MC-YOLOv5: A Multi-Class Small Object Detection Algorithm.

Area of Science:

• Computer Vision
• Deep Learning
• Object Detection

Background:

• Detecting multi-class small objects is a significant challenge in computer vision.
• The standard YOLOv5 algorithm is not optimized for small object detection, leading to suboptimal performance.
• Existing methods struggle with precision and missed detections in dense, small object scenarios.

Purpose of the Study:

• To develop an improved algorithm, MC-YOLOv5, specifically for accurate multi-class small object detection.
• To enhance the feature extraction capabilities for better small object representation.
• To optimize network architecture for improved detection in dense and small object environments.

Main Methods:

• Introduced an improved Convolutional Block (CB) module for capturing subtle edge information in small objects.
• Developed a Shallow Network Optimization (SNO) strategy to expand receptive fields and reduce missed detections.
• Implemented an anchor frame-based decoupled head for faster training and increased efficiency.
• Evaluated the algorithm on VisDrone2019, Tinyperson, and RSOD datasets.

Main Results:

• MC-YOLOv5 demonstrated superior performance on multi-class small object detection datasets.
• On the VisDrone2019 dataset, MC-YOLOv5 achieved an 8.2% increase in mAP50 and a 5.3% improvement in mAP50-95 compared to YOLOv5L.
• The algorithm showed a 7% increase in F1 score, a 1.8 ms faster inference time, and a 35.3% reduction in computational requirements.
• Similar performance gains were observed across other tested datasets.

Conclusions:

• MC-YOLOv5 is a feasible and effective solution for accurate multi-class small object detection.
• The proposed innovations significantly improve detection precision, reduce missed detections, and enhance overall efficiency.
• MC-YOLOv5 offers a viable alternative to standard algorithms for specialized small object detection tasks.