In the realm of fine-grained sentiment analysis, target-oriented sentiment analysis (TSA) is an important task. The TSA task is to categorize the sentiment of a particular target word in a multimodal post or remark. But the text in multimodal data is short and comes with complicated images that might not be relevant. Also, the current neural network models in this realm are mostly focused on combining multimodal data and don’t look at how different themes affect the target’s emotional tendencies. Therefore,we propose an ICLB (Image Caption Latent Dirichlet Allocation Bert) model. The model includes an image caption module to generate sentences that describe images, augment the text corpus for the language model, and extract detailed information about entities from intricate images. Using topic-modal modules to improve focused communication of emotional messages. We evaluated our method on two standard datasets, Twitter-15 and Twitter-17. The results show that the model is better at target-oriented multimodal sentiment analysis than the most recent top-of-the-line models.