Enhanced Express Package Trademark Recognition via a Novel PTD-YOLO Algorithm

• Published in: Traitement du signal Vol. 40; no. 3; pp. 1083 - 1091
• Main Authors: Gao, Yanghua, Lou, Weidong, Lu, Hailiang, Zhang, Li, Wang, Yaliang
• Format: Journal Article
• Language: English
• Published: Edmonton International Information and Engineering Technology Association (IIETA) 01.06.2023
• Subjects: Ablation; Accuracy; Algorithms; Methods; Modules; Neural networks; Object recognition; Packaging; Parameterization; Target detection; Trademarks; Vision systems
• ISSN: 0765-0019; 1958-5608
• DOI: 10.18280/ts.400324

The rise in prominence of the logistics industry necessitates a boost in its efficiency. A notable hurdle to this lies in the classification of goods, based on the unique trademark of express packages, a problem with a direct bearing on delivery efficiency. Traditional methodologies for inspecting packaging appearances struggle with accuracy in recognizing a variety of scales, necessitating the use of multiple detection systems. Additionally, they fail in accurately ascertaining the precise location and size of the express packaging trademark. To rectify this, the study presents the development and application of a detection technique christened PTD-YOLO (Packing Trademark Detection algorithm based on YOLO, PTD-YOLO). This technique bolsters the YOLO v5 algorithm through improvements in three key areas. The first is the restructuring of the FSRP (Focus module with Structural Re-Parameterization) module, aimed at enhancing pre-backbone features. The second involves the integration of a novel prediction head, designed to bolster the ability of PTD-YOLO in detecting smaller-scale targets. Lastly, an attention mechanism has been incorporated within the head part, to better distinguish relevant features of detected objects. The performance of the PTD-YOLO has been validated via rigorous ablation and comparative experiments, proving its effectiveness and reliability.