 |
Description  |
|
|
FIELD OF THE INVENTION
The present invention relates generally to image processing and
specifically to archiving and retrieving of images, such as digitized
photographs, works of art, or graphic illustrations, by use of a natural
language such as English.
DESCRIPTION OF RELATED ART
Numerous schemes have been used in the past for archiving images and
selecting images for retrieval from such archives. Before computers became
widely available, simple index cards were often used to keep track of
stock photographs, and personnel within photo agencies often relied on
their own experience to retrieve photographs that corresponded to a
potential customer's request.
Such methods of archiving and retrieving stock photographs provided
imperfect results, and were difficult, time-consuming and expensive to
implement. As image libraries grew, the shortcomings of conventional
archiving and retrieval techniques became yet more pronounced.
The advent of photo Compact Disc ("CD") technology allowed certain advances
to be made in this field. With CDs, a customer may purchase rights to use
a large number of pictures that may be stored on a single disc and
selectively browsed using a CD-ROM drive. However, the number of images
available on a CD is still somewhat limited, and most CD-based photo
portfolios require a relatively large up-front payment for all of the
images on the CD, regardless of how many the user may be interested in.
Finally, image quality on CD-based photo portfolios is not always
production quality.
Some on-line systems have recently become available that include photo CD
technology, such as the KODAK PICTURE EXCHANGE ("KPX") and the COMSTOCK
BULLETIN BOARD SERVICE. Such services typically include relatively large
libraries of images, and permit conventional keyword search techniques.
However, none of the known systems provide an easy to use, natural
language search capability, nor do they allow for automating the process
of pricing, ordering, and delivering selected images.
It would be desirable to allow users to select images from a library based
on conceptual characteristics of such images, to obtain immediate pricing
information regarding selected images, and to order and obtain
production-quality versions of such images directly.
DISCLOSURE OF INVENTION
In accordance with the present invention, images are archived and retrieved
by associating metadata with an image, the metadata including
bibliographic data, a caption, and a set of suggestions evoked by the
image, removing ambiguities from the metadata, storing the image and
metadata in a database with other images and metadata, and selecting
certain images from the database that have metadata corresponding to a
user's search request.
In one aspect of the invention, a natural language processing technique is
used in connection with the selection of images based on the user's search
request.
In another aspect of the invention, an image is watermarked so as to allow
the image to be viewed for selection on a computer monitor, but not to be
printed in a usable format or downloaded for digital publishing.
In yet another aspect of the invention, the user may order and obtain
delivery of selected images directly over a computer connection.
In still another aspect of the invention, ambiguities in the metadata are
removed by highlighting portions of the metadata that are recognized as
having multiple senses, providing a list of possible senses for those
portions, and allowing the user to select the appropriate sense.
In accordance with the present invention, apparatus (100) for image archive
and retrieval includes an ingestion center (110), an image center (120),
and user workstations (130-132).
BRIEF DESCRIPTION OF THE DRAWINGS
FIG. 1 is a physical block diagram of apparatus for pattern recognition in
accordance with the present invention.
FIG. 2 is a functional block diagram of the ingestion center portion of the
apparatus of FIG. 1, in accordance with the present invention.
FIG. 3 is a functional block diagram of the image center portion of the
apparatus of FIG. 1, in accordance with the present invention.
FIG. 4 is a flow diagram illustrating disambiguation processing, in
accordance with the present invention.
FIG. 5 is a flow diagram illustrating watermarking, in accordance with the
present invention.
FIG. 6 illustrates an index card screen used in conjunction with the
ingestion center portion of the apparatus of FIG. 1, in accordance with
the present invention.
FIG. 7 illustrates an interpreter screen used in conjunction with the
ingestion center portion of the apparatus of FIG. 1, in accordance with
the present invention.
FIG. 8 is a flow diagram of upload processing, in accordance with the
present invention.
FIG. 9 is a flow diagram of index server upload processing, in accordance
with the present invention.
FIG. 10 is a flow diagram of search engine processing in accordance with
the present invention.
FIG. 11 illustrates communications layers of the apparatus of FIG. 1, in
accordance with the present invention.
