| Summary: | 碩士 === 國立臺灣科技大學 === 資訊工程系 === 107 === Facial expression recognition (FER) is an important issue in the field of computer vision.There are methods that perform FER, but they have reduced performance if the source (training) and target (testing) dataset have a large discrepancy.Domain adaptation is usually performed to handle this problem, however, in practical applications, the target dataset is not always readily available, and the model needs to be adapted for each new target dataset.Our FER method uses data augmentation rather than domain adaptation to generate robust facial expression classifier networks.We performed data augmentation using Generative Adversarial Networks (GAN) to generate synthetic face images with different facial expressions defined by Actions Units (AU), which are anatomically-based movements of certain facial muscle groups (e.g cheek raise).We augmented a high variation dataset (e.g. contains a variety of head poses, illumination, perspective, subject ethnicity) to create a new dataset, with a large amount of datapoints, for training a robust network.To improve the classification performance of our network, we utilized non-local blocks to capture long-range spatial relationship, and island loss to decrease intra-class (same facial expression) variations and increase inter-class (different facial expressions) differences.Our network can be trained on a single dataset, and our experiments show that our network has state-of-the-art performance in facial expression classification on different datasets.
|
|---|
