Graph-Based Keyword Spotting

iebukes 电子书 73 次浏览 没有评论
WSPC.Graph-Based Keyword Spotting

Graph-Based Keyword Spotting

Handwritten historical documents around the world are endangered by an
increasing state of degradation, and thus, many libraries have started to
preserve their historical treasures by digital means in recent decades. As
a result, large numbers of handwritten historical manuscripts have been
made available digitally for a broader audience. However, we observe a gap
between the availability and accessibility of such documents. That is, automatic full transcriptions of ancient manuscripts are often not feasible due
to wide variations in handwriting and noisy documents. Keyword spotting
(KWS) has been proposed as a flexible and more error-tolerant alternative
to full transcriptions. Basically, KWS allows us to retrieve arbitrary query
words in handwritten historical documents.
In most cases, these methods are based on a statistical representation
of handwriting images. That is, certain characteristics of handwriting images are represented by means of feature vectors. In contrast to this, few
approaches can be found where handwriting is represented by means of
structural representations (i.e. strings, trees, or graphs). This book explores the possibilities and limitations of graph-based representations for
KWS in handwritten historical manuscripts. In particular, it introduces
and thoroughly researches a novel graph-based KWS framework. First,
handwritten historical document images are preprocessed and segmented
into single word images. Based on preprocessed word images, graphs are
extracted by means of different graph representations. The actual keyword
spotting is then based on a pairwise matching of a query graph with all
document graphs. The resulting graph dissimilarities are used to form a
retrieval index that in the best possible case consists of all n instances of
the query word as its top-n results.
Representing handwriting with graphs means that the graphs are affected by subtle variations, and thus, only inexact graph matching can be

下载地址:Graph-Based Keyword Spotting

 
 扫描二维码,关注微信公众号,发送“FWMM”获取下载访问密码,关注我,永不迷路  
                微信公众号二维码