What is a "Humming Window" (Digital Image or Speach Processing ?)

[DM2]

I'm doing some research that involves digital processing of video images. There is a statement in one manufacturers patent (US 6184792) which states:

For the purpose of the present embodiment, these measures are computed by performing a Fast Fourier Transform (FFT) on the temporally varying, pixel intensities. The measure of the static component is taken to be the Zero FFT term, (i.e., mean brightness value), While the sum of the three FFT terms centered around 5 HZ are taken as the measure of the dynamic component. However, similar end results were obtained when using Digital Signal Processing techniques with Humming Windows (that is not to suggest that Humming Window is the only technique possible). In addition, the dynamic component can be determined by simply counting how many times the intensity signal crosses its mean value within each analysis cycle.

I attempted to Google the term, but with little luck. I did find a PDF publication on the web titled [link gendocs.ru/docs/10/9175/conv_1/file1.pdf]"Speech Technologies"[/url], where on pdf page 181 the term is used in the context of:

Frame Length 25 ms. (Humming Window)

Regards,
DM

"Real world Knowledge isn't dropped from a parachute in the sky but rather acquired in tiny increments from a variety of sources including panic and curiosity."

 

[geekEE]

Sounds like they meant Hamming Window

 

[IRstuff]

Windows galore:

http://en.wikipedia.org/wiki/Hamming_window#Hamming_window

Aside from the quaint application described in the patent, I can't really see this being a practical solution. What exactly are you trying to do?

TTFN
faq731-376

Need help writing a question or understanding a reply? forum1529

Of course I can. I can do anything. I can do absolutely anything. I'm an expert!

 

[2dye4]

As far as an explanation of what a hamming window is or does here is a go at it.

When a FFT is performed it operates on only a slice of the data ( equal to the length fed in ).

This means that the original data was effectively multiplied by a square window of ones to get the
data for the FFT. Unless this window coincides with an exact period of a completely periodic data
sequence then it represents a distortion of the data.

The hamming window softens this distortion because it rolls off near each end toward zero at the end
and causes less disturbance in the freq domain result.

http://www.ni.com/white-paper/4844/en/

For images in two dimensions it is called a point spread function.

http://en.wikipedia.org/wiki/Point_spread_function

 

[IRstuff]

I think the OP is actually talking about a time series analysis of each pixel.

TTFN
faq731-376

Need help writing a question or understanding a reply? forum1529

Of course I can. I can do anything. I can do absolutely anything. I'm an expert!

 

[DM2]

Thanks for everyone's response. Both of the links I provided, spell the word as "Hu...", not "Ha...". Based on the reading of several of the internet links provided by others, I'm inclined to believe it is "Ha..".

To put this into perspective, the patent is for a video based smoke and flame detection system. According to the manufacture, the system works by taking an image from a camera, and processing that image to see if a specific pixel in a given frame, is brighter than the same pixel in the following frame. For the sake of this discussion, let's assume the manufacture has all of their algorithms right so they don't get false positives from some flipping on and off a light switch, or playing with a dimmer switch (It would be interested to see how the system deals with a failing fluorescent light however).

My background is mechanical so you guys may have to dumb it down a little further for me.

Here is my understanding, in my words, how the process works...
1. The camera is providing a video feed, that's captured and broken down into frames which are converted to bitmaps.
2. These frames/bitmaps (i.e. 1, 2, 3, 4, 5) are examined for individual pixel change in brightness (mind you the term brightness is what's sighted in the patent)
3. The process is repeated with the next set of frames (2, 3, 4, 5, 6) and the process is repeated.
4. If a change is detected in the pixel brightness over several sets of frames, a "Fire Alarm" is triggered.

Is each sent of frames (1 to 5) the "Hamming Window"?

Regards,
DM

"Real world Knowledge isn't dropped from a parachute in the sky but rather acquired in tiny increments from a variety of sources including panic and curiosity."

 

[IRstuff]

No! A window is what you apply to the time series. However, an FFT is bit extreme for something like this, considering that you have to do this on each, and every, pixel in the camera, AND re-do everything the very next frame.

TTFN
faq731-376

Need help writing a question or understanding a reply? forum1529

Of course I can. I can do anything. I can do absolutely anything. I'm an expert!

 

[DM2]

I just noticed the grammar error, it should have said "Is each set of frames (1 to 5) the "Hamming Window". Am I more on track now?

I'm not sure they're using FFT in the process. They may be using a some other method of identifying the change in pixel color, brightness, etc, or may be somehow averaging it. I also believe they're using edge detection in someway as well.



Regards,
DM

"Real world Knowledge isn't dropped from a parachute in the sky but rather acquired in tiny increments from a variety of sources including panic and curiosity."

 

[DM2]

further...
I say "...set..." because each frame is taken in a give time period...
1.00000 Second mark
1.03125
1.06250
...
...
...
1.31250 (frame 10 of 32 frames / second)

Regards,
DM

"Real world Knowledge isn't dropped from a parachute in the sky but rather acquired in tiny increments from a variety of sources including panic and curiosity."

 

[IRstuff]

No! A Hamming window is a series of coefficients that you multiply against the time series.

