期刊论文

中文

华中科技大学学报(自然科学版)

1671-4512

2017

45

11

22-28,67

10.13245/j.hust.171105

61702208:国家自然科学基金 41671377:国家自然科学基金 61502195:国家自然科学基金 CCNU17QN0003:中央高校基本科研业务费专项 CCNU2016A02020:中央高校基本科研业务费专项

本校为第一机构

针对现有图书页面检索方法检索精度低的问题,利用任务无关数据集训练卷积神经网络,提出了一种基于卷积神经网络的图书页面检索方法.首先将待检图书页面图像进行图像分割和畸变校正,降低背景的干扰和几何畸变的影响;然后将校正后的图像输入卷积神经网络提取图像特征;最后使用夹角余弦距离来度量待检图像和候选图像的相似度.实验结果表明:本方法在测试数据集上的Top-5命中率为97.31%,而直接使用任务无关数据集训练的卷积神经网络的Top-5命中率仅为58.47%.本方法避免了耗费大量的时间和精力去收集大规模图书页面图像数据库,而且利用卷积神经网络强大的图像特征描述能力,取得了优异的图书页面检索精...

As existing book page retrieval approaches based on handcraft features usually can not achieve the high precision required by many applications, a convolutional neural network based book page retrieval method was presented. Firstly, to avoid the interference of the background and the geometry distortion, the query book page image was separated from background and corrected from geometry distortion. Then, the corrected image was passed into a convolution neural network to extract the image features. Finally, the cosine distance was used to measure the similarity of the query image and the candi...