FIG. 12 illustrates a match list screen used in conjunction with the image
center portion of the apparatus of FIG. 1, in accordance with the present
invention.
FIG. 13 illustrates an image browser screen used in conjunction with the
image center portion of the apparatus of FIG. 1, in accordance with the
present invention.
FIG. 14 illustrates an information screen used in conjunction with the
image center portion of the apparatus of FIG. 1, in accordance with the
present invention.
FIG. 15 illustrates a lightbox screen used in conjunction with the image
center portion of the apparatus of FIG. 1, in accordance with the present
invention.
FIG. 16 illustrates a pricing screen used in conjunction with the image
center portion of the apparatus of FIG. 1, in accordance with the present
invention.
BEST MODE FOR CARRYING OUT THE INVENTION
Referring now to FIG. 1, there is shown a system 100 for archiving and
retrieving images in accordance with the present invention. Briefly, the
system 100 is comprised of three major functional blocks. Ingestion center
110 is used to enter images and associated characterizing data (described
below as "metadata") into system 100. Image center 120 is used to store
the image data and metadata, and to process queries for images based on
the metadata. User workstations 130-132 are used to allow remote users to
enter queries for images, to view the images sent by image center 120, to
obtain pricing information on such images, to order such images, and to
obtain delivery of such images.
More specifically, ingestion center 110 includes a data entry processor
112, disambiguation processor 114, and natural language processing ("NLP")
database 116. Image and basic bibliographic information provided by stock
photography agents are converted into digital format if not provided in
that form by the agents, and conventionally input to ingestion center 110
using data entry processor 112. Typically, the basic bibliographic
information provided by the agents includes the artist's name, source,
copyright holder, location, artist's project name or series, dates,
release information, and any notes relating to the photograph supplied by
the artist. The data entry processor 112 permits input of the image data
and this textual data to 110, and also allows an operator known as a
"captioner" to verify the quality of both the image data and the
bibliographic data, to write a short caption, or description, of the
salient features of the image, and to select certain attributes of the
image. The caption may be a set of regular English language sentences, as
opposed to merely a listing of unconnected keywords. The attributes may
include, for example, the type of image (photograph, computer-generated
graphic, video clip or other multimedia object, background pattern,
portrait, abstract, aerial, or special effect), predominant hue, and image
orientation (landscape or portrait). The captioner also provides as part
of the metadata a "suggests" text field describing the emotional
suggestions evoked by the image. If not already provided by the photo
agency with the bibliographic data, the captioner may obtain and add to
the bibliographic data information concerning the prices and other terms
under which such image may be licensed. Collectively, the bibliographic
data, the caption, the attributes and the suggests field are known as the
"metadata" associated with the image.
A disambiguation processor 114 takes as input the metadata of the image and
identifies for the captioner any portions of the metadata that are capable
of multiple interpretations, based on information previously stored in NLP
database 116. The captioner may then select which interpretation is
desired, or may enter a new interpretation. The disambiguation processor
114 also serves to standardize the form of the caption, so that all
captions use conjunction and disjunction in a standard way, and so that
all captions are written in the same anaphoric reference style.
Furthermore, if the style of captions is standardized, the location of
information within a caption may even provide useful information. If, for
example, the most important descriptive information is consistently placed
in the first sentence of a caption, that information can be weighted more
heavily in making decisions about the relative "closeness" of a query to
the caption.
In a preferred embodiment, data entry processor 112, disambiguation
processor 114, and NLP database 116 are implemented using conventional
client/server computer systems, with client workstations being personal
computers such as the APPLE MACINTOSH or IBM-compatible personal computers
and servers being conventional computers such as the SUN SPARCSTATION.
Memory map B-trees are used to implement NLP database 116, as described in
R. Sedgewick, ALGORITHMS IN C++, Reading, Mass.: Addison-Wesley (1992),
the teachings of which are incorporated herein by reference. NLP database
116 contains information about words, their senses, and how those senses
are linked together. Word senses are represented as unique numbers. An
"expansions" portion of NLP database 116 represents each link as a
database record incorporating (i) the word sense, represented by a unique
number, (ii) a word sense to which the sense in (i) is linked, represented
by another unique number, and (iii) the type of link between the senses in
(i) and (ii), e.g., "synonym", "antonym", "a kind of", "a part of."
