FPGA Central

Hello DSP gurus.
I am busy researching the possibilities of detecting and recognizing text in
motion video (MPEG/WMV stream).
So far i've found lots op university-projects and a hand full of commercial
applications that kind-of do what I seek.

I am not interested in spending months/years of research and development to
build my own software. I am looking for a currently stable and functional
SDK to provide my already existing A/V research and archiving software with
text-recognition.
I seek for a system that does a recognition once, so I can put the metadata
in a database for later use (in information searches).

Is anyone here familiar with this, and can point me in the right direction?

I'd appreciate any response!

Best regards,

Rob Vermeulen
Arbor Audiocommunications BV
rvermeulenatarbor-audiodotcom

I have no experience with this, but SRI's ConTEXTract
may be worth a look. I heard about it when I was
browsing Virage's web site a while ago (they were
offering it bundled with one of their products).

http://www.esd.sri.com/automation/video_recog.html

In article <[email protected]>,
Rob Vermeulen <rvermeulen@arbor-audio-antispam-.com> wrote:
>Hello DSP gurus.
>I am busy researching the possibilities of detecting and recognizing text in
>motion video (MPEG/WMV stream).
>So far i've found lots op university-projects and a hand full of commercial
>applications that kind-of do what I seek.
>
>I am not interested in spending months/years of research and development to
>build my own software. I am looking for a currently stable and functional
>SDK to provide my already existing A/V research and archiving software with
>text-recognition.
>I seek for a system that does a recognition once, so I can put the metadata
>in a database for later use (in information searches).
>
>Is anyone here familiar with this, and can point me in the right direction?
>
>I'd appreciate any response!
>
>Best regards,
>
>Rob Vermeulen
>Arbor Audiocommunications BV
>rvermeulenatarbor-audiodotcom
>
>
>
>

Thanks,

I already found that one while googling around. Indeed, quite impressive.

Cheers,

Rob
"David Gelbart" <[email protected]> wrote in message
news:ceo12r$6hs$[email protected]..
> I have no experience with this, but SRI's ConTEXTract
> may be worth a look. I heard about it when I was
> browsing Virage's web site a while ago (they were
> offering it bundled with one of their products).
>
> http://www.esd.sri.com/automation/video_recog.html
>
> In article <[email protected]>,
> Rob Vermeulen <rvermeulen@arbor-audio-antispam-.com> wrote:
> >Hello DSP gurus.
> >I am busy researching the possibilities of detecting and recognizing text
in
> >motion video (MPEG/WMV stream).
> >So far i've found lots op university-projects and a hand full of
commercial
> >applications that kind-of do what I seek.
> >
> >I am not interested in spending months/years of research and development
to
> >build my own software. I am looking for a currently stable and functional
> >SDK to provide my already existing A/V research and archiving software
with
> >text-recognition.
> >I seek for a system that does a recognition once, so I can put the
metadata
> >in a database for later use (in information searches).
> >
> >Is anyone here familiar with this, and can point me in the right
direction?
> >
> >I'd appreciate any response!
> >
> >Best regards,
> >
> >Rob Vermeulen
> >Arbor Audiocommunications BV
> >rvermeulenatarbor-audiodotcom
> >
> >
> >
> >
>
>

Rob Vermeulen wrote:

> Hello DSP gurus.
> I am busy researching the possibilities of detecting and recognizing text in
> motion video (MPEG/WMV stream).
> So far i've found lots op university-projects and a hand full of commercial
> applications that kind-of do what I seek.
>
> I am not interested in spending months/years of research and development to
> build my own software. I am looking for a currently stable and functional
> SDK to provide my already existing A/V research and archiving software with
> text-recognition.
> I seek for a system that does a recognition once, so I can put the metadata
> in a database for later use (in information searches).
>
> Is anyone here familiar with this, and can point me in the right direction?
>
> I'd appreciate any response!
>
> Best regards,
>
> Rob Vermeulen
> Arbor Audiocommunications BV
> rvermeulenatarbor-audiodotcom
>

I would suggest posting also to comp.speech.research .
IIRC there was recent discussion of identifying speech occurring in
presence of other audio.

