A dynamic features method for retrieval and emotional polarity of digital media images
Xin Xin, Muddassira Arshad · PeerJ Computer Science · 2025
On digital media platforms, too many images are produced. To process these images efficiently, we propose an image retrieval and sentiment polarity analysis method based on dynamic features. First, considering that a single visual modality contains too few semantic features, we introduce an image captioning method that enhances semantic information, thereby adding additional modality information to the model. Then, based on the above-described features, we propose an image retrieval and sentiment polarity analysis method using multimodal dynamic features, which enables the retrieval of results and the analysis of sentiment polarity for the image. Experiments demonstrate that our method achieves Acc@1 = 0.951 and mean average precision (mAP) = 0.907, outperforming comparable baselines by up to 3–5% in retrieval accuracy, while keeping the average processing time at only 112.4 ms per image. These results confirm that the proposed framework delivers both high accuracy and real-time efficiency, significantly advancing the state of the art in multimodal image retrieval and sentiment polarity analysis