DICKT—Deep Learning-Based Image Captioning using Keras and TensorFlow

Abstract

This study evaluates a caption generation model's performance using the BLEU Score metric. The model generates descriptions for images, compared to reference captions with single and dual references. Results show a high BLEU Score, suggesting human-like captions. However, BLEU primarily measures linguistic similarity and n-gram overlap, missing full human-generated caption richness. The findings reveal the model's potential to convey image essence in text, but highlight BLEU Score limitations. TensorFlow and Keras are used for model development, acknowledging their widespread use but also their limitations. The research offers insights into caption generation model capabilities and urges a broader perspective on caption quality beyond quantitative metrics. While higher BLEU Scores are generally preferred, a "good" score varies with dataset and context. The study emphasizes a need for a more comprehensive approach to assess the quality and creativity of machine-generated captions.