Internship Daily

After we tested a lot of images for classification, we classified all photos and adjusted a ratio. We specified 5 class which are Close, Normal-Close, Normal, Normal-Wide and Wide Eyes. Some results shown below .

In this stage, we are trying to classify different eyes shapes. The most important thing in this classification is distance between of two eyes. For example, if the space between two eyes is smaller than the width of one eye, this is close set eye or if vice versa, this is wide set eye. For normal set eyes, we specified a tolerance value and tested over approx. a hundred images. Then, we classified these images which called wide-close- normal eyes. In next stage, we will try in our program each image and make a histogram to get a correct result for classification. For instance this is close set eyes. Normal Set Eyes Wide Set Eyes Close Set Eyes

To take a correct images are very important to get  good results.We have tested a lot of photos. We couldn't get correct result some photos to detect types of eyes. So, most important thing, it must be straight face pose when you take a photo. We are trying to avoid to take a picture with wrong angle.

There are lots of multiple kind of eyes on people. We can`t understand which kind of eyes at first view. In this samples, we will show you how to detect type of eyes. Even if there are a lot of variable eyes, firstly we will try to simple classification. We will consider only 3 types of eyes which are CAT EYE, OVAL EYE, DROOPY EYE. We shouldn`t  try to low definition images to get accurate results. 1. CAT EYE 2. ROUND/OVAL EYE 3. DROOPY EYE

If you want to measure any facial landmarks correctly, you should rotate facial image which is not straight. The person who taken an photo may not stand straight. To avoid this situation, firstly we drew two lines throught eyes and nose. We rotate the image according to eye line. After that, we found the angle between two lines. This angle should be about 90 degrees for correct and straight face photos. If this angle is between 85-95 degrees, this photo can use to measuring and we rotate the image other but if it is not, the program gives an warning which is "This photo was rejected. Please take a photo again". So this method will give much more correct. Taken image Rotated image As you can see, the program gave a warning because the photo taken has an incorrect angle. Taken photo Rotated and original photo The program didn't give a warning. This photo is suitable.

We have tried to some codes and got a good result. The hardest thing for us was to compile Dlib library. We spent a lot of time but its worth :) My result shown as follow: LEFT EYEBROW AND EYES BLACK SKİN