Lunch seminar

Finally I did my presentation in this afternoon.

It is not a big deal that my bosses have to be in the meeting. But they insisted on their presence. So, we postponed our lunch seminar many times (>3). It is about how to code robust and accurate.

There is nothing to say about the presentation. I didn't practice much, except reading it aloud to my 11 months baby in the morning, which was stopped by my wife because she thought it was a torture to the baby as well as herself.

The fact proves that my English isn't good enough to present the idea without preparation. I'd better spend more time on this if I really want to improve my speaking skill.

Anyway, my boss gave me the job to lead the lunch seminar, which is exactly what I want.

Some interesting readings I've done in the last week

One topic is about Web 2.0. I didn't know much about it. However, this is a post regarding to use Web 2.0 for a successful business. Although I was quite disappointed with my ability to start a new business, at least I can dream. So, list it here.

http://www.thisisgoingtobebig.com/2005/08/10_steps_to_a_h.html

Regarding to this article, a Chinese blogger reviewed many chinese sites for the future development:

http://in.comengo.net/index.php/archives/web20-hero-in-china/

With the links on those 2 articles, I found an interesting website called del.icio.us, which can be used to tag your interesting blogs.

Anyway, it is a very productive browsing. Open more minds to me, give me more dreams. The bad thing is that I lose my focus again: what's my working progress?

Xu..., my boss doesn't know this, keept secret!

Image normalization

There are 3 ways to normalize regular data:
1. Get the maximum and minimum from the dataset, follow the equation to normlize each element (x - min)/(max - min).

2. Use a known maximum value to normalize data: x/Max;

3. Z normalization: estimate the mean and standard deviation, and normalize the population to a normal distribution with 0 mean and 1 standard deviation: (x - mean)/deviation.

But how to normalize images?

The usage is given as follows:

We have a classifier optimized on certain type of images, which means the histogram the classifier can handle is determined to that type of distribution. For images from other sources, we need to transform the original image to an image that fit the histogram of the trained source. Now, we get a desired image and an original version of thesame image, use histogram specification to generate a look up table to do the remapping.

A question ticked in: after transforming, the useful features whitened out because histogram specification tries to even the histogram, thus those features with with less population get compressed.

What shall I do? use linear transformation? Is it a good idea?