| 面向压缩图像复原的网络增强训练方法 |
| |
| 引用本文： | 廖理心,赵耀,韦世奎.面向压缩图像复原的网络增强训练方法[J].信号处理,2022,38(6):1192-1201. |
| |
| 作者姓名： | 廖理心 赵耀 韦世奎 |
| |
| 作者单位： | 1.北京交通大学信息科学研究所，北京 100044 |
| |
| 基金项目： | 国家重点研发计划2018AAA0102100国家自然科学基金62120106009中央高校基本科研业务费专项资金资助2018JBZ001 |
| |
| 摘 要： | 高质量的数据是深度卷积神经网络成功的关键因素之一。在计算机视觉领域，常用图像数据集通常以JPEG格式存储。这种有损压缩技术不可避免地会导致原始数据信息的丢失，进而造成利用压缩数据训练的卷积神经网络的性能降低。因此，为了增强卷积神经网络的性能，本文提出了一种面向压缩图像复原的增强训练方法，通过复原压缩图像实现卷积神经网络的性能增强。该方法具体为一个包含复原模块和任务模块的联合增强框架。复原模块致力于恢复有损压缩技术造成的信息丢失；任务模块专注于基于任务需求增强压缩图像。两个模块联合训练，使得压缩图像的复原增强更具有目的性。本文通过图像分类任务的实验表明，与压缩图像相比，该方法能有效地复原压缩图像，增强卷积神经网络的性能。此外，该方法中两个模块间的低耦合性和可替代性保证了该方法的适用性。
|
| 关 键 词： | 压缩图像 复原 联合增强 |
| 收稿时间： | 2021-10-08 |
Enhancing Compressed Images for Training Better CNNs |
| |
| Affiliation: | 1.Institute of Information Science, Beijing Jiaotong University, Beijing 100044, China2.Beijing Key Laboratory of Advanced Information Science and Network Technology, Beijing 100044, China |
| |
| Abstract: | ? ?High-quality data is one of the key factors for the success of deep convolution neural network. In the field of computer vision， commonly used image data sets are usually stored in JPEG format. Because of the lossy compression technology， part of the original data information is inevitably lost. It further results in the performance degradation of convolutional neural network trained with compressed data. Therefore， in order to enhance the convolutional neural network， this paper proposes a joint enhancement framework， including the restore module and the task module. The restore module aims to recover the information loss caused by the lossy compression technology. According to the task requirements， the task module focuses on enhancing the compressed image. The joint training of these two modules makes the restoration and enhancement of compressed images more purposeful. Experiments on the image classification task show that our method can effectively restore the compressed image and enhance the performance of convolutional neural network. In addition， the low coupling and substitutability of the two modules in the method make the method applicable. |
| |
| 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 |
