首页|图像信息量度量

图像信息量度量

扫码查看
图像不仅承载着丰富的视觉内容,同时还蕴含潜在的高级语义,是重要的信息传递媒介。度量图像信息量则是对图像信息进行定性描述和定量计算的抽象过程,属于对图像处理领域中的科学问题进行理论构建的关键环节。然而,在现有的研究工作中,虽然对"图像信息量"的度量有所提及,但实际上关注的往往是在图像质量或图像特征的层级,而忽略了图像所具有的高级语义以及人对图像的认知。因此,为了衡量图像的信息提供价值,本文基于"图像信息量因人、因任务而异"这一常识,通过建模"先验知识"相关的图像记忆并提取"给定任务"相关的图像特征,尝试对图像信息量进行度量。首先,提出图像信息量度量的理论框架,具体表现为:在"能够获得全世界所有的图像"和"能找到一种能够准确表达图像语义的特征"两个假设条件下,以信息论为基础,通过构建超球来描述样本点的邻域稠密度,由此建模语义信息的概率,进而得到理论假设条件下的图像信息量。其次,探讨了理论假设无法满足的实际情况,将"全世界所有的图像"这一假设条件退化为"给定具体的数据集",并将"一种能够准确表达图像语义的特征"这一假设退化为"给定任务相关的特征",进一步通过实验展示了在实际情况下如何有效获得图像信息量的数值结果。最后,指出了本文度量方法的限定对象和度量结果的边界,并展望了未来对相关体系进行丰富和完善的可行方向。
Measuring the information of images
Images possess not only abundant visual content but also inherent high-level semantics,rendering them a pivotal medium for information transmission.Measuring the information of images is an abstract process of qualitative description and quantitative calculation of image information,constituting a fundamental aspect of scientific theories in the field of image processing.While the measurement of"image information"has been mentioned in previous research,it primarily centers on the quality or features of images,neglecting the exploration of the factors linked to high-level semantics and human cognition.Hence,to gauge the informational value that images offer to individuals,this paper seeks to measure image information by modeling the iconic memory associated with individuals'"prior knowledge"and extracting image features pertinent to"the given tasks".This approach is grounded in the widely accepted belief that"the information of images differs among individuals and across various tasks".Firstly,a theoretical framework for measuring the information of images is proposed.It is demonstrated as follows:under the assumptions of"having access to all images in the world"and"possessing a feature that can fully express the semantic content of images",a hypersphere is constructed to describe the density of the neighborhoods of the sample point.By modeling the probability of semantic information within this framework,the information of images is obtained based on information theory under the theoretical assumptions.Secondly,the practical scenarios when the assumptions specified in the theoretical framework are not feasible are discussed.It is illustrated by reducing the assumption of"having access to all images in the world"to"the given specific dataset"and reducing the assumption of"possessing a feature that can fully express the semantic content of images"to"the given task-related features".Then the experiments were conducted to show how to effectively obtain numerical results of image information in practical scenarios using the proposed method.Lastly,the limitations of the measurement method and the boundaries of the measurement results are pointed out,and feasible directions for enriching and improving the relevant systems in the future are contemplated.

image informationinformation measurementinformation capacityinformation theoryimage processing

李学龙、何如玢

展开 >

西北工业大学光电与智能研究院,西安 710072

智能交互与应用工业和信息化部重点实验室(西北工业大学),西安 710072

图像信息量 信息量度量 信容 信息论 图像处理

国家重点研发计划国家自然科学基金

2022YFC280800061871470

2024

中国科学F辑
中国科学院,国家自然科学基金委员会

中国科学F辑

CSTPCD北大核心
影响因子:1.438
ISSN:1674-5973
年,卷(期):2024.54(6)