MMAI 09 Final Project

高紹航, 戴瑋彥,李佳蓉

Our goal is to re-ranking the image search results from image search engines like google image search & flickr, based on image quality, i.e. the images with higher quality will present to users in the search result.

We'll generate some training data from web, and extract features from images. Possible features like the exposure value, EXIF data, color distributions, face scale & positions..etc.

dataset: http://lear.inrialpes.fr/~jegou/data.php
label: good, fair, bad

auto exposure
http://www.imageval.com/public/Products/ISET/ISET_Introduction/AutoExposure.htm
(反推圖片是under/over exposed & 離ideal exposure的distance)

RGB→HSV
H:色相
S:飽和度
V:亮度

平均飽和度
平均亮度

Contrast
http://en.wikipedia.org/wiki/Contrast_%28vision%29

模糊程度 (?)

色彩數量

以上都可以分整張圖, 中央, 整張圖分九塊取最高/最低

supervised or unsupervisd，今晚你要選哪一道?
example of quality re-ranking ***
把scope 縮小成spot or object quality → 對風景或人物或景點或有名的東西做quality re-rank

有篇paper叫做「把我們拍的照片學習成攝影大師拍的照片的paper」，來自SIGGRAPH 2006，或許可以用?
color features
texture features

看起來效果會很差，還是不要好了QQ

imageCLEF2008→似乎都是concept detection的task，還是不要好了QQ
自己撈→人要label，應該不難

user study, human expert, labeling by 廉價研究生
p@n, NDCG

If a searcher searches images about “golden bridge” and types it into Google Image Retrieval, he is likely to get the following ranking results as in the first column: We can see that the ranking result is not very good, as we can see the landmark Golden Bridge and the watch Golden Bridge are interleaved in a way somehow contradicting to human’s expectation. If we can re-rank it into the second column, we can see that, for the landmark Golden Bridge, if the image contains the specific sharp shape of the bridge, it is more likely to meet human’s expectation, and thus it ranked topper. Also, for watch Golden Bridge, a image would be ranked topper if it captures the unique center of the watch. In addition to the re-rank, we can see that in the third column, the images are separated into 2 categories, the landmark and the watch. If we can also do a clustering technique in the search result, it would make it easier for searchers to see their search results.

大師的照片，try feature，看feature表現與顯著程度→猜測哪些feature重要

EXIF的資訊
清晰度
亮度
Color Distribution

哪些feature重要?
1) label一堆去train
2) 看variance
3) 看histogram分佈(反差大→黑的跟白的很多?)
4) 描述object的feature (texture, shape, object detection)，摻入構圖元素 (井字構圖，三分法之類的)