文本与模糊形状特征驱动的图像检索与分析

需积分: 10 185 浏览量更新于2024-07-18 收藏 10.94MB PDF 举报

《图像检索与模糊形状特征分析：新兴研究与机会》是一本由P. Sumathy（印度巴腊蒂达桑大学）、P. Shanmugavadivu（印度甘地格拉姆农村研究所）和A. Vadivel（印度特里奇鲁帕利国立技术学院）合著的英文电子书，收录在《多媒体与交互式技术进展》（Advances in Multimedia and Interactive Technologies, AMIT）系列中。该书由美国IGI Global出版社发行，版权归属于2018年的IGI Global，所有内容未经书面许可不得复制、存储或以任何形式（电子或机械，包括影印）分发。本书聚焦于图像检索领域中的一个重要分支，即结合文本和模糊形状特征进行图像内容的查找和分析。作者们探讨了如何利用这些特征来提高图像检索的精度和效率，尤其是在处理大量视觉数据时，如何通过文本描述（如关键词或元数据）辅助识别图像内容，以及如何通过模糊形状匹配技术来适应实际应用场景中可能存在的一致性和不精确性。模糊形状特征分析部分可能涵盖了诸如形状描述符（如SIFT、SURF等）、形状匹配算法（如RANSAC、BFMatcher等）以及形状相似度度量方法的研究。书中可能会介绍如何通过这些技术对图像中的物体进行识别，即使它们的形态有所变化，也能找到相似的匹配。书中还可能涉及深度学习和机器学习在图像检索中的应用，特别是在卷积神经网络（CNN）和深度特征提取方面的最新进展，这些技术能够自动学习并提取图像的高级特征，从而提升检索的准确性。此外，该书可能讨论了在实际应用中的挑战，如多模态数据融合（文本和视觉特征的协同）、实时性需求、大规模数据下的计算效率优化，以及如何通过评估指标（如召回率、精确度、F1分数等）来衡量系统的性能。《图像检索与模糊形状特征分析：新兴研究与机会》提供了一个深入理解图像检索技术前沿的窗口，对于研究人员、工程师和行业从业者来说，它既是理论指导，也是实践参考，有助于推动图像检索领域的技术创新和发展。

Text-Based Image Retrieval

Content-Based Image Retrieval Applications

Various applications such as medicine, fingerprint identification, biodiversity

information systems, digital libraries, crime prevention, historical research

uses image retrieval concept and some of applications are given below for

better understanding.

Medical Applications: The medical industry relies more on diagnostic

images, which are produced by medical instruments. For instance,

a medium-sized hospital usually performs procedures that generate

medical images, which require hundreds or even thousands of gigabytes

within a small space of time. It is understood that processing such huge

amount of data is hard and time-consuming and thus the researchers

have motivated in the field of Content-Based Image Retrieval. In fact,

the medical domain is a well-known area, where Content Based Image

Retrieval finds its application.

Biodiversity Information Systems: Biologists uses various types of

information for biodiversity studies, including spatial data and images of

living beings. The researchers have developed Biodiversity Information

Systems (BIS) to perform query for understanding about species and

their habitats. This is being carried out in the query by combining textual,

content-based and geographical queries. For example, image of a fish

Figure 3. Sample object based retrieval (object dominant) (a) sample retrieval set

(b) extracted object

Text-Based Image Retrieval

can be submitted as query with a semantic “Retrieve all database images

containing fish whose fins are shaped like those of the fish in this image”.

Digital Libraries: In recent times, the image content is effectively used in

several digital libraries for supporting services based on image content.

One example is the digital museum of butterflies, aimed at building a

digital collection of Taiwanese butterflies. It is known that most of the

digital library systems include a module responsible for content-based

image retrieval based on color, texture, and patterns. In addition, content-

based image retrieval digital library can also support geographical image

retrieval. The system processes air photos that are retrieved based on

texture properties. The name of the place associated with retrieved

images can be attached by cross referencing with a Geographical Name

Information System (GNIS) gazetter.

