VGG Models in Image Captioning: Which Architecture Delivers Better Descriptions?

The recent advancements and a flurry of deep learning architectures in the fields of computer vision and natural language processing have greatly benefited the task of creating natural language descriptions for images in recent years. These advancements have proved to be the turning point with popular practical applications in artificial intelligence, as well as for driving cross-disciplinary developments. However, not many studies have investigated the performance of different convolutional and LSTM architectures for image captioning, motivating our study to bridge this gap. This study compares the performance of VGG16 and VGG19 models on the Flickr8k dataset. Image features are extracted using pretrained VGG16 and VGG19 models, textual input is tokenized, and a unique Encoder-Decoder architecture is built. We determine the effect of training length on captioning proficiency by conducting comprehensive tests with varied epochs. The BLEU ratings reveal detailed insights into the models' language generating capabilities, demonstrating comparable performance across VGG16 and VGG19. Our work fills this gap by assessing the efficiency of several neural designs on a given dataset. The findings highlight the complexities of model convergence, and the need for different evaluation measures and defined benchmarks in the picture captioning sector. We determine the relationship between training epochs and model performance in this study, and provides useful observations for researchers and practitioners. We systematically examine the design of various convolutional neural networks, and provide a basis for future research in the ever-advancing world of image captioning.