DEEP LEARNING FOR PHOTO EMOTION RECOGNITION

Abstract [English]

Photographic emotion recognition has become an important field of research application in the interface of computer vision, affective computing, and deep learning, and has been applied in digital media analysis, human-computer interaction, mental health assessment, and content behavioral AI. In comparison to object or scene recognition, photo emotion recognition is the recognition of subjective affective reactions that visual stimuli trigger, which means that the task is an inherently difficult and situation-specific task. This paper introduces a deep learning-based emotion recognition model of the expressions of photographic images incorporating the psychological theories of emotions with the state-of-the-art convolutional neural network models. The framework of the proposed solution is also based on the known models of emotions, such as valence-arousal dimensions, discrete categories of emotions, which allow mapping visual patterns and affective semantics systematically. Hierarchy Visual features like color distributions, texture gradients, lighting and composition balance are represented by hierarchical feature extraction to achieve low level perceptual features of the visual image and high-level semantic features of the visual image. It uses a properly selected and annotated dataset of emotions, that are backed up by strong preprocessing and data augmentation techniques to increase generalization. The deep neural network applies convolutional learning of features and attention mechanism to highlight emotional regions of the image. Large-scale experiments are performed based on regularized training, validation, and testing conditions, and performance is measured against various baseline models in terms of accuracy, precision, recall, and F1-score measures.