If your Blender FD and you happen to run across this i have a node setup that i created after watching about 5-6 videos on keying theory that not only pulls amazing keys but it works on objects that are blurred with the greenscreen and corrects the color spill and the blurred spill.
Download this file. Delete the .blend extension (it is a zip file) and unzip.
Open the blend file. You see a switch that is ON selecting your method of green screen keying. Select the Viewer1 and you see
your result. Select Viewer2 and you see the colors on the head and lamp of the original are different than in your result.
Now turn OFF the switch and you see the colors are correct.
Also in your method the hair of the guy is lost.
Can you review the nodes and see if something is lacking or I build your method correctly?
@Bao2
It definitely works, quite nicely too I might add. You just have to play with the Blur factors and the Color ramps like he said. Thanks for making that .blend to save me some time. 
@einstein**-7
Thanks for sharing this! It works quite nicely. The Key extracting group that is. The dedicated color spill removal node still works best for spill removal, I think.
Firstly, nice test file! Your setup was perfect just needed to adjust the ramps and blur amounts had to drop the blur due to res and hair size. Below are the links
my best matt and despill
blender best
my best matt and other despill
tweaked node setup
As you can see my matt has a little more hair than the blender one and i’d have to agree that the blender despill is overall better. Mine was created specifically for fast objects that blured with the backkey and worked well. Didn’t do it here but you can also put a blur on the matt input - before any of the nodes- and get even better results because the grain is eliminated before pulling the matt.
One thing to remember is that the jpg you put in your test file is essentially a 4:4:4 frame-- how to tell? Put it through a converter that splits in into the ycbcr space pop viewers on each channel and notice that the resolution of each is equal. If that was a standard video frame from say, a sony handycam or t2i it would be a 4:2:1 or 4:2:0 meaning that the cb would be half the resolution of the y and cr half of the cb. Trying to pull a key from that with blender’s keyers would be a disaster; i’ve tried.