The Effective Evaluation of Emotions in the Visual Emotion Images Using Convolutional Neural Networks

This paper develops a model for recognizing visual image emotions. The integration of contrastive-center loss optimization is proposed in this paper. This effectively improves the recognition of emotions when training a convolutional neural network against the baseline. The proposed contrastive-center loss function optimizes deep neural networks by enhancing feature discriminability. This loss function includes two key components: intra-class compactness and inter-class separability. We have suggested controlling the impact of the inter-class separability on the loss function. Moreover, we suggest combining cross-entropy and contrastive-center loss to calculate the total loss. In addition, we have proposed to apply the dimensionality reduction (visualization) for interactive evaluation of how the objects (images) in the test set are arranged and how this arrangement, as well as the classification as a whole, can be improved by choosing the best combination of the strength of contrastive-center loss impact on the total loss. The efficiency of the developed model improvements is examined on three datasets: WEBEmo, FI-8, and EmoSet-118K. Our research allows us to improve the performance of visual emotion classification: for the WEBEmo dataset by 1.6%, the FI-8 dataset by 2.2%, and for the EmoSet-118K dataset by 2.52% higher accuracies than the baseline case.