Dear Wang,

Actually this is the correct way to find out faces because if we go for to find out the real boundaries, it would be time consuming and it will take high resource overhead. So, it can be possible but not feasible. On the other hand the other methods give us the ROI that is actually important for any kind of recognition or localization processing.

Still if you want facial boundaries for any purpose then in my point of view convert images into grayscale then start your process from the chin area after that go anti-clock wise by normalising the boundaries by converting area into binary image (ONLY NEARBY AREA). It is an experimental thing I have no personal experience of it.

Hello

I think there are several different models to do that:

1. Classification pixels (color) to face colors. Then make some morphology operations.

2. Modern face recognition extract feature points. This points are uses to make graph model of face, 3D model of face and etc. Try to find open API of face recognition and try extract latent variables....

This is actually a very hard task, since the facial boundaries are often not discernible (especially the chin-neck region). Then there is the problem what you would define as the true boundary: Should the hair be included? What about the ears? The chin?

If you want to exclude the background (e.g. for recognition), you can just shrink the bounding box so that it only contains the "inner" face (eyes, nose, mouth). If you want to know the shape of the face, have a look at facial feature detection, e.g. Cootes' active appearance models and related algorithms (one that is very fast and reasonably easy to understand is "Face alignment by Explicit Shape Regression" by Cao et al.)

I think this is a very hard task, and it's highly dependent on the face images you're working with and your purpose.

To use segmentation methods based on frequency space analysis is tricky, because the contrast among different persons varies. For instance, if you have a brunette person (dark hair, light skin), this method can be accurate, but if you have a blond person or a black person (light hair/light skin or dark hair/dark skin), the contrast is very low, and frequency based methods won't work.

Anther approach, used by many authors, is to model skin color, and use a segmentation method based on this. This also have problems mainly due to varied lighting conditions and shadows (that modify the actual skin color, and makes it less homogeneous).

I would recommend a combination of both approaches. I have done this myself, in order to segment the silhouette of a person. I segmented the hair first, and the skin area after, using a combination of the aforementioned approaches. Unfortunately, this paper is written in spanish, and I don't have an english version. Anyways, it is available here:

https://www.researchgate.net/publication/257943438_EVALUACIN_DE_FONDO_UNIFORME_EN_IMGENES_DE_ROSTROS?fulltextDialog=true

Another attempt that may help would be to use ASM/AAM methods. By detecting facial landmarks along the face shape, you can actually obtain the boundary you're looking for, and it works even in the presence of occluded regions (i.e., hair around that boundary). The problem is that it takes a great deal to create training data, although there are datasets that provides training information already collected.

I hope this enlightens a little the path for you...

Conference Paper EVALUACIÓN DE FONDO UNIFORME EN IMÁGENES DE ROSTROS