Image captioning models have significantly improved by integrating various CNN-based feature extractors, demonstrating their efficacy in capturing visual semantics. Building on this progress, our study investigates the impact of using different CNN architectures for feature extraction to represent crucial aspects of visual content. Specifically, we explore two prominent model families: Inception and EfficientNet. Within these families, we evaluate the performance of Inception V3, Inception V4, and EfficientNet B0, B4, and B5. Our studies show that Inception V3 surpasses all other models, offering higher accuracy. We subsequently analyze the performance and accuracy of the remaining models to determine their relative strengths. In addition, we use Transformer models to analyze the extracted data, creating image captions consecutively and turning visual representations into meaningful textual descriptions. To improve contextual relevance, we include an Attention mechanism that focuses on key areas of the image during caption generation. The proposed approach significantly improves accuracy over the prior version, bringing Inception V3’s performance from 42% to a higher benchmark. EfficientNet B0 also shows improved accuracy. This technique improves classification accuracy and decreases computation time, demonstrating its effectiveness for efficient image captioning.