AMFMER: A multimodal full transformer for unifying aesthetic assessment tasks

Computational aesthetics aims to simulate the human visual perception process via the computers to automatically evaluate aesthetic quality with automatic methods. This topic has been widely studied by numerous researchers. However, existing research mostly focuses on image content while disregarding high-level semantics in the related image comments. In addition, most major assessment methods are based on convolutional neural networks (CNNs) for learning the distinctive features, which lack representational power and modeling capabilities for multimodal assessment requirement. Furthermore, many transformer-based model approaches suffer from limited information flow between different parts of the assumed model, and many multimodal fusion methods are used to extract image features and text features, and cannot handle multi-modal information well. Inspired by the above questions, in this paper, A novel Multimodal Full transforMER (AMFMER) evaluation model without aesthetic style information is proposed, consisting of three components: visual stream, textual stream and multimodal fusion layer. Firstly, the visual stream exploits the improved Swin transformer to extract the distinctive layer features of the input image. Secondly, the textual stream is based on the robustly optimized bidirectional encoder representations from transformers (RoBERTa) text encoder to extract semantic information from the corresponding comments. Thirdly, the multimodal fusion layer fuses visual features, textual features and low-layer salient features in a cross-attention manner to extract the multimodal distinctive features. Experimental results show that the proposed AMFMER approach in this paper outperforms current mainstream methods in a unified aesthetic prediction task, especially in terms of the correlation between the objective model evaluation and subjective human evaluation.