Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Technology

Background:

• Low-light conditions significantly degrade apple detection accuracy in orchards due to noise and non-uniform exposure.
• Edge cues essential for precise localization are blurred, hindering automated harvesting and monitoring systems.

Purpose of the Study:

• To develop a compact, end-to-end framework for robust apple detection under low-light conditions.
• To bridge the illumination domain gap while preserving critical geometric information for improved localization.

Main Methods:

• Proposes Knowledge Distillation with Geometry-Consistent Feature Alignment (KDFA), integrating image enhancement and detection.
• Employs Cross-Domain Mutual-Information-Bound Knowledge Distillation to align features between daylight and low-light images.
• Utilizes Geometry-Consistent Feature Alignment with Laplacian smoothness and bipartite graph correspondences across feature lattices.

Main Results:

• Achieved 51.3% mean Average Precision (mAP) on a challenging low-light apple detection benchmark.
• Set a new state-of-the-art performance, outperforming existing methods in low-light scenarios.
• Demonstrated effective bridging of the illumination domain gap and preservation of geometric consistency.

Conclusions:

• KDFA framework significantly improves apple detection performance in low-light orchard environments.
• The proposed method offers a robust solution for precision agriculture applications requiring accurate object localization under adverse lighting.
• Future work can explore adaptability to other agricultural object detection tasks and varying illumination conditions.