Published Paper Details:

Online advertisements are one of the most common ways brands connect with people, combining images, text, and design to capture attention, measuring audience reaction but also for predicting how well an ad might perform. Traditional methods often focus only on sentiment or single aspects of ads, which leaves out many important factors like call-to-action text, visual appeal, or the likelihood of engagement. This paper presents a multimodal framework that brings together both visual and textual information from advertisements. This method uses a ResNet18 network to learn visual features from ad images and a DistilBERT model to capture the meaning of the text extracted from them. These features are merged through a fusion block and then passed into multiple task-specific modules. The framework can identify ad themes, emotions, and trustworthiness, detect objects and visual styles, and predict outcomes such as click-through rate (CTR) and audience engagement by using ADS-DS-1M dataset.