repository
Real-Time Lightweight Human Parsing Based on Class Relationship Knowledge Distillation
- Author(s)
- LANG YUQI
- Alternative Author(s)
- LANG YUQI
- Advisor
- 황원준
- Department
- 일반대학원 인공지능학과
- Publisher
- The Graduate School, Ajou University
- Publication Year
- 2023-08
- Language
- eng
- Alternative Abstract
- In the field of computer vision, understanding human objectives is a crucial and chal- <br>lenging task, as it requires recognizing and comprehending human presence and behavior in <br> <br>images or videos. Within this domain, human parsing is an extremely challenging task, as <br>it necessitates accurately locating the human region and dividing it into multiple semantic <br>areas. This is a dense prediction task that demands powerful computational capabilities <br>and high-precision models. Recently, with the continuous development of computer vision <br> <br>technologies, human parsing has been widely applied to other tasks related to human ob- <br>jectives, such as pose estimation, and human image generation. These applications are <br> <br>expected to play an increasingly important role in future artificial intelligence research. <br> <br>To achieve real-time human parsing tasks on devices with limited computational re- <br>sources, we have designed and introduced a lightweight human parsing model. We chose <br> <br>Resnet18 as the core network structure and simplified the traditional pyramid module used <br> <br>to obtain high-definition contextual information, thus significantly reducing the complex- <br>ity of the model. Additionally, to enhance the parsing accuracy of the model, we integrated <br> <br>a spatial attention fusion strategy. Our lightweight model exhibits efficient performance <br>and achieves high segmentation accuracy on the commonly used dataset for human parsing <br>tasks, Look into Person (LIP). Although traditional models perform excellently in terms of <br>segmentation accuracy, their high complexity and abundance of parameters restrict their <br>use on devices with limited computational resources. To further improve the accuracy of <br> <br>our lightweight network, we also implemented knowledge distillation techniques. The tra- <br>ditional knowledge distillation method uses the Kullback-Leibler (KL) divergence to match <br> <br>the prediction probability scores of teacher-student models. However, this approach may <br>be ineffective at learning useful knowledge when there is a significant difference between <br>the teacher and student networks. Therefore, we adopted a new distillation standard, <br>based on inter-class and intra-class relationships in prediction results, which significantly <br>improves parsing accuracy. Empirical evidence has shown that, while maintaining high <br>segmentation accuracy, our lightweight model has substantially reduced the number of <br>parameters, thereby achieving our expected goals.
- Fulltext
- Appears in Collections:
- Graduate School of Ajou University > Department of Artificial Intelligence > 3. Theses(Master)
- Files in This Item:
- There are no files associated with this item.
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.
-
- License
- STATISTICS
- Total Visit :5,015,702
- Total Download :2,095
- Today View :5,886
