Previous | Next --- Slide 43 of 70
Back to Lecture Thumbnails
BryceSummers

Consider different types of visual media, including Films, photography, comics, line drawings etc. Which of them tend to have lower frequency components? Which types of media tend to have high frequency signals? If you were implementing a classification algorithm for flickr or some other website, then do you think you could use the fourier transform to determine which type of media a given image is from?

Khryl

I think it all comes down to how complex the image is. Real word scenes such as films and photography tend to have more objects in the image and are usually more vivid and colorful, so they will have larger proportion of high frequency signals. Comics and line drawings tend to be less sophisticated and use less color, and they will have larger proportion of low frequency signals.

I think we can use fourier transform to determine the type of media. But the result may not be very accurate. Because sometimes films and photography can be monotonous while comics are very sophisticated and colorful. I think it's better to treat the frequency information as additional features when implementing a classification algorithm.

BryceSummers

@Khryl I don't think that the colorfulness of a scene does not necessarily impact the relative magnitude of the frequencies. The amount of objects also does not necessarily coorelate with the frequency bins as well.

Consider the images above on this webpage, does it look like the "real world" image of Professor Kayvon has more low or high frequency signals?

Khryl

Hi Bryce, I think color is important because in photography the scene(the signal) is more densely and accurately sampled comparing to comics. For example, a person's shirt may contain various kinds of red color, photography can tell them apart thus forming more edges while comics may brush the shirt with single color. In this sense, photography will contain more details and edges. If comics can also depict the same details, then I guess their difference will be insubstantial.

BryceSummers

Hello @Khryl, you are on definatly right about the edges being important.

Consider the following image that has only low frequency bands: Low Frequency

Also, consider the following image with large amounts of high frequency bands: High Frequency

The first image looks like it has quite a few more edges than the latter. Does this support your hypothesis?

Low frequency signals usually manifest in the coloring of an image, whereas high frequency signals manifest in the steepness of the changes in color.

BryceSummers

Here is a wikipedia site about a one dimensional waveform that has a very high energy frequency spectrum. Square Wave

Khryl

@Bryce, I see your point. I guess it's because in LowFrequency the edge color change is not as steep as it seems like. But in HighFrequency the edge color change is very steep and the slope can be infinite which causes very high frequency.

BryceSummers

@Khryl Wonderful! I very much like your observation about the potentially infinite slope. I also like your wording, especially the subtlety "not as steep as it seems".

Here is a restatement of the original question:

When comparing real world images like professor Kayvon with realistic lighting ( or realistic synthesized images ) to drawn images such as comics, is there a substantial difference apparent in the frequency spectrums?

(Related Hint / Observation : Flikr often mis-classifies my rendered images, probably because it assumes that everything is a photograph. )

Khryl

@Bryce, these images really awaken me. The strokes in comics create very steep and distinct edges (most obviously from the shadow on muscles), while realistic synthesized images has much smoother edge on shadows and they will never render shadows as comics do. I guess the difference in frequency spectrum between realistic image and comics is not substantial because they both have steep edges.

BryceSummers

@Khryl ;-)