instructor : Xiaoming Jin
This project is on the machine intelligence(code here). With the development on multimedia and internet, image retrival based on content is promising area. It is comprised of two parts, the extraction of image feature and the artificial neural network.
As for the image feature extraction, we would like explore from color, texture, shape and so on. We implied four algorithms, HSV color histogram, SIFT, perceptual hashing and Wavelet perspectively. In terms of HSV color histogram, we obtained the best result when dimensions of HSV was equal to 2:1:1. We can see the accuracy on several portion of dimesions.
| H | S | V | Accuracy |
|---|---|---|---|
| 8 | 8 | 0 | 35.2% |
| 6 | 6 | 6 | 36.8% |
| 12 | 6 | 6 | 38.1% |
We can see the accuracy of each method as well. As the accuracy in overall was larger than any single one, we obtained that the artificial neural network had learned more higher feature by itself.
| Surf | Wavelet | pHash | HSV |
|---|---|---|---|
| 34.5% | 23.9% | ~20% | >= 40% |
The artificial neural network was trained by 4000 images and tested on the rest 1000 images. The accuracy was around 70%.
In overall, we built the system on image retrival based on the artificial neural network. Numerous methods on image feature extranction were implementedW.
Group member: Chengpeng Wang, Yanzhe Yang, Jiayu Zhang