Thanks for your reply,

but what has this got to do with speech recognition ? :-)
Perhaps you misinterpret the word 'Text' which in this case means graphical
written (drawn/rendered) words. I am not looking for a speech recognition
algorithm; this I already have.

But any input is welcome!

Cheers,

Rob

"Richard Owlett" <[email protected]> wrote in message
news:[email protected]..
> Rob Vermeulen wrote:
>
> > Hello DSP gurus.
> > I am busy researching the possibilities of detecting and recognizing
text in
> > motion video (MPEG/WMV stream).
> > So far i've found lots op university-projects and a hand full of
commercial
> > applications that kind-of do what I seek.
> >
> > I am not interested in spending months/years of research and development
to
> > build my own software. I am looking for a currently stable and
functional
> > SDK to provide my already existing A/V research and archiving software
with
> > text-recognition.
> > I seek for a system that does a recognition once, so I can put the
metadata
> > in a database for later use (in information searches).
> >
> > Is anyone here familiar with this, and can point me in the right
direction?
> >
> > I'd appreciate any response!
> >
> > Best regards,
> >
> > Rob Vermeulen
> > Arbor Audiocommunications BV
> > rvermeulenatarbor-audiodotcom
> >
>
> I would suggest posting also to comp.speech.research .
> IIRC there was recent discussion of identifying speech occurring in
> presence of other audio.
>

Ooops. You right.
I'm biased towards problems *I* wish to solve ;]

Now do you have lead on end user "phoneme recognizer" as opposed to
"speech recognizer"?

Rob Vermeulen wrote:
> Thanks for your reply,
>
> but what has this got to do with speech recognition ? :-)
> Perhaps you misinterpret the word 'Text' which in this case means graphical
> written (drawn/rendered) words. I am not looking for a speech recognition
> algorithm; this I already have.
>
> But any input is welcome!
>
> Cheers,
>
> Rob
>
> "Richard Owlett" <[email protected]> wrote in message
> news:[email protected]..
>
>>Rob Vermeulen wrote:
>>
>>
>>>Hello DSP gurus.
>>>I am busy researching the possibilities of detecting and recognizing
>
> text in
>
>>>motion video (MPEG/WMV stream).
>>>So far i've found lots op university-projects and a hand full of
>
> commercial
>
>>>applications that kind-of do what I seek.
>>>
>>>I am not interested in spending months/years of research and development
>
> to
>
>>>build my own software. I am looking for a currently stable and
>
> functional
>
>>>SDK to provide my already existing A/V research and archiving software
>
> with
>
>>>text-recognition.
>>>I seek for a system that does a recognition once, so I can put the
>
> metadata
>
>>>in a database for later use (in information searches).
>>>
>>>Is anyone here familiar with this, and can point me in the right
>
> direction?
>
>>>I'd appreciate any response!
>>>
>>>Best regards,
>>>
>>>Rob Vermeulen
>>>Arbor Audiocommunications BV
>>>rvermeulenatarbor-audiodotcom
>>>
>>
>>I would suggest posting also to comp.speech.research .
>>IIRC there was recent discussion of identifying speech occurring in
>>presence of other audio.
>>
>
>
>

> Now do you have lead on end user "phoneme recognizer" as opposed to
> "speech recognizer"?

I use the Nexidia SDK (formerly known as Fasttalk) for speech recognition.
Indeed this is phoneme based recognition which is language/dialect
dependent.

I haven't got the capacity & time to develop my own. Although it is very
interesting matter.
But I do like to combine technologies in to something that is worth twice as
much as the sum of the parts :-)

