A Comparative Evaluation of Image Caption Synthesis Using Deep Neural Network

Image caption generation is a crucial challenge in deep learning and natural language processing, involving identifying the context of an image and providing appropriate captions. In this study, we aimed to evaluate and compare the performance of two different model architectures using pre-trained CNN models for image classification and sequential LSTM models for caption generation. Specifically, we used RestNet50 and inceptionV3 CNN models with word2Vec and GloVe word embeddings, respectively, to generate captions. We evaluated the models based on two criteria: calculating the BLEU score for each generated caption and comparing the BLEU score with the inceptionResNetV2 state-of-the-art model. Our results showed that the second model architecture with inceptionV3 and GloVe-based model outperformed the first model and closely followed the benchmark BLEU score of the state-of-the-art model. Therefore, our study provides evidence that the choice of pre-trained CNN model and word embedding technique can significantly impact the performance of image caption generation, with the proposed architecture offering an accurate and efficient solution.