Despite the diversity of methods developed in recent years, the implementation of an efficient system for automatic recognition of facial emotions remains a technological challenge that has not been fully resolved. Many problems have not yet been resolved. The occlusion problem remains a challenge today for the research community, certain features characterizing several different emotions may seem similar, etc. High performing and precise techniques are therefore necessary to perfectly distinguish between two different emotions, even though they might be difficult to distinguish. The objective of this work is the development of an automatic method for recognizing basic facial emotions (joy, anger, sadness, disgust, surprise, fear and neutral) in video streams. The method of deep learning, known for its great performance in image classification, becomes essential. In order to be able to benefit from several feature maps at the same time, we propose to use two techniques: bilinear pooling (B-CNN), and Fusion Feature Net (F-CNN). This technique is more efficient and more precise than conventional techniques, whether based on deep learning or not.