摘要
随着计算机移动网络技术快速发展,实现屏幕内容图像的分发和处理达到信息共享尤为重要。本文提出了一种基于深度学习卷积神经网络(Convolution neural network,CNN)的屏幕内容图像质量评估方法。人类视觉系统针对文字区域和图像区域具有不同的关注点,所以不能从单一的特征去判断图片的质量,通过设计更深层次的卷积神经网络从而达到多特征提取的目的。本模型使用全卷积网络(Full convolution network,FCN)将屏幕内容图像分割成文字区域和图片区域,利用深层次CNN分别对文本区域和图片区域进行质量评估,通过融合策略将分数整合在一起得到图像的整体质量得分。本模型允许局部质量和局部权重的联合学习,以数据为主导进行驱动,并且不依赖手工特征和图像统计的知识。在两个公开数据集上的实验结果表明,所提出的方法性能相比于目前的方法在主观感知上取得了更高的一致性。
With the rapid development of computer mobile network technology,it is particularly important to realize the distribution and processing of screen content images to achieve information sharing.The method of screen content image quality evaluation based on deep learning Convolution neural network(CNN)is proposed.The human visual system has different concerns for the text area and the image area,so it cannot judge the quality of the image from a single feature.The purpose of multi-feature extraction can be achieved by designing a deeper convolution neural network.The model in this paper first uses the Full convolution network(FCN)to divide the screen content image into text area and image area,then uses the deep level CNN to evaluate the quality of the text area and image area respectively,and finally integrates the scores through the fusion strategy to get the overall quality score of the image.The model in this paper allows the joint learning of local quality and local weight,driven by data,and does not rely on the knowledge of manual features and image statistics.Experiments on two public data sets show that the proposed algorithm achieves higher consistency in subjective perception than the current methods.
作者
杨昭
息佳琦
陈智超
汪国强
YANG Zhao;XI Jiaqi;CHEN Zhichao;WANG Guoqiang(College of Electronic Engineering,Heilongjang University,Harbin 150080,China)
出处
《黑龙江大学自然科学学报》
CAS
2024年第1期99-108,共10页
Journal of Natural Science of Heilongjiang University
基金
国家自然科学基金资助项目(51607059)
黑龙江省自然科学基金资助项目(QC2017059)。
关键词
屏幕内容图像
深度学习
特征提取
神经网络
图像质量评估
screen content image
deep learning
feature extraction
neural network
image quality assessment
