Object Detection of Flexible Objects with Arbitrary Orientation Based on Rotation-Adaptive YOLOv5

Area of Science:

• Computer Vision
• Deep Learning
• Object Detection

Background:

• Accurate detection of flexible objects with arbitrary orientation is crucial for power grid maintenance and inspection.
• Existing methods using horizontal bounding boxes (HBB) suffer from low accuracy due to foreground-background imbalance.
• Multi-oriented detection algorithms using irregular polygons have limitations due to boundary problems during training.

Purpose of the Study:

• To develop a robust object detection method for flexible objects with arbitrary orientation in power grid monitoring.
• To enhance detection accuracy and overcome limitations of existing bounding box strategies.
• To propose a suite of rotation-adaptive YOLOv5 models for diverse practical applications.

Main Methods:

• Introduced a rotation-adaptive YOLOv5 (R_YOLOv5) utilizing rotated bounding boxes (RBB).
• Employed a long-side representation method to add degrees of freedom (DOF) for bounding boxes.
• Addressed boundary problems using classification discretization and symmetric function mapping, optimizing the loss function for training convergence.

Main Results:

• Four R_YOLOv5 models (R_YOLOv5s, R_YOLOv5m, R_YOLOv5l, R_YOLOv5x) were developed.
• Achieved high mean average precision (mAP) on DOTA-v1.5 (up to 0.745) and a self-built FO dataset (up to 0.713).
• R_YOLOv5x outperformed ReDet by 6.84% on DOTA-v1.5 and original YOLOv5 by at least 2% on the FO dataset.

Conclusions:

• The proposed R_YOLOv5 with RBB effectively detects flexible objects with arbitrary orientation, addressing foreground-background imbalance and boundary issues.
• The developed models demonstrate superior recognition accuracy and generalization ability compared to existing methods.
• R_YOLOv5 offers a promising solution for object detection in challenging power grid inspection scenarios.