A BRIEF HISTORY OF TBIR

Techniques for Processing HTML Docs

In TBIR, the images are manually annotated with the text descriptions. It

is well known that WWW contains a large number of images embedded in

HTML documents. It is unrealistic to annotate the images manually. Moreover,

designer annotations can be subjective and ambiguous. An alternative way

is to consider the textual data associated with the image such as filename,

caption, surrounding text and HTML TAGs for annotation. Usually, more

descriptive text is available around the images and they are implicitly

provided by the page designer. All these surrounded text include some form

of descriptions generated by the designer about the images and is considered

as a closer semantic representation. By exploiting such rich semantic textual

descriptions, the TBIR has been extensively used in popular image search

engines such as Google and Alta Vista. The textual data present along

with HTML TAGs, such as image filename, ALT TAG can be explored for

capturing the semantics. However, it is found to be difficult to select the exact

feature of a web page to specifically describe the image. Those parts may

not always provide hundred percent relevant text descriptions of the image

that minimizes precision of search engine.

The retrieval of images from the web has received a lot of attention recently.

Most of the early systems have employed text based approach, which exploits

Text-Based Image Retrieval

how images are structured in web documents. Sanderson and Dunlop have

modeled image contents using a combination of texts from associated HTML

pages (Sanderson & Dunlop, 1997). The content is modeled as a bag of

words without any structure and it is considered as one of the inefficiencies

of this approach for indexing. Shen et al 2000 have built chain related terms

and used more information from web documents and proposed a scheme for

unifying keywords with low-level visual features. The assumption made in

this method is that some of the images in the database have been already

annotated in terms of short phrases/keywords. These annotations are assigned

either using surrounding texts of the images in HTML pages or by speech

recognition or manual annotation. During retrieval, user’s feedback is obtained

for semantically grouping keywords with images. Color moments, Color

histograms, Tamura’s features, Co-occurrence matrix features and Wavelet

moments are extracted for representing low-level features. An approach for

unifying keywords and visual features of images (Feng et al 2001) has been

proposed, in which annotations are assigned using surrounding texts of the

images in HTML pages. An attempt has been made by Zhao and Grosky to

improve the semantic correlation between keywords in the document title

in HTML documents and image features for improving retrieval in news

documents (Zhao & Grosky, 2002). Collections of 20 documents from a

news site have been used and 43 keywords along with HSV based color

histogram are constructed. While constructing histogram, saturation and hue

axes are quantized into 10 levels to obtain H×S histogram with 100 bins.

However, it is found that this technique performs well for a small number

of web pages and images only. In general, image search results returned by

search engine contain multiple topics and hence organizing the results into

different semantic clusters would help user. A method has been proposed

for analyzing the retrieval results from web search engine (Feng et al 2004)

which is a bootstrapping approach to automatically annotate a large number

of images from the web. It is demonstrated that the co-training approach

combines the information extracted from image contents and its associated

HTML text. The method VIsion-based Page Segmentation (VIPS) (Deng

et al 2003) has been used to extract semantic structure of a web page based

on visual presentation. The semantic structure is represented as a tree with

nodes, where every node represents the degree of coherence to estimate the

visual perception.

Microsoft Research Asia has developed a system for Web Image Retrieval,

the purpose of which is to cluster the search results of conventional web, so

that the users can find the desired images quickly. Initially, an intelligent

Text-Based Image Retrieval

vision based segmentation algorithm is designed to segment a web page

into blocks and from the block containing image, the information related

to textual and link images are extracted. Later, image graph is constructed

using block-level link analysis techniques. Three types of representations

are obtained for every image such as visual feature based representation,

textual feature based representation and graph based representation. For

each category, several images are selected as non-representative images that

help the users to quickly understand the main topics of the search results.

