The automatic recognition of human emotions is used to support several
computing paradigms, like afective, positive and pervasive computing.
Histograms of oriented gradients (HOG) have been successfully used with such a
purpose, by processing facial images. However, the results of using HOG vary
depending on the position of the facial components in the image used as input.
This paper presents an extension to the HOG method, which was named
Landmark-based Histograms of Oriented Gradients (LaHOG), that not only
calculates HOG blocks in the whole face, but also in specific positions around
selected facial landmarks. In this sense, the new method is more robust than
its predecessor. In order to evaluate the capabilities and limitations of this
proposal, we used it to recognize emotions in face images from the FACES
database. In such a process we used two classification strategies: support
vector machines and logistic regression. The results show that the extended
method significantly surpasses the performance of HOG in the tested database.