VPDet: Refined region CNN for hair follicle recognition in arbitrary angle

• Published in: IET image processing Vol. 18; no. 12; pp. 3156 - 3167
• Main Authors: Gu, Xinyu, Zhang, Xiaoxu
• Format: Journal Article
• Language: English
• Published: Wiley 01.10.2024
• Subjects: image recognition; object detection
• ISSN: 1751-9659; 1751-9667
• DOI: 10.1049/ipr2.13162

In response to the growing demand for hair‐loss treatments, this study introduces the vector proposal detector (VPDet), a groundbreaking solution in hair transplant robotics. VPDet, distinct from traditional approaches, addresses the complex challenges of hair follicle detection, notably the variability in hair growth orientations and the intricacies of hair clustering. The method innovatively leverages the linear nature of hair, spanning a full 360‐degree orientation spectrum. The VPDet framework, a novel two‐stage object detection system, incorporates the vector proposal network and vector align blocks. These elements are crucial in transforming conventional anchor boxes into anchor vectors, thereby generating reference vectors across various scales and angles. The vector align block, a key innovation, uniquely combines vector and adjacent feature data to align features through shared maps. The extensive experiments, conducted on the FDU_HairFollicleDataset and an extended dataset, exhibit a remarkable enhancement in model performance, with a 51.3% increase in precision and a 20.8% boost in F1 score. The results not only demonstrate VPDet's superior capability in hair follicle recognition but also its potential in posture recognition for vector‐characteristic objects. This approach represents a significant advancement in both the field of hair transplant robotics and vector‐based object detection. A novel end‐to‐end target detection framework for hair follicle recognition in hair transplantation scenarios is proposed. Vector proposal network and vector align modules are designed for the precise generation of region proposals, thereby enhancing hair follicle detection at arbitrary angles. The method framework is applied to the FDU‐HairFollicleDataset and demonstrates the effectiveness and accuracy of the approach through experimentation.