研究目的
To propose a new approach for lip image segmentation in natural scenes with acceptable computational complexity for mobile devices, employing a complex teacher network and a compact student network with the same structure, enhanced by a remedy loss and an alternative knowledge distillation scheme.
研究成果
The proposed compact student network, trained with an alternative knowledge distillation scheme and remedy loss, achieves superior lip segmentation performance in natural scenes with less computational cost than the teacher network and other state-of-the-art approaches, making it suitable for mobile applications.
研究不足
The approach may be sensitive to the initial segmentation errors of the teacher network and requires careful tuning of the remedy loss parameters.
1：Experimental Design and Method Selection:
The study employs two networks, a teacher and a student network, with the same structure but different complexities. The student network is trained using a novel knowledge distillation scheme that includes a remedy loss to correct segmentation errors from the teacher network.
2：Sample Selection and Data Sources:
A dataset of 49 people captured under natural scenes by various cellphone cameras is used, with images of 40 people for training and the rest for testing.
3：List of Experimental Equipment and Materials:
The study utilizes RGB color images processed by the proposed networks.
4：Experimental Procedures and Operational Workflow:
The student network is trained with the aid of the teacher network using an alternative knowledge distillation scheme, incorporating cross-entropy loss, distillation loss, and remedy loss.
5：Data Analysis Methods:
Segmentation performance is evaluated using pixel accuracy, mean accuracy, mean IU, and frequency weighted IU metrics.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
