Abstract
Bridging the gap between visual data and human language has been increasingly looked at through the task of automatically generating descriptive captions for images. This places the work within important scopes of accessibility, multimedia search, and human-computer interaction. For this work, we propose a hybrid deep learning model that fuses high-level scene context with localized object information for quality captions. Global image features are obtained through an Xception network, while You Only Look Once, version 8 (YOLOv8) is used to derive object-specific fine details. These visual features are merged and passed to a Bahdanau attention mechanism, which feeds an LSTM decoder to generate context-aware captions. The proposed method was tested on the Flickr8k dataset using BLEU and METEOR metrics; it showed promising improvements over traditional single-stream approaches. Results speak well of the model’s ability to deliver better interpretability and accuracy in image captioning.