SF-YOLOv5: A Lightweight Small Object Detection Algorithm Based on Improved Feature Fusion Mode.

Area of Science:

• Computer Vision
• Object Detection
• Deep Learning

Background:

• Detecting small objects in computer vision remains a significant challenge.
• Existing algorithms often lack specific optimizations for small objects, leading to low accuracy and wasted resources, especially in dense scenes.
• Current methods struggle with the trade-off between detection accuracy and computational efficiency for small object detection.

Purpose of the Study:

• To develop an improved and lightweight object detection algorithm optimized for small objects.
• To enhance the accuracy and real-time capabilities of small object detection systems.
• To reduce the computational cost and resource requirements of object detection models.

Main Methods:

• An improved detection algorithm based on YOLOv5 was proposed, featuring a lightweight design by clipping feature maps.
• A novel feature fusion method, PB-FPN (Pyramid Bi-level Feature Pyramid Network), was developed to improve small object detection capabilities.
• Spatial Pyramid Pooling (SPP) was integrated into the backbone and feature fusion network, enhancing performance and connecting to the prediction head.

Main Results:

• The proposed SF-YOLOv5 algorithm demonstrated significant improvements in detection accuracy and real-time performance.
• Compared to YOLOv5, SF-YOLOv5 achieved a 1.6% increase in mAP@0.5 and a 0.8% increase in mAP@0.5:0.95.
• The new model reduced network parameters by 68.2%, computational resources (FLOPs) by 12.7%, and inference time by 6.9%.

Conclusions:

• The developed SF-YOLOv5 algorithm effectively addresses the challenges of small object detection in computer vision.
• The integration of feature map clipping, PB-FPN, and SPP results in a more accurate, efficient, and lightweight model.
• SF-YOLOv5 offers a promising solution for real-time small object detection with reduced computational overhead.