The primary components of image center 120 include upload processor 122,
database processor 124, browse processor 126, and order processor 128.
Upload processor 122 is used to take the image and metadata information
from ingestion center 110 and store it in a form usable by database
processor 124. Database processor 124 stores images and metadata for use
by browse processor 126 and order processor 128, and also archives image
data for long-term storage. In a preferred embodiment, database processor
stores images in three separate resolutions for "thumbnail",
quarter-sized, and full-sized display as desired, and uses VHS-type
videotape as the medium for archiving full-resolution images. Browse
processor 126 permits a user to see both image data and metadata for
selected images, and order processor 128 permits the user to learn pricing
information for such images, to order such images, and to obtain delivery
of such images. In a preferred embodiment, processors 122-128 are
implemented using conventional client/server architecture as described
above in connection with the components of ingestion center 110.
User workstations 130-132 permit users of system 100 to access the image
center 120 for the purpose of posting image queries, for reviewing the
results of such queries, for selecting images to order, for completing
ordering transactions, and for receiving print-quality images. For
purposes of illustration, three user workstations 130-132 are shown but it
should be recognized that any number of such workstations may be used in
accordance with the invention. In a preferred embodiment, workstations
130-132 are implemented using conventional personal computers such as
described above in connection with the client computers of ingestion
center 110, programmed to provide the functionality described herein.
The components of system 100 illustrated in FIG. 1 are further described
below in connection with the other figures.
Referring now also to FIG. 2, there is shown a functional block diagram of
ingestion center 110. In operation, a digitized picture 250 and
bibliographic data 252 (shown in simplified form in FIG. 2) are applied to
an image/bibliographic input and quality control service 202. Service 202,
implemented primarily by data entry processor 112, permits input of the
image and agency-supplied bibliographic data into system 100. Service 202
also displays the image and data so that a captioner may perform quality
control to ensure that the image is right-side up and in focus, that the
colors are correct, and that the bibliographic data 252 was scanned or
otherwise input correctly and matches the image 250.
After processing by service 202 is complete, the image and bibliographic
data are applied to a caption/suggests field entry service 204. This
service 204 permits a captioner to enter the caption and suggests field
information as described in connection with FIG. 1. In a preferred
embodiment, service 204 is implemented using data entry processor
disambiguation processor 114, but other processors, e.g., data entry
processor 112, could also provide this functionality.
The data are next applied to a disambiguation of caption and suggests field
tool 206. This tool 206 provides processing, described in greater detail
in connection with FIG. 4, that checks the spelling of words in the
bibliographic data, allows for supplementation of information in the
bibliographic data (e.g., to provide more complete location information),
"tags" words in the caption and suggests field as being particular parts
of speech, checks the spelling of words in the caption and suggests field,
links logically connected adjacent words in the captions and suggests
field as "multiwords" (e.g., "United States" and "home run"), and removes
ambiguities from the caption and the suggests field by allowing the
captioner to select a word sense that most closely matches the concept or
intended meaning of any particular word in the context. For instance, the
word "crane" has both noun and verb meanings, and the noun meaning is also
ambiguous between a "heavy equipment" sense and a "bird" sense. The
captioner is presented with a list of possible senses and asked to
indicate which sense is intended. From that point on, the word is marked
with the intended sense so that requests for images related to the other
senses of the word do not pull up that image.
The disambiguation tool 206 uses a semantic net of word senses, including a
hierarchy of synonyms and related words. This net acts as a thesaurus to
link related words in both the disambiguation service 206 and in the image
center 120 so that even if the end user does not request "crane", but
instead requests "heron", images captioned using the bird sense of crane
may be retrieved.
The disambiguation tool 206 permits a captioner to add new words and new
senses to the semantic net. In a preferred embodiment, disambiguation tool
206 is implemented using disambiguation processor 114 and NLP database
116. Referring now also to FIG. 4, greater detail is provided concerning
the data flow of disambiguation tool 206.