TTFN
faq731-376

Need help writing a question or understanding a reply? forum1529

Of course I can. I can do anything. I can do absolutely anything. I'm an expert!

 

[MacGyverS2000]

Take pixel (0,0) from frames 0-31. A pixel can be 256 shades of gray (an 8 bit value), so you have a string of 32 8-bit bytes. The Hamming window will also be 32 bytes in length, with each fractional value in the window ranging from 0.0 to 1.0. The window will be mostly 1.0s in the middle, tapering to 0.0 at the ends. You apply the window by multiplying the first pixel with the first Hamming window value, the second with the second, etc. until you have a new 32-byte long sequence of numbers. That new sequence is the original pixel sequence smoothed at the ends by the Hamming window.

There are other window shapes, but the process is the same.

Dan - Owner

http://www.Hi-TecDesigns.com

 

[DM2]

MacGyverS2000 and IRstuff,
Thanks for your help, now I have a better understanding.

Regards,
DM

"Real world Knowledge isn't dropped from a parachute in the sky but rather acquired in tiny increments from a variety of sources including panic and curiosity."

 

[John2025]

Thanks for explanation MacG. If I understand, the window function basically weights the center of the time span values to be more important, but I'm wondering what the purpose would be. If the process is cyclic, like 2dye4 mentioned, how would you know the center is better? Maybe a triggered sensor? Or is it simply to let the sensor to settle?

 

[BrianE22]

If there is a complete cycle (no fractions) of every frequency in the signal then you would not need windowing. For example, if you have 3/4 of a complete cycle of a 10 hertz signal and you string together those waveforms you have a big discontinuity at the junctions. The FFT would give you content at 10 hertz and also the frequency content of the discontinuity.

 

[IRstuff]

> The second link in the OP references Hamming window 15 times, while Humming window is used only once, so, clearly a typo

> An FFT really not ideal for detecting a singular, aperiodic event. Fourier analysis is really about PERIODIC functions, which is why a Hamming, or other, window is necessary, since you are transforming a time-domain signal into a frequency domain signal.

> I would guess that many fires do not assert themselves within a 1, or even 10, second timeframe. Smoldering fires build very slowly over time.

TTFN
faq731-376

Need help writing a question or understanding a reply? forum1529

Of course I can. I can do anything. I can do absolutely anything. I'm an expert!

 

[DM2]

IRstuff,

many fires do not assert themselves within a 1, or even 10, second timeframe

Your statement hold true, if the intended purpose of the device is for "SMOKE" detection. The pass fail criteria for flame detection is to detect the fire within 30 seconds due to the speed at which some fires propagate (fuel dependent). This particulate device started out as a "Smoke" detector, and is being marketed to also detect flames. It therefore is tested against different standards and must respond within 30 seconds. Most flame detectors respond within 10 seconds.

My research is to support a white paper on flame detectors. Image based detection systems started their development cycle in the early 90's. At present there is only one manufacture that has developed an image based product proven to be useful for flame detection (the patent link above is not from that manufactures). Marketing literature and video's from the referenced company seem to demonstrate the technologies capability, however they have a low installation base. Further, hardware that the detector is centered on needs some, improvement in order to be used in the typical harsh environments and explosive atmospheres a typical flame detector is installed in.

Regards,
DM

"Real world Knowledge isn't dropped from a parachute in the sky but rather acquired in tiny increments from a variety of sources including panic and curiosity."

 

[IRstuff]

Flames have other characteristics. Using only temporal intensity isn't the only option. Running FFTs on every single pixel is very processor intensive.

TTFN
faq731-376

Need help writing a question or understanding a reply? forum1529

Of course I can. I can do anything. I can do absolutely anything. I'm an expert!

 

[2dye4]

Seems to me you should pass the frames through a differentiator on frame to frame basis.

IE each pixel out of the differentiator is the difference between the previous intensity for that pixel and the current intensity.

The on this frame that is the differentiated one a smoke event as opposed to an overall light intensity change
would be reflected in an uneven distribution in the differentiated values.

I would calculate the mean of the differentiated values then evaluate the percentage of pixels exceeding this value
by two or three standard deviations of the pixel noise. If the image has changed in only a fraction of the pixels
then test to see if they lie in a region. If so sound the alarm.

Of course you need much more testing and refinement.

 

[MacGyverS2000]

Windowing is not the same as weighting. With weighting, you re considering one portion of the signal to be more important than another. The purpose of windowing is to shape the signal for further processing... in this case, you are reducing end points of each frame to a common value (zero) to avoid discontinuities during processing.

Dan - Owner

http://www.Hi-TecDesigns.com

 

[2dye4]

The hamming window is applied to the sequence of frames on each pixel. Think of it as N separate time sequences
that each get multiplied by the hamming window where N is the number of pixels.

The purpose is to keep artifacts that result from just picking a window of data and not using a very very long data sequence.

The artifacts would be false frequency values appearing in the output which could trigger the detection.

The hamming window significantly reduces the artificial frequencies that would appear.

One part of the detection is looking for a flicker rate that is assumed to be relatively constant across many instances of fire.

This sounds like fun work. I will trade ya ..