Vehicle Logo Recognition Based on Enhanced Matching for Small Objects, Constrained Region and SSFPD Network

Area of Science:

• Computer Vision
• Robotic Systems
• Artificial Intelligence

Background:

• Vehicle Logo Recognition (VLR) is crucial for vehicle identification and behavior analysis in robotic systems.
• Current VLR methods struggle with accurate candidate region extraction, especially for small logos and in complex environments.
• Inaccurate logo candidate extraction significantly impacts overall recognition accuracy.

Purpose of the Study:

• To develop an advanced VLR method addressing limitations in existing approaches.
• To enhance the accuracy of small vehicle logo detection and recognition.
• To improve VLR performance in challenging, complex environmental conditions.

Main Methods:

• Proposed a constrained region extraction method using car head and tail segmentation for precise logo candidate localization.
• Introduced an enhanced matching technique involving data augmentation by repeatedly copying and pasting small objects to improve small object detection.
• Developed a Single Shot Feature Pyramid Detector (SSFPD) network, utilizing a reduced ResNeXt model and Feature Pyramid Networks for improved classification and detailed feature retention.

Main Results:

• The proposed VLR method achieved 93.79% accuracy on the Common Vehicle Logos Dataset.
• The method attained 99.52% accuracy on another public vehicle logo dataset.
• Demonstrated superior performance compared to existing VLR methods, particularly for small logos and complex scenarios.

Conclusions:

• The enhanced VLR method effectively overcomes limitations of current techniques.
• The combination of constrained region extraction, enhanced matching, and the SSFPD network significantly improves logo recognition accuracy.
• The proposed approach offers a robust solution for vehicle logo recognition in real-world applications.