The data with ambiguities removed is next applied to captions/suggests
field quality control service 208, wherein the captioner is provided with
an opportunity to again check the accuracy of the caption and suggests
field information. In a preferred embodiment, caption/suggests field
quality control service is also implemented using disambiguation processor
114.
Referring now to FIG. 6, there is shown an index card screen 600 by which
data entry for the quality control services 202, 208 and caption/suggests
field entry services 204 may be accomplished. Index card screen 600
displays image 250 in a picture display area 602, allows the captioner to
review and modify bibliographic data in bibliographic data area 610,
allows the captioner to add or review a caption in caption area 604,
allows the captioner to add or revise suggests field information in
suggests field area 606, and allows the captioner to add or revise
photographer's notes in photographer's notes area 608. Screen 600 also
provides an area 612 in which the captioner may specify the image
characteristics (e.g., predominant hue, type of image). Furthermore,
screen 600 provides a set of navigation buttons 618 by which the captioner
may move among images, reject an image that is of faulty quality, or
change default values for the data. In a preferred embodiment, screen 600
is implemented to operate in conjunction in the conventional windowing
environment of a UNIX workstation such as a SUN SPARCSTATION or an IBM
model RS6000 workstation. In a preferred embodiment, a separate screen
similar to screen 600 is used for data entry and correction of pricing and
delivery information for images, but it should be recognized that this
information could be provided on screen 600 if desired.
Referring now to FIG. 7, there is shown an interpreter screen 700 that is
presented to the captioner when the disambiguation tool 206 is invoked by
the captioner's selection of a "next" choice from navigation buttons 618
on screen 600. Interpreter screen 700 includes an area 702 for display of
tokens, or words, in a caption, an area 704 for display of the various
sense choices known in NLP database 116 for a selected word in display
area 702, an area 706 for more detailed interpretation of a selected one
of the sense choices that is displayed in area 704, and an area 712 for
display of the part of speech of the selected token. A user commands area
710 displays button choices that the captioner may invoke to add or
subtract information, to form or break multiwords, and to add new sense
definitions. A navigation commands area 708 displays button choices that
the captioner may invoke to finalize disambiguation selections, to ignore
tokens that are flagged as being potentially ambiguous, or to finish or
cancel a session. In operation, the caption displayed in the tokens area
may include a number of words that the disambiguation tool determines to
be ambiguous, and may also mark selected groups of words as multiword
candidates.
Referring now also to FIG. 4, the process of disambiguation corresponding
to the screens in FIGS. 6 and 7 begins by invoking 401 the disambiguation
tool 206. A conventional spell-checker is then invoked 402 to correct any
spelling errors that appear in the caption. Next, a check 403 is made to
determine whether the caption contains any likely multiwords. If so, the
multiwords are marked 404 by the captioner underlining them on screen 600.
In one embodiment, the ingestion center 110 automatically provides
suggestions for multiwords and provides the captioner an opportunity to
modify those selections as desired. For example, a caption may include the
term "blue collar", and it is up to the captioner to determine whether
this term merely describes the color of a shirt collar and should
therefore be considered as two separate words, or whether it relates to
manual labor, in which case it should be considered as a multiword. The
system may at times present a large number of choices for multiwords. For
example, if the caption includes "Mt. Rushmore National Park, South
Dakota", a number of possible multiwords may be presented, ranging from no
multiwords, to a combination of the multiwords "Mt. Rushmore", "National
Park", and "South Dakota", to a large single multiword containing the
entire phrase. The proper selection is left to the captioner's discretion,
and should be made in a manner that will be most helpful in searching for
images. In the above example, one likely selection would be of the
multiwords "Mt. Rushmore National Park" and "South Dakota".
