基于特征增强知识蒸馏的目标检测压缩算法
Object detection compression algorithm based on feature enhanced knowledge distillation
邓兴隆 1王逸涵 1罗建桥 1熊鹰 1李柏林1
作者信息
- 1. 西南交通大学机械工程学院,四川成都 610031
- 折叠
摘要
本文提出一种基于特征增强知识蒸馏的目标检测压缩算法(ODCA).首先,使用坐标注意力增强教师网络中间层特征对前景目标的表征能力;然后,采用二进制掩码在空间上区分前景和背景,并定义带权重的空间信息损失以平衡前景和背景特征蒸馏强度;最后,使用卷积操作统一教师和学生网络通道数,构造通道信息损失学习特征的通道间分布信息.在Pascal VOC数据集上的实验结果表明,所提算法将学生网络的平均精度均值从68.3%提高至75.7%.
Abstract
An object detection compression algorithm(ODCA)based on feature enhancement knowledge distillation is proposed.Firstly,coordinate attention is used to enhance the representation ability of intermediate layer features of the teacher network for foreground target.Then,binary masks are used to spatially distinguish foreground and background,and a weighted spatial information loss is defined to balance the strength of foreground and background feature distillation.Finally,the convolution operation is used to unify the number of network channels of teachers and students,and the distribution information between channels of the learning features of channel information loss is constructed.Experimental results on the Pascal VOC dataset show that the proposed algorithm improves the mean average precision(mAP)of the student network from 68.3%to 75.7%.
关键词
深度学习/目标检测/模型压缩/知识蒸馏/YOLOv5Key words
deep learning/object detection/model compression/knowledge distillation/YOLOv5引用本文复制引用
基金项目
四川省科技计划(重点研发项目)(2021YFN0020)
出版年
2024
NETL
NSTL
万方数据