卷积网络和SIFT特征融合的图像自动标注研究  

Research on Automatic Image Annotation Based on Convolutional Network and SIFT Feature Fusion

作　　者：彭棉珠 PENG Mianzhu(Yangen University,Quanzhou,China,362014)

机构地区：[1]仰恩大学,福建泉州362014

出　　处：《福建电脑》2021年第10期12-16,共5页Journal of Fujian Computer

摘　　要：近年来,图像自动标注成了当下机器学习最热门的研究方向之一。图像自动标注技术能够将互联网上海量的图像信息转换为文本信息,方便进行图像检索、图像分类等应用。现在主流的图像自动标注模型大部分都采用基于编码器—解码器框架的深度学习网络构建而成。本文主要是在编码器的基础上进行研究改进,从而提出了将卷积网络和SIFT特征进行融合的网络模型。该模型结合了卷积网络强大的特征提取能力以及SIFT特征对于图像的旋转、大小比例缩放、明暗度的变化保持不变性的特性,使得模型具有更加强大的图像特征提取能力。最后在公开数据集上和现存有代表性的模型进行了对比实验。实验证明,在Encoder部分加入SIFT特征进行特征融合,可以提高网络模型提取图像特征的能力,使得编码器部分提取到的图像信息更加准确。In recent years,automatic image annotation has become one of the most popular research fields in machine learning.Image automatic labeling technology can convert image information from the Internet into text information,which is convenient for image retrieval,image classification and other applications.At present,most of the mainstream automatic image annotation models are constructed using deep learning networks.These models are based on the encoder-decoder framework.This article mainly study and improve encoder.A network model that integrates convolutional network and SIFT features is proposed,which combines the powerful feature extraction capabilities of convolutional networks and the feature that SIFT features remain invariant to image rotation,size scaling,and changes in brightness.The proposed model have more powerful image feature extracting capabilities.Results of compared experiments with existing representative models on the public data set show that adding SIFT features to the Encoder part for feature fusion can improve the network model’s ability to extract image features and make encoding.The image information extracted by the filter part is more accurate.

关 键 词：图像标注 卷积神经网络 SIFT特征 编解码器 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]