基于语义感知的图像压缩算法研究  

Research on Semantic-Aware Image Compression Algorithms

作　　者：宋媛萌 贾正正 贾召弟 王宇辰 杨少华 韩卓航 

机构地区：[1]北华航天工业学院计算机学院,河北 廊坊 [2]北华航天工业学院航空宇航学院,河北 廊坊

出　　处：《计算机科学与应用》2024年第12期67-75,共9页Computer Science and Application

基　　金：北华航天工业学院2023年硕士研究生科研创新项目(项目号:YKY-2023-40);北华航天工业学院2023年硕士研究生科研创新项目(项目号:YKY-2023-39);北华航天工业学院2024年硕士研究生科研创新项目(项目号:YKY-2024-70);北华航天工业学院2024年硕士研究生科研创新项目(项目号:YKY-2024-40)。

摘　　要：图像压缩的目的是尽量保持图像质量的前提下减少图像数据的存储空间。传统的图像压缩方法主要依赖于对图像像素进行编码和量化,无法利用图像中高级语义信息。本文提出了一种基于语义感知的图像压缩算法,具体步骤为:首先,通过卷积神经网络对图像进行语义分析,然后,通过语义感知模块提取出图像的语义级别,即主要语义区域和次要语义区域。最后,将语义级别带入VAE图像压缩网络模型中,根据语义区域信息,对图像中的主要语义区域进行轻度压缩,而对次要区域进行更大幅度的压缩,以确保在减少文件体积的同时,最大程度地保留图像中的关键信息和视觉质量。在Kodak等公开数据集上进行实验,实验表明基于语义感知的图像压缩算法在提供更好的图像质量方面具有显著优势。The goal of image compression is to reduce the storage space of image data while preserving image quality as much as possible. Traditional image compression methods primarily rely on encoding and quantizing image pixels, and are unable to leverage advanced semantic information within the image. This paper proposes a semantic-aware image compression algorithm, with the following steps: First, a convolutional neural network is used to perform semantic analysis on the image. Then, a semantic-aware module extracts the semantic levels of the image, identifying primary and secondary semantic regions. Finally, these semantic levels are incorporated into a VAE image compression network model, where the primary semantic regions of the image are lightly compressed, and the secondary regions are more heavily compressed, ensuring that while reducing file size, the key information and visual quality within the image are retained to the greatest extent possible. Experiments conducted on public datasets such as Kodak demonstrate that the semantic-aware image compression algorithm has a significant advantage in providing superior image quality.

关 键 词：语义感知网络 VAE 图像压缩 深度学习 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]