Comparative Analysis on Machine Learning Models for Image Captions Generation
Anirudh B. Mitta, Ajay Hegde, Beluvigi Shreegagana, Mohammad Rooh Ullah, Asha Rani K P, Saahithya Gowrishankar · 2023
This research paper conducts a comparative analysis on five distinct machine learning models employed for generating signature images using shared data and multiple evaluation metrics. The primary aim is to evaluate the effectiveness of these models and identify their strengths and weaknesses. The assessment of the generated image quality is performed through various criteria including BLEU, ROUGE, METEOR, and CIDEr. The results reveal that each model exhibits varying performance across different metrics, with no single model outperforming the others. However, certain models demonstrate higher scores on specific metrics, indicating their potential suitability for particular applications. The study emphasizes the importance of employing multiple evaluation metrics to accurately assess the quality of generated labels. It underscores the effectiveness of different machine learning models in this task. For instance, the CIDEr metric proves valuable when multiple tags are required for an image, as it considers tag diversity. Conversely, metrics like BLEU and ROUGE are more suitable for applications that prioritize clear and precise text by evaluating the clarity and overlap between generated and reference labels. This research underscores the significance of considering multiple evaluation methods to ensure the accurate assessment of label quality. The findings have practical implications for the development of efficient sign-making systems, providing guidance on the selection of appropriate models and measurements tailored to specific applications.