This article introduces an algorithm for facial expression recognition (FER) using deep Siamese Neural Networks (SNNs) that preserve the local structure of images in the embedding similarity space. We designed the network to reveal the input pairs similarity by comparing features through a designed metric. Furthermore, we developed a novel image pairing (i.e., positive and negative pairs) strategy technique to train our Siamese model. Our Siamese model comprises of a verification framework and an identification framework to learn a joint embedding space. The verification path reduces the intra-class variations by minimizing the distance between the extracted features from the same class, while the identification path increases the inter-class variations by maximizing the distance between the features extracted from different classes. We apply transfer learning to only use the identification model for facial expression classification. We evaluated our algorithm using AffectNet, FER2013, and Compound Facial Expressions of Emotion (CFEE) datasets, where better results are achieved compared to other deep learning-based approaches.

This article introducesBReG-NeXt, a residual-based network architecture using a function wtih bounded derivative instead of a simple shortcut path (a.k.a.identity mapping) in the residual units for automatic recognition of facial expressions based on the categorical and dimensional models of affect. Compared to ResNet, our proposed adaptive complex mapping results in a shallower network with less numbers of training parameters and floating point operations per second (FLOPs). Adding trainable parameters to the bypass function further improves fitting and training the network and hence recognizing subtle facial expressions such as contempt with a higher accuracy. We conducted comprehensive experiments on the categorical and dimensional models of affect on the challenging in-the-wild databases of AffectNet, FER2013, and Affect-in-Wild. Our experimental results show that our adaptive complex mapping approach outperforms the original ResNet consisting of a simple identity mapping as well as other state-of-the-art methods for Facial Expression Recognition (FER). Various metrics are reported in both affect models to provide a comprehensive evaluation of our method. In the categorical model, BReG-NeXt-50 with only 3.1M training parameters and 15 MFLOPs, achieves 68.50 and 71.53 percent accuracy on AffectNet and FER2013 databases, respectively. In the dimensional model, BReG-NeXt achieves 0.2577 and 0.2882 RMSE value on AffectNet and Affect-in-Wild databases, respectively.

Residual-based neural networks have shown remarkable results in various visual recognition tasks including Facial Expression Recognition (FER). Despite the tremendous efforts have been made to improve the performance of FER systems using DNNs, existing methods are not generalizable enough for practical applications. This paper introduces Bounded Residual Gradient Networks (BReG-Net) for facial expression recognition, in which the shortcut connection between the input and the output of the ResNet module is replaced with a differentiable function with a bounded gradient. This configuration prevents the network from facing the vanishing or exploding gradient problem. We show that utilizing such non-linear units will result in shallower networks with better performance. Further, by using a weighted loss function which gives a higher priority to less represented categories, we can achieve an overall better recognition rate. The results of our experiments show that BReG-Nets outperform state-of-the-art methods on three publicly available facial databases in the wild, on both the categorical and dimensional models of affect.

This paper presents a novel algorithm for an end-to-end facial expression recognition(FER) system based on deep Siamese neural networks equipped with a supervised loss function. Our method learns a powerful FER system by dynamically modulating verification signal over identification/classification signal. The identification signal increases the inter-class variations by maximizing the distance between the features for different classes, while the verification signal reduces the intra-class variations by minimizing the distance between features for the same class. We have evaluated our method on the AffectNet dataset [10] and achieved promising results compared to other deep learning models.

In this paper, we introduce a digital implementation of the 3-D shearlet transform and illustrate its application to problems of video denoising and enhancement. The shearlet representation is a multiscale pyramid of well-localized waveforms defined at various locations and orientations, which was introduced to overcome the limitations of traditional multiscale systems in dealing with multidimensional data. While the shearlet approach shares the general philosophy of curvelets and surfacelets, it is based on a very different mathematical framework, which is derived from the theory of affine systems and uses shearing matrices rather than rotations. This allows a natural transition from the continuous setting to the digital setting and a more flexible mathematical structure. The 3-D digital shearlet transform algorithm presented in this paper consists in a cascade of a multiscale decomposition and a directional filtering stage. The filters employed in this decomposition are implemented as finite-length filters, and this ensures that the transform is local and numerically efficient. To illustrate its performance, the 3-D discrete shearlet transform is applied to problems of video denoising and enhancement, and compared against other state-of-the-art multiscale techniques, including curvelets and surfacelets.