MAE-YOLO improves small object detection for intelligent inspection | JoVE Visualize

Area of Science:

Computer Vision
Artificial Intelligence
Machine Learning

Background:

Intelligent inspection technology is crucial for industrial maintenance, traffic management, and equipment upkeep, focusing on small, critical targets.
Traditional deep learning models like YOLO struggle with high false detection and miss rates for small objects in complex backgrounds, hindering real-time applications.
Existing lightweight models often compromise edge detail for efficiency, limiting their effectiveness in precise industrial inspection tasks.

Purpose of the Study:

To propose MAE-YOLO, a novel intelligent inspection model designed for accurate and efficient detection of small objects in challenging industrial scenarios.
To enhance the detection accuracy and robustness of small object identification while maintaining a lightweight model architecture for real-time performance.
To improve the trade-off between model size and detection accuracy compared to existing state-of-the-art lightweight object detection methods.

Main Methods:

Developed a multi-scale edge space feature extraction module (MSESTE) to optimize edge and spatial feature extraction for improved small target detection.
Introduced an adaptive multi-scale context fusion network (AMCFN) to effectively integrate features across different scales, enhancing model adaptability.
Proposed an adaptive cavity shared detection head (ELCIN) to reduce false and missed detections, optimizing parameter efficiency.

Main Results:

MAE-YOLO achieved a 2.6% accuracy improvement over the original YOLOv8n model on the VisDrone2019 dataset.
The model size was reduced by 24% to 4.7 MB compared to YOLOv8n on a self-collected dataset, while maintaining high detection accuracy.
MAE-YOLO demonstrated a superior accuracy-size trade-off (35.1% mAP@50 at 4.57 MB), outperforming SOD-YOLO (30.08% mAP@50) by preserving fine-grained object boundaries.

Conclusions:

MAE-YOLO effectively addresses the limitations of traditional methods in small object detection within complex industrial environments.
The proposed model offers a compelling solution for real-time intelligent inspection, balancing high accuracy with a significantly reduced model footprint.
The integration of edge-aware feature extraction and efficient detection heads provides a robust and adaptable framework for diverse industrial applications.