基于距离损失函数的特征融合模型
Feature Fusion Model Based on Distance Loss Function
林益文 1杨啸 1陈青 1邱新媛 1任维泽1
作者信息
- 1. 中国核动力研究设计院,四川 成都 610213
- 折叠
摘要
卷积神经网络在图像识别任务中表现出出色的学习语义特征的能力,实现了相关目标对象的高精度检测,然而其通常只关注图像最具识别能力的特定区域,忽视了部分有价值的语义特征.为提高卷积神经网络的识别能力,提出一种基于距离损失函数的特征融合模型.该方法利用欧氏余弦复合距离损失函数迫使基础模型学习具有差异的特征概念,通过并置特征融合法整合差异化的特征概念进行目标识别.实验使用了多种基准卷积神经网络骨架、数个流行数据集和不同样本量进行多因素交叉分析,从准确率数据和类激活图两个方面证实了该方法能够丰富基础模型语义特征的多样性,提升融合模型的识别性能,并且具有有效性和普遍性,同时利用数学统计方法也揭示了该方法的应用特征与优势.
Abstract
Convolutional neural networks have shown excellent ability to learn semantic features in image recognition tasks and realized highly accurate detection of relevant target object.However,it usually only focuses on specific regions of the image with the most recognition power,ignoring some valuable semantic features.To improve the recognition capability of convolutional neural networks,we propose a feature fusion model based on distance loss function.The Euclidean-Cosine distance loss function is used to force the base models to learn differentiated feature concepts,and the concatenation feature fusion method is implemented to integrate differentiated feature concepts for the image recognition.Experiments are conducted using different benchmark convolutional neural networks,several popular datasets and different sample sizes for multi-factor cross-tabulation analysis.It's confirmed that the proposed method can enrich the diversity of semantic features of the base models,and improve the recognition performance of the fusion model in terms of accuracy values and class activation maps.The validity and generality of the proposed method can also be guaranteed in this way.Meanwhile,the application characteristics and advantages of the method are also revealed using mathematical statistical methods.
关键词
深度学习/卷积神经网络/语义特征/距离函数/模型融合Key words
deep learning/convolutional neural network/semantic feature/distance function/model fusion引用本文复制引用
出版年
2023
国家科技期刊平台
NSTL
维普
万方数据