Image captioning is an important research area in artificial intelligence that integrates computer vision and natural language processing (NLP) to automatically generate descriptive textual interpretations of images. Conventional image captioning systems typically employ Convolutional Neural Networks (CNNs) for visual feature extraction and Long Short-Term Memory (LSTM) networks for generating sequential text descriptions. Although effective, CNN-based approaches may have limitations in capturing global contextual relationships within images.

This research introduces an improved image captioning framework that utilizes Vision Transformers (ViTs) as the feature extraction backbone instead of traditional CNN architectures. By leveraging self-attention mechanisms, Vision Transformers can effectively model long-range dependencies and capture comprehensive contextual information from visual data. The extracted image representations are subsequently provided to an LSTM network, which generates coherent and meaningful captions in a sequential manner.

The proposed model is evaluated using widely accepted image captioning performance metrics, including BLEU and METEOR scores. Experimental findings indicate that the Vision Transformer-based approach produces more accurate, descriptive, and context-aware captions compared to conventional CNN-LSTM models. The enhanced caption generation capability of the proposed framework makes it suitable for various real-world applications, including assistive technologies for visually impaired individuals, automated image annotation, content management systems, and intelligent multimedia retrieval.

Keywords— Image Captioning, Vision Transformer (ViT), Long Short-Term Memory (LSTM), Natural Language Processing (NLP), Computer Vision, Text Generation, BLEU Score