YOLO-SK: A lightweight multiscale object detection algorithm.

Area of Science:

• Computer Vision
• Deep Learning
• Object Detection

Background:

• YOLOv5 struggles with multiscale object detection due to insufficient use of scale information and irrelevant contextual data.
• This limitation particularly impacts performance on resource-constrained devices, leading to prediction errors.

Purpose of the Study:

• To introduce an improved object detection model, YOLO-SK, based on YOLOv5s.
• To enhance the model's ability to handle objects with significant scale variations and improve accuracy on low-performance devices.

Main Methods:

• Developed YOLO-SK by integrating a weighted dense feature fusion network and an SK attention prediction head.
• Implemented dynamic feature fusion across scales with autonomous learning parameters and cross-layer connections.
• Incorporated SIoU loss function and Ghost Convolution for improved accuracy and efficiency.

Main Results:

• YOLO-SK demonstrated significant improvements in prediction accuracy compared to the baseline YOLOv5s.
• Achieved a 2.6% increase in mAP@.5 and a 4.8% increase in mAP@.5:.95 on PASCAL VOC datasets.
• Maintained comparable model complexity while enhancing performance.

Conclusions:

• The proposed YOLO-SK model effectively addresses the limitations of YOLOv5 in multiscale object detection.
• The weighted dense feature fusion and SK attention mechanisms are key to enhancing feature representation and prediction accuracy.
• YOLO-SK offers a promising advancement for precise multiscale object identification on low-performance devices.