hth,

Rob

> Rob Vermeulen wrote:
> > Thanks for your reply,
> >
> > but what has this got to do with speech recognition ? :-)
> > Perhaps you misinterpret the word 'Text' which in this case means
graphical
> > written (drawn/rendered) words. I am not looking for a speech
recognition
> > algorithm; this I already have.
> >
> > But any input is welcome!
> >
> > Cheers,
> >
> > Rob
> >
> > "Richard Owlett" <[email protected]> wrote in message
> > news:[email protected]..
> >
> >>Rob Vermeulen wrote:
> >>
> >>
> >>>Hello DSP gurus.
> >>>I am busy researching the possibilities of detecting and recognizing
> >
> > text in
> >
> >>>motion video (MPEG/WMV stream).
> >>>So far i've found lots op university-projects and a hand full of
> >
> > commercial
> >
> >>>applications that kind-of do what I seek.
> >>>
> >>>I am not interested in spending months/years of research and
development
> >
> > to
> >
> >>>build my own software. I am looking for a currently stable and
> >
> > functional
> >
> >>>SDK to provide my already existing A/V research and archiving software
> >
> > with
> >
> >>>text-recognition.
> >>>I seek for a system that does a recognition once, so I can put the
> >
> > metadata
> >
> >>>in a database for later use (in information searches).
> >>>
> >>>Is anyone here familiar with this, and can point me in the right
> >
> > direction?
> >
> >>>I'd appreciate any response!
> >>>
> >>>Best regards,
> >>>
> >>>Rob Vermeulen
> >>>Arbor Audiocommunications BV
> >>>rvermeulenatarbor-audiodotcom
> >>>
> >>
> >>I would suggest posting also to comp.speech.research .
> >>IIRC there was recent discussion of identifying speech occurring in
> >>presence of other audio.
> >>
> >
> >
> >

Rob Vermeulen wrote:

> Hello DSP gurus.
> I am busy researching the possibilities of detecting and
> recognizing text in motion video (MPEG/WMV stream).
> So far i've found lots op university-projects and a hand full of
> commercial applications that kind-of do what I seek.
>
> I am not interested in spending months/years of research and
> development to build my own software. I am looking for a currently
> stable and functional SDK to provide my already existing A/V
> research and archiving software with text-recognition.
> I seek for a system that does a recognition once, so I can put the
> metadata in a database for later use (in information searches).
>
> Is anyone here familiar with this, and can point me in the right
> direction?
>
> I'd appreciate any response!
>
> Best regards,
>
> Rob Vermeulen
> Arbor Audiocommunications BV
> rvermeulenatarbor-audiodotcom

Maybe it's requiring too much work from your side, but it might be
at least worth mentioning...

If your software is able to provide a still image of the text, then
the OCR tools which are used for scanners, might be applicable.
I think of "gocr" which should be integrable.

Bernhard

Bernhard,

Thanks for this.
I have been searching for OCR SDK's the past few days but gocr did not show
up in Google. I'll look it up and see if it is usable.

Yes, I can produce still images which I can feed to an OCR algorithm. I
already tested it on several algorithms, even on my own written OCR routine
(which is basically simple, I found out).

The problem with OCR is that the algoritm only works accurately when text is
placed on a solid background, which isn't the case in video material most of
the time. I want to detect subtitles and other "overlayed text" but also
"scene text" such as license places on cars and company logos on buildings.
The text in the last category can also appear in every angle, rotated in
every direction and even in perspective.
What I need is more than just OCR. It must first do text-detection,
classification, de-blurr filters and other preprocessing things before it
recognizes characters.

But I haven't looked into gocr yet, so it still might surprise me :-)

I'm still open for other suggestions.

Cheers,

Rob


"Bernhard Holzmayer" <[email protected]> wrote in message
news:[email protected]..
>
> Maybe it's requiring too much work from your side, but it might be
> at least worth mentioning...
>
> If your software is able to provide a still image of the text, then
> the OCR tools which are used for scanners, might be applicable.
> I think of "gocr" which should be integrable.
>
> Bernhard

Rob Vermeulen wrote:

...

> The problem with OCR is that the algoritm only works accurately when text is
> placed on a solid background, which isn't the case in video material most of
> the time. I want to detect subtitles and other "overlayed text" but also
> "scene text" such as license places on cars and company logos on buildings.
> The text in the last category can also appear in every angle, rotated in
> every direction and even in perspective.
> What I need is more than just OCR. It must first do text-detection,
> classification, de-blurr filters and other preprocessing things before it
> recognizes characters.

There may be image enhancement processes that you could use to isolate
the letter outlines. Once you have that, you can put it on any
background you like. Some OCR programs may be able to use that
information directly. I can't begin to estimate the resources needed.
They could well be excessive.

