Haar级联分类器训练问题

Question

I've created multiple haar cascaded classifiers of face. I used a different number of positives and negatives each time.

For example,

1st classifier: 5000 positive and 3000 negatives
2nd classifier: 3000 positive and 3000 negatives (deleted 2000 redundant/similar images)

the efficiency of both these classifiers was almost same...

Problems:

1. Isn't there a method by which I can delete all redundant images in my database prior to training?

2. What are the ideal lighting and background conditions for training Classifier?

3. How many images in database are considered ideal for best performance or does it depend on the type of data in the set?

Regards,

Saleh...

Answer 1

All the best for your work.

Answers:

1. I want to know how you have deleted the redundant images while training the second classifier. I can not tell you the exact solution. One solution could be : Take a simple haar feature, get the feature vectors (say F1 and F2) for two images . If the correlation between the F1 and F2 is zero (or less than some threshold), then images are similar. You have to test this. And if it works then please let me know.

2. It depends on the application. If one wants to use a classifier in a scenario with changing illuminations and backgrounds, then such images should be included in training classifiers.

3. Training database should contain many images (typically thousands of images). The important thing is the variations in the images in terms of appearance, illusion, shadows etc. Variation in database makes classifier more robust.