PROJECT TITLE :

Occlusion Aware Facial Expression Recognition Using CNN With Attention Mechanism

ABSTRACT:

It is difficult to recognise facial expressions in the outdoors because of a variety of unrestricted circumstances. Existing face classification systems are nearly flawless for evaluating limited frontal faces, but they fall short in analysing partially occluded faces that are more common in the real world. Acknowledging the occlusion regions of the face, we present a convolution neutral network (CNN) with an attention mechanism (ACNN) that focuses attention on the most discriminative areas of the face. ACNN is a comprehensive framework for learning. It mixes the representations from different parts of the face that are relevant (ROIs). An adaptive weight is calculated for each representation based on the region's unobstructedness and relevance via a proposed gate unit. With respect to different regions of interest (ROIs), we offer two versions of ACNN (pACNN and GLOBALD): (gACNN). Local facial patches are all that pACNN is interested in focusing on. Local representations at the patch level are integrated into a global representation at the image level using gACNN's help. All of the ACNNs are tested on both real and synthetic facial occlusion datasets, such as the two largest in-the-wild facial expression datasets, as well as their modified versions with synthetic facial occlusions, to see how well they perform. ACNNs enhance recognition accuracy on both non-occluded and occluded faces, according to experimental results. When compared to CNNs without Gate Units, ACNNs are capable of transferring attention away from occluded patches to other related but unimpeded ones, according to visualisation results. Using the cross-dataset evaluation technique, ACNNs outperform other state-of-the-art approaches on numerous commonly used in-lab facial expression datasets.