Jerry
--
Engineering is the art of making what you want from things you can get.
¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯ ¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯

Jerry Avins wrote:

> Rob Vermeulen wrote:
>
> ...
>
>> The problem with OCR is that the algoritm only works accurately when
>> text is
>> placed on a solid background, which isn't the case in video material
>> most of
>> the time. I want to detect subtitles and other "overlayed text" but also
>> "scene text" such as license places on cars and company logos on
>> buildings.
>> The text in the last category can also appear in every angle, rotated in
>> every direction and even in perspective.
>> What I need is more than just OCR. It must first do text-detection,
>> classification, de-blurr filters and other preprocessing things before it
>> recognizes characters.
>
>
> There may be image enhancement processes that you could use to isolate
> the letter outlines. Once you have that, you can put it on any
> background you like. Some OCR programs may be able to use that
> information directly. I can't begin to e stimatetheresourcesneeded.
> They could well be excessive.
>
> Jerry
???

I can't begin to estimate the resources needed.

What happened?

Jerry
--
Engineering is the art of making what you want from things you can get.
¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯ ¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯

Jerry Avins wrote:

> Jerry Avins wrote:
>
>> Rob Vermeulen wrote:
>>
>> ...
>>
>>> The problem with OCR is that the algoritm only works accurately when
>>> text is
>>> placed on a solid background, which isn't the case in video material
>>> most of
>>> the time. I want to detect subtitles and other "overlayed text" but also
>>> "scene text" such as license places on cars and company logos on
>>> buildings.
>>> The text in the last category can also appear in every angle, rotated in
>>> every direction and even in perspective.
>>> What I need is more than just OCR. It must first do text-detection,
>>> classification, de-blurr filters and other preprocessing things
>>> before it
>>> recognizes characters.
>>
>>
>>
>> There may be image enhancement processes that you could use to isolate
>> the letter outlines. Once you have that, you can put it on any
>> background you like. Some OCR programs may be able to use that
>> information directly. I can't begin to e stimatetheresourcesneeded.
>> They could well be excessive.
>>
>> Jerry
>
> ???
>
> I can't begin to e stimatetheresourcesneeded.
>
> What happened?

Again!

I can't begin to estimate the resources needed.

I can't begin to
estimate
the
resources
needed.

Jerry
--
Engineering is the art of making what you want from things you can get.
¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯ ¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯

On 2004-08-04 20:30:05 +0200, Jerry Avins <[email protected]> said:

> Jerry Avins wrote:
>
>> Jerry Avins wrote:
>>
>>> There may be image enhancement processes that you could use to isolate
>>> the letter outlines. Once you have that, you can put it on any
>>> background you like. Some OCR programs may be able to use that
>>> information directly. I can't begin to e stimatetheresourcesneeded.
>>> They could well be excessive.
>>>
>>> Jerry
>>
>> ???
>>
>> I can't begin to e stimatetheresourcesneeded.
>>
>> What happened?
>
> Again!
>
> I can't begin to estimate the resources needed.
>
> I can't begin to
> estimate
> the
> resources
> needed.
>
> Jerry


Maybe you have some weird text enhancement processes messing up your
posting... :-)

Btw: only the quoted text appears messed up, the original looks good on
my end (individual.net).

Cheers,
Stephan

--
Stephan M. Bernsee
http://www.dspdimension.com

"Jerry Avins" <[email protected]> wrote in message
news:4110fdd0$0$2825$[email protected]..

> There may be image enhancement processes that you could use to isolate
> the letter outlines. Once you have that, you can put it on any
> background you like. Some OCR programs may be able to use that
> information directly. I can't begin to estimate the resources needed.
> They could well be excessive.

I'm familiar with image processing / filtering methods and I think I could
figure out a way to separate text from the rest of the image.
(Probably a combination of edge detection & classification).
Perhaps this will make gocr work better, because I only got garbage out when
feeding it with frames from my video database.

But I think there's a different approach necessary. Some approach that I'm
not familiar with.

And there's also the 'speed' component. I'd prefer if a routine worked
realtime and even faster. I've already got an archive of 10000 hours of
video which need to be searched. I'd hate to see that it takes 20000 hours
to process 10000 hours of material.

Anyway, its worth sorting this out and I like the matter, so I'll be doing
research on this for quite some time, I think :-)

If someone else can come up with good suggestions, I'd appreciate it.

Best regards,

Rob