基于多结构教师蒸馏的服装图像分类方法
CLOTHING IMAGE CLASSIFICATION BASED ON KNOWLEDGE DISTILLATION OF MULTI-STRUCTURE TEACHERS
张晓滨 1刘昊1
作者信息
- 1. 西安工程大学计算机学院 陕西 西安 710048
- 折叠
摘要
针对当前大多数服装图像分类模型结构复杂、参数量大的问题,结合多教师知识蒸馏方法,提出一种基于多结构教师蒸馏的服装图像分类方法.该方法要点在于,选取学习了不同类型知识的多个蒸馏模型作为多教师网络,根据每个教师的模型表现自适应分配权重,协同作为监督来指导目标模型的学习,实现服装分类模型的轻量化改进.在服装数据集DeepFashion上的实验表明,该方法相对于同结构的服装分类模型约有 1.14 百分点的准确率提升,且模型本身仅有0.27M的参数量.
Abstract
In order to solve the problems of complex structure and large number of parameters in current garment image classification models,this paper proposes a garment image classification method based on multi-teacher knowledge distillation.The key point of this method is to select multiple distillation models with different types of knowledge as a multi-teacher network,assign weight adaptively according to the model performance of each teacher,and cooperate as a supervisor to guide the target network,so as to realize the lightweight improvement of the clothing classification model.Experiments on DeepFashion show that the proposed method has an accuracy improvement of about 1.14 percentage points compared with the clothing classification model with the same network structure,and the model itself has only 0.27 M parameters.
关键词
模型压缩/知识蒸馏/多教师知识蒸馏/服装图像分类Key words
Model compression/Knowledge distillation/Multi-teacher knowledge distillation/Clothing image classification引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据