Integrating Color Perception in Multimodal Sentiment Analysis: A Hybrid Model Using DeBERTa and DenseNet

With the surge of multimedia content, multimodal sentiment analysis has emerged as a significant research direction in the field of artificial intelligence. However, existing models often prioritize high-level semantic features extracted by Convolutional Neural Networks (CNNs) when processing the visual modality, while paying insufficient attention to low-level visual features, such as color and texture, which play a crucial role in abstract or artistic images. To address this limitation, this paper presents a multimodal sentiment analysis model that integrates low-level color perception with high-level semantic representation. The model first utilizes pretrained DeBERTa and DenseNet-201 networks to extract deep semantic features from text and images, respectively. The innovative aspect of this model is its explicit computation of the average brightness of the image as a key low-level color feature, which is then concatenated with the high-level semantic features to form a more comprehensive image representation. Finally, the model optimizes the features using the Convolutional Block Attention Module (CBAM) and inputs them into the OTE multimodal fusion framework for sentiment polarity classification. In a series of comparative experiments and ablation studies, the proposed method demonstrates significant advantages in both accuracy and minority class F1 score, achieving a relative accuracy increase of 0.102 compared to the baseline. The experimental results indicate that integrating low-level color features effectively enhances the model's ability to perceive sentiment in diverse visual content, providing an effective and interpretable improvement pathway for multimodal sentiment analysis.