Processing then invokes 405 a disambiguation/part of speech tagger and
allows the captioner to select a word for processing. A check 406 is made
to determine whether the selected word is recognized, i.e., is known by
the NLP database 116. If not, an unknown word handler is invoked 412 so
that the disambiguation tool 206 enters a learning mode, and the captioner
is prompted 413 to enter the word type, i.e., a proper name, a location,
or other. In a preferred embodiment, unknown words of a caption are
immediately displayed using a different color, e.g., red, from recognized
words to ensure that the captioner provides such unrecognized words with
special attention. In typical operation of the preferred embodiment, all
words in a caption and in a suggests field are selected by the captioner
for disambiguation, but it should be recognized that operation in which
only some words are so processed is also possible.
If the captioner indicates that the type of the unknown word is a name, the
word is stored 415 as a name in the NLP database 116, and processing
returns to 405 for disambiguation of subsequent words in the caption. If
the word type is a location, the word is stored 414 as a location in the
NLP database 116, and processing returns to 405 for disambiguation of
subsequent words in the caption. If the captioner indicated any other word
type, the captioner is prompted to identify 416 the part of speech of the
word (e.g., noun, verb, adjective, adverb, date, keyword, helping word)
and to associate 417 the word with a word that is known by the NLP
database 116. Keywords are acronyms, company names, newly-defined terms in
common usage, slang, and words that do not fall into the other categories.
Examples of keywords might be names of musical groups such as "Peter, Paul
& Mary" and of cultural movements such as "New Age" or "grunge". Helping
(or "function") words are determiners such as prepositions, conjunctions
and possessive pronouns when used in a manner that would not assist in
image searching. For instance, a caption that reads "A boy runs past a
house" should have the word "past" marked as a helping word.
In a preferred embodiment, three types of associations are provided for
unknown words. The first, a "kind of" association, links genus terms with
species terms. For instance, Halloween is a "kind of" holiday. The second
type of association is a synonym association. For instance, the words
"trauma" and "shock" may be linked in this manner. The third type of
association is a "sister term" association, and is used to link two
species within the same genus. For instance, the multiword terms "tank
top" and "tee shirt" would be linked as sister terms. This information is
used to determine the unknown word's placement in NLP database 116. For
instance, if NLP database 116 already recognizes "tank top" as a kind of
shirt, linking "tee shirt" as a sister term for "tank top" establishes tee
shirt" as also being a kind of shirt. It should be recognized that
additional or other types of associations may be provided.
The captioner may use the information stored in NLP database 116 as a
dictionary or thesaurus by free associating other possible words with the
unknown word in interpretation area 706. Once the captioner has typed in a
proposed associated word that is recognized as being in NLP database 116,
the disambiguation tool will allow that proposed associated word to be the
sense with which the unknown word is tagged. For instance, if the word
"biker" appears in a caption and is unknown in NLP database 116, the
captioner may try free associating the term "bicyclist". If that term is
unknown as well, the captioner may try the term "cyclist." If cyclist is
in the NLP database 116, the captioner can choose that "biker" be tagged
with the sense "cyclist" in that caption. Image searching can also be
enhanced by free associating proper names or dates with other terms in NLP
database 116. For example, the captioner may associate the proper name
"Abraham Lincoln" with the noun "president". Similarly, nouns may be
associated with verbs, for instance "explosion" with "combust". As a more
complete example, the captioner may mark the words "boat people" as a
multiword, indicate that the part of speech is a noun, and associate this
multiword with the known term "refugees".
All of the new information provided by the captioner while disambiguation
tool is in a learning mode is recorded in a log for future use in
disambiguation and, once uploaded to image center 120, for use in image
searching.
If check 406 indicates that the word is recognized in the NLP database 116,
then a check 407 is made to determine whether the part of speech assumed
by disambiguation tool 206 for the word is correct. This check 407 is
accomplished by prompting the captioner to indicate whether there is an
error in the assumed part of speech, which is conventionally determined by
word order and statistical information concerning usage of each word. If
there is an error, the captioner indicates 408 the correct part of speech.
