Efficient adaptive rotated object detection for 1D and QR barcodes.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Object detection, particularly for rotated objects like barcodes, presents significant challenges.
• Existing methods often struggle with computational efficiency and accuracy in complex scenarios.

Purpose of the Study:

• To introduce EA-OBB, a novel lightweight framework for rotated object detection, specifically targeting 1D and Quick Response (QR) barcodes.
• To enhance both the accuracy and computational efficiency of rotated object detection.

Main Methods:

• The EA-OBB framework is built upon the YOLO11 architecture, incorporating innovative modules: KWConv for dynamic feature extraction, ORPNCSPELAN for efficient multi-path aggregation, and LADH-OBB for precise rotation angle regression.
• The Taylor Pruning algorithm is integrated to reduce model parameters and computational costs, optimizing for resource-constrained environments.

Main Results:

• Experiments on the RotBar dataset show EA-OBB outperforms existing methods in rotated object detection.
• EA-OBB achieves a superior balance between precision, recall, and computational complexity.

Conclusions:

• EA-OBB offers an effective and efficient solution for detecting rotated 1D and QR barcodes.
• The framework demonstrates significant improvements in accuracy and computational performance, making it suitable for real-world applications.