ER7ST-algorithm for extracting facial expressions

Abstract EN

his paper, proposes a n^ algorithm for recognition offacial expressions, called ER7ST.

Studied expressions are anger, disgust, fear, happiness natural, sadness and surprise.

Proposed method is based on extraction of the essential objects, then finding of characteristic points positions of each object.

Detected points slant and its average are calculated on the basis of fixed points.

Mug database is considered as data source for training and testing.

We collect set of images.

ER7ST is designed to define the work area of face that contains characteristic expressions, which its centre is face centre and its dimensions are 8^16 supposed squares.

ER7ST algorithm discovers the essential objects depending on the coordinate of minimum and maximum points of each object in defined area, considering the length of object is larger than the major-axis of formed ellipse on studied object.

Object gradient is ranked between [-60, +60] degrees.

Our algorithm detects ConvexHull points upon detected objects and then its slant is calculated.

Slant vectors are formed; some calculations are done to be a good input to network.

Net contains input, three hidden layers and output layer.

After training on set of faces and testing on new data, the recognition rate was promising.

Algorithm can be maintained with different types of images and it did not need to scale.

Finally, recognition rate is ranked between 60 % and 95 %, experiments have shown that method was efficient and results were very encouraging in this field, especially network can be trained on new situations of expressions.