Other types of image search

http://www.xcavator.net/
http://www.fotosearch.com/
http://www.yangsky.com/products/picseer/index.htm

Some famous image search engines

http://www.faganfinder.com/img/
http://www.search-engine-index.co.uk/Images_Search/
http://images.google.com/
http://www.picsearch.com/
http://www.altavista.com/image/default
http://www.ask.com/?tool=img
http://www.exalead.com/image/results?q=
http://www.pixsy.com/
http://www.netvue.com/
http://www.airtightinteractive.com/projects/simple_image_search/app/
http://www.ithaki.net/images/
http://yotophoto.com/

Learning from Small Number of Examples

In this project, there is a need that user may select several images as query images. I choose discrimintive model as our approach.
Based on the libsvm provided by Prof. Lin Chih-Jen, the algorithm is listed as below:
1. Read the checked images and un-checked images.
2. The checked images are regarded as positive examples. And the un-checked images and other images in dataset are regarded as pseudo-negative examples.
3. If the number of positive examples is N, then 2N pseudo-negative examples are generated.
4. For each loop do
- Construct a bag (training set) that contains N positive examples and 2N pseudo-negative examples.
- Choose the whole image dataset as testing set.
- Call svm_train to generate the model. Due to some parameters (g, C) are adjustable, I use the tools provided by Prof. Lin Chih-Jen to find out the best performance.
- Use the predict call to generate the classification results. The prob. parameter must be set to active.
- Collect the probability table.
5. Use MIN fusion to obtain the final result.
6. Filter out the positive examples in descending order and output to system.

In order to incrementally update tag-tag and image-image relations, we implement an efficient function to modify related tables.

Usage Guide:call the function: Add_NewTag($imageID, $flickrID, $tag)
(Please include the file "common/AddNewTag.php" first.)
You have to provide $imageID or $flickrID.
Besides, the name of added $tag is also needed.

Seven tables will be updated when one tag is added into an image.
Tag related tables:
* TBL_ImageTag
* TBL_InvertedFile
* TBL_TagTagMatrix
* TBL_TagRelationSVD

Image related tables:
* TBL_InvertedFileTrans
* TBL_ImageImageMatrix
* TBL_ImageRelation

All tables can be updated within 1~2 seconds.

- Chi-Yao

Rent Progress

Following items are finished.

Tag-Tag Relation
Image-Image Relation

Incremental Upate of adding new tags will be done soon.

Singular Value Decomposition

Reference Website: http://www.alglib.net/matrixops/general/svd.php

Final Image Categories in the Database

each type 500 images:

1. scenery
2. landscape
3. landmark
4. night

5. street
6. building
7. river
8. lake
9. beach
10. mountain
11. forest
12. sunset
13. sunrise
14. sky

The first 4 categories are somewhat general.