A Novel Multi-purpose Deep Architecture for Facial Attribute and Emotion Understanding

Facial expression estimation has for years been studied benefiting a wide array of application areas ranging from information retrieval and sentiment analysis to video surveillance and emotion analysis. Methods have been proposed to tackle the problem of facial attribute recognition using deep architectures yielding high accuracies, however less efforts exist to focus on the performance of these architectures. Here in this work, we make use of Squeeze-Net [6] for the first time in the literature to perform facial emotion recognition benchmarked on Celeb-A and AffectNet datasets. Here we extend Squeeze-Net by introducing a new 5 \(\times \) 5 convolution kernel after the last fully-connected layer offered by Squeeze-Net, merging the 1 \(\times \) 1 and 3 \(\times \) 3 outputs from the last fully-connected layers, to perform a more domain-specific feature extraction. We run extensive experiments using widely-used datasets; i.e. Celeb-A and AffectNet, using AlexNet and Squeeze-Net in addition to our proposed architecture. Our proposed architecture, an extension to Squeeze-Net, yields results inline with state of the art while offering a simple architecture involving less complexity compared to state of the art, reporting accuracies of 90.47% and 56.38% compared to 90.94% and 52.36%, in Attribute Prediction and Expression Prediction respectively.