However, the time for processing tree based representation for indexing and

retrieval is found to be on the higher side. Rough Set Decomposition based

model has been proposed for decomposition in information retrieval (Wu and

Hu, 2010) that consists of three different knowledge granules in incomplete

information system. However, while WWW documents are presented along

with images as input, the semantic of images are exactly captured and thus

retrieval performance is low.

Tollari et al (2005) have proposed a method for mapping textual keywords

with visual features of images. In order to map textual and visual information,

semantic classes containing a few image samples are obtained. Each class is

initialized with exactly one image and then two classes are merged based on

a threshold value on distance between them. The merging process is stopped

when distance between all classes are higher than the second threshold. Finally,

the semantic classes are divided into two partitions as reference and test set.

The reference partition provides example documents of each class, which

are used to estimate the class of any image of test partition using textual,

visual or combination of textual and visual information. The performance of

this method is high with less number of images and corresponding textual

annotations. Han et al have proposed memory learning framework for image

retrieval application (Han et al 2005). This method uses a knowledge memory

model for storing the semantic information by accumulating user-provided data

during query interactions. The semantic relationship is predicted by applying

a learning strategy among images according to the memorized knowledge.

Image queries are finally performed based on seamless combination of low-

level features and learned semantics. Similarly, a unified framework has also

been proposed, which uses textual keywords and visual features for image

retrieval applications (Feng et al 2001 & Hu et al 2000). The framework has

built a set of statistical models using visual features of a small set of manually

labeled images that represents semantic concepts. This semantic concept has

been used for propagating keywords to unlabeled images. These models are

updated regularly when more images implicitly labeled by users become

Text-Based Image Retrieval

available through relevance feedback. This process performs the activity

of accumulation and memorization of knowledge learned from relevance

feedback of users. The efficiency of relevance feedback for keyword query

is improved by an entropy-based active learning strategy. An integrated patch

model has been suggested by Xu and Zang (2007) which is a generative model

for image categorization based on feature selection.

Feature Selection Strategy

Feature selection strategy is categorized into three steps for extracting

representative feature and also to remove the noise feature. Initially, salient

patches present in images are detected, clustered and keyword vocabulary is

constructed. Later, the region of dominance and the salient entropy measure are

calculated for reducing the non-common noises of salient patches. Categories

of the images are described by an integrated patch model with the help of

keywords. The textual keywords appearing in web pages are recently being

used for identifying unsuitable, offensive and pornographic web sites (Hu

et al 2007). In this framework, the web pages are categorized as continuous

text pages, discrete text pages and image pages using decision tree with

respect to their content and these pages are handled by respective classifiers.

Statistical and semantic features are used for identifying pornographic nature

of continuous text web pages. Similarly, Bayesian classifier and object’s

contour of images are being used to identify the pornographic content of

discrete and image web pages respectively. Recently, the keyword is being

extracted from the HTML pages and low-level features are extracted from

the images present in HTML documents for retrieval applications (Vadivel et

al 2009). Various low-level features have been extracted and integrated with

the high-level features for relevant image retrieval. However, the association

and importance of the keywords with the HTML page or with the image is

not measured.

Annotation-Based Information Retrieval (ABIR)

Annotated Based Information Retrieval (ABIR) annotates each image manually

using textual data and standard database management systems are employed

for the retrieval task (Kiling & Alpkocak, 2011). However, annotating large

image database manually is a tedious task and it is difficult to capture the

semantics of the images. The low-level feature of images is replaced by the

剩余192页未读，继续阅读

J-10

粉丝: 18
资源: 483

文本与模糊形状特征驱动的图像检索与分析

imageRetrieval图像检索

ImageRetrieval:使用C ++的图像检索系统

Image Retrieval using CNN and Low-level Feature Fusion

Feature Representation Learning for Unsupervised Cross-domain Image Retrieval

有哪些研究shape analysis的综述论文？

PermissionError: [Errno 13] Permission denied: 'D:/lm/graduation/ImageRetrieval/static/image_database/bird'

推荐20篇关于多特征服装检索的文献

Zero-Shot Object Proposals

最新资源