If the part of speech is determined 409 to be as a function (or "helping")
word, the word is ignored 410 for purposes of disambiguation and
processing returns to 405. A function word, as opposed to a content word,
is a word that only links together or supports words that describe things,
actions, and properties. For example, content words would include "house",
"walk" or "crooked", while function words would include "the", "and",
"could", and "if". If the word is not a function word, the captioner is
prompted 411 to indicate the correct sense of the word and thereby mark
that instance of the word with the desired sense. This prompting 411 takes
place even if the NLP database 116 is currently aware of only one sense of
the word, in order to give the captioner an opportunity to add a new sense
for that word to the NLP database 116.
Disambiguation tool 206 is implemented in a preferred embodiment in a
conventional manner using disambiguation processor 114 and NLP database
116. Further information on known techniques of natural language
processing for text-only retrieval systems are found, for example, in T.
Strzalkowski and B. Vauthey, Information Retrieval Using Robust Natural
Language Processing, PROCEEDINGS OF THE 30TH ANNUAL MEETING OF THE
ASSOCIATION FOR COMPUTATIONAL LINGUISTICS, 28 Jun.-2 Jul. 1992, Newark,
Del., pp. 104-111; P. Nelson, Site Report for the Text REtrieval
Conference, TREC: The First Text REtrieval Conference (TREC-1), D. K.
Harman, ed., Computer Systems Laboratory, National Institute of Standards
and Technology, Gaithersburg, Md., NIST Special Publication 500-207
(hereinafter, "TREC"), pp. 287-296 (1993); D. Evans, et al., CLARIT TREC
Design, Experiments, and Results, TREC, pp. 251-286 (1993); T.
Strzalkowski, Natural Language Processing in Large-Scale Text Retrieval
Tasks, TREC, pp. 173-187 (1993); S. Abney, Parsing by Chunks,
PRINCIPLE-BASED PARSING: COMPUTATION AND PSYCHOLINGUISTICS, Berwick et
al., eds., Dordrecht: Kluwer Academic Publishers, pp. 257-78 (1991), the
teachings of which are incorporated herein by reference.
In a preferred embodiment, each word sense is represented using a unique
identifying number. An ambiguous word, such as "crane", may have several
senses, and disambiguation refers to selecting the correct sense in a
particular context or, in other words, discarding those senses of the word
that are not appropriate in the current context. It should be recognized
that this disambiguation may be performed either manually, i.e., with the
captioner selecting a proper sense for each word, or may be performed
automatically, e.g., with a system that uses statistical information to
select the most likely sense in a given context.
The output of ingestion center 110 includes image 250 and metadata 262 and
may, as desired, be in the form of a data stream on a bus connecting
ingestion center 110 to image center 120, or may be written onto storage
media such as magnetic or optical disks or tapes.
Referring now to FIG. 3, there is shown a functional block diagram of image
center 120. Image 250 and metadata 262 are applied to uploading,
archiving, watermarking and indexing service 302 for initial processing.
Service 302 transfers full-resolution images, e.g., 250 for long-term
storage onto a conventional medium such as magnetic tape; generates
browse-resolution images, watermarks such images and stores them for
browsing service 308; stores metadata and any additions to the semantic
net resulting from disambiguation for index querying service 306, and
stores licensing and pricing information for use by purchase and delivery
service 310 to permit on-line delivery of a full-resolution image 350. In
a preferred embodiment, separate databases within database processor 124
are used to provide such storage, but it should be recognized that any
conventional storage scheme could be used for storage of the
browse-resolution images, the metadata, the semantic net information, and
the licensing and pricing information.
Still referring to FIG. 3, a user workstation, e.g., 130, communicates with
image center 120 by connection to credentials verification service 304.
Service 304 verifies an user's credentials by checking an input user
identification number, organization identification number, user name, and
password. Users are assigned a permission level to indicate whether they
are authorized only to search for images or to both search for and
purchase rights to images. Service 304 also maintains audit trails of
system usage, such as connect time and login attempts, both for billing
purposes and for tracing attempted unauthorized use of system 100. In a
preferred embodiment, credentials verification service 304 is implemented
partially on user workstation 130 and partially on image center 120,
specifically database processor 124. It should be recognized, however,
that other equivalent implementations could be used to achieve the
function of credentials verification service 304. In an alternative
e | | |
