Automatically captioning images using deep learning, datasets and estimation parameters: A Review

Automatic image captioning bridges natural language processing and computer vision by generating textual descriptions of visual content. This survey critically examines the evolution from early template-based methods to advanced deep learning architectures, emphasising transformer and encoder-decoder frameworks. We compare benchmark datasets such as MSCOCO, Flickr30k, Flickr8k, and PASCAL 1k, analysing their scale, diversity, and domain coverage. Quantitative evaluation of models is discussed through key metrics—including CIDEr, ROUGE, METEOR, and BLEU—highlighting their correlations with human judgment and limitations in contextual understanding. Finally, we identify emerging trends, such as multimodal pretraining and reinforcement learning-based optimization, that promise to improve caption fluency and semantic alignment. This work provides a data-driven overview of the field and outlines concrete research directions for enhancing spontaneous image captioning performance.