Elena Battini Sönmez

Convolutional neural networks with balanced batches for facial expressions recognition

This paper considers the issue of fully automatic emotion classification on 2D faces. In spite of the great effort done in recent years, traditional machine learning approaches based on hand-crafted feature extraction followed by the classification stage failed to develop a real-time automatic facial expression recognition system. The proposed architecture uses Convolutional Neural Networks (CNN), which are built as a collection of interconnected processing elements to simulate the brain of human beings. The basic idea of CNNs is to learn a hierarchical representation of the input data, which results in a better classification performance. In this work we present a block-based CNN algorithm, which uses noise, as data augmentation technique, and builds batches with a balanced number of samples per class. The proposed architecture is a very simple yet powerful CNN, which can yield state-of-the-art accuracy on the very competitive benchmark algorithm of the Extended Cohn Kanade database.

Classification with emotional faces via a robust sparse classifier

We consider the problem of emotion recognition in faces as well as subject identification in the presence of emotional facial expressions. We propose alternative solutions for this identification and recognition problems using the idea of sparsity, in terms of Sparse Representation based Classifier (SRC) paradigm. In both cases, the problem is formulated as finding the most parsimonious set of representatives from a training set, which will best reconstruct the test image. For emotion classification, we considered the six fundamental states and the SRC performance was compared with that of the Active Appearance Model (AAM) algorithm [1]. For face recognition displaying various emotions, in order to test the robustness of SRC, we considered gallery faces of subjects having one or more expression variety while the probe faces had a different expression. We experimented with both the whole faces or faces observed with multiple blocks. The SRC algorithm, while not demanding any training, performed surprisingly well in both emotion identification across subjects and subject identification across emotions.

Face classification via sparse approximations

2D face classification problem is very difficult because of the illumination, pose, expression and occlusion factor. The sparse Representation based classification (SRC) provides a significant amount of robustness against to illumination changes and noise. In this study, sparse approximation is used on face recognition and applied to “Extended Yale Face B” database. Results of sparse approximation classification are compared with Fisher Linear Discriminator.

Emotion recognition in the wild using deep neural networks and Bayesian classifiers

Group emotion recognition in the wild is a challenging problem, due to the unstructured environments in which everyday life pictures are taken. Some of the obstacles for an effective classification are occlusions, variable lighting conditions, and image quality. In this work we present a solution based on a novel combination of deep neural networks and Bayesian classifiers. The neural network works on a bottom-up approach, analyzing emotions expressed by isolated faces. The Bayesian classifier estimates a global emotion integrating top-down features obtained through a scene descriptor. In order to validate the system we tested the framework on the dataset released for the Emotion Recognition in the Wild Challenge 2017. Our method achieved an accuracy of 64.68% on the test set, significantly outperforming the 53.62% competition baseline.

A Model for an Emotional Respondent Robot

The aim of this study is to design an emotional regulation model based on facial expressions. It is argued that emotions serve a critical function in intelligent behavior and some researchers posed the questions of whether a robot could be intelligent without emotions. As a result, emotion recognition and adequate reaction are essential requirements for enhancing the quality of human robot interaction. This study proposes a computational model of emotion capable of clustering the perceived facial expression, and using cognitive reappraisal to switch its internal state so as to give a human-like reaction over the time. That is, the agent learns the person’s facial expression by using Self Organizing Map, and gives it a meaning by mapping the perceived expression into its internal state diagram. As a result, the presented model implements empathy with the aim to enhance human-robot communication.

A survey of product recognition in shelf images

Nowadays, merchandising is one of the significant method which allows to increase the sales. Therefore, activities such as monitoring the number of products on the shelves, completing the missing products and matching the planogram continuously have become important. An autonomous system is needed to automate operations such as product or brand recognition, stock tracking and planogram matching. In the literature, it is seen that many studies have been carried out in order to address this issue. This survey classifies and compares all existing works with the aim to guide researchers working on merchandising.

Critical parameters of the sparse representation-based classifier

In recent years, the growing attention in the study of the compressive sensing (CS) theory suggested a novel classification algorithm called sparse representation-based classifier (SRC), which obtained promising results by casting classification as a sparse representation problem. Whereas SRC has been applied to different fields of applications and several variations of it have been proposed, less attention has been given to its critical parameters, that is, measurements correlated to its performance. This work underlines the differences between CS and SRC, it gives a mathematical definition of five measurements possible correlated to the performance of SRC and identifies three of them as critical parameters. The knowledge of the critical parameters is necessary to fuse multiple scores of SRC classifiers allowing for classification. The authors addressed the problem of two-dimensional face classification: using the Extended Yale B dataset to monitor the critical parameters and the Extended Cohn-Kanade database to test the robustness of SRC with emotional faces. Finally, the authors increased the initial performance of the holistic SRC with a block-based SRC, which uses one critical parameter for automatic selection of the most successful blocks.

Face classification via sparse approximation

We address the problem of 2D face classification under adverse conditions. Faces are difficult to recognize since they are highly variable due to such factors as illumination, expression, pose, occlusion and resolution. We investigate the potential of a method where the face recognition problem is cast as a sparse approximation. The sparse approximation provides a significant amount of robustness beneficial in mitigating various adverse effects. The study is conducted experimentally using the Extended Yale Face B database and the results are compared against the Fisher classifier benchmark.

Biometric face recognition using randomfaces

This paper investigates the use of the Compressive Sensing (CS) technique to the classification issue. In this context, CS is used as a means to probe the nonlinear manifold on which faces under various illumination effects reside. The scheme of randomly sampled faces (Randomfaces) with nearest neighbor classifier are compared with two classical feature extraction approaches, as Eigenfaces and Fisherfaces. It is shown that randomfaces outperform the eigenface approach in classifying faces under illumination disturbances and their performance approaches that of the Fisherfaces.