| 深度学习和迭代量化在图像检索中的应用研究 |
| |
| 引用本文： | 甄俊杰,应自炉,赵毅鸿,黄尚安.深度学习和迭代量化在图像检索中的应用研究[J].信号处理,2019,35(5):919-925. |
| |
| 作者姓名： | 甄俊杰 应自炉 赵毅鸿 黄尚安 |
| |
| 作者单位： | 五邑大学智能制造学部 |
| |
| 基金项目： | 国家自然科学基金(61771347)；广东省特色创新类项目(2017KTSCX181)；广东省青年创新人才类项目(2017KQNCX206)；江门市科技计划项目（江科〔2017〕268号）；五邑大学青年基金(2015zk11)
|
| |
| 摘 要： | 基于内容的图像检索的关键在于对图像进行特征提取和对特征进行多比特量化编码 。近年来，基于内容的图像检索使用低级可视化特征对图像进行描述，存在“语义鸿沟”问题；其次，传统量化编码使用随机生成的投影矩阵，该矩阵与特征数据无关，因此不能保证量化的精确度。针对目前存在的这些问题，本文结合深度学习思想与迭代量化思想，提出基于卷积神经网络VGG16和迭代量化(Iterative Quantization, ITQ)的图像检索方法。使用在公开数据集上预训练VGG16网络模型，提取基于深度学习的图像特征；使用ITQ方法对哈希哈函数进行训练，不断逼近特征与设定比特数的哈希码之间的量化误差最小值，实现量化误差的最小化；最后使用获得的哈希码进行图像检索。本文使用查全率、查准率和平均精度均值作为检索效果的评价指标，在Caltech256图像库上进行测试。实验结果表明，本文提出的算法在检索优于其他主流图像检索算法。
|
| 关 键 词： | 图像检索 深度学习 VGG16 迭代量化 哈希码 |
| 收稿时间： | 2019-01-02 |
Application research of deep learning and iterative quantization in image retrieval |
| |
| Affiliation: | School of Intelligent Manufacturing, Wuyi University |
| |
| Abstract: | The key of content-based image retrieval lies in feature extraction and multi-bit quantization coding. In recent years, content-based image retrieval uses low-level visual features to describe images, and there is a "semantic gap" problem. Besides, the traditional quantization coding uses the randomly generated projection matrix, which is independent of the feature data, so it can not guarantee the accuracy of quantization. Aiming at these problems, this paper proposes an image retrieval method based on convolution neural network VGG16 and iterative quantization (ITQ). We use the Network Model VGG16 pre-trained on an open dataset. The method of ITQ is used to train the Hashing function, the minimum quantization error between the feature and the hash code with the bits of set number continuously approximated, so the quantization error is minimized. Finally, the obtained hash code is used for image retrieval. In this paper, recall rate, precision rate and mean average precision are used as the index to evaluate the retrieval results, and the results are obtained based on the Caltech256 image database. Experimental results show that the proposed algorithm is better than other mainstream image retrieval algorithms. |
| |
| Keywords: | |
|
| 点击此处可从《信号处理》浏览原始摘要信息 |
|
点击此处可从《信号处理》下载全文 |
|
