IMAGE CAPTION GENERATION: COMPARATIVE ANALYSIS AND DEEPLEARNINGBASED FRAMEWORK

Abstract

Image caption generation is an important research area in artificial intelligence that combines computer vision and natural language processing to automatically generate textual descriptions for images. The ability to produce meaningful captions has several real-world applications, including assistance for visually impaired individuals, multimedia management, image retrieval, surveillance systems, and human–computer interaction. This research work presents a deep learning-based image caption generation framework using a CNN–LSTM encoder–decoder architecture. A convolutional neural network is utilized for extracting important visual features from images, while a Long Short-Term Memory network is employed for sequential caption generation. The proposed model is trained using the Flickr8K dataset containing images paired with multiple human-annotated captions. Various preprocessing techniques such as image normalization, tokenization, vocabulary generation, and sequence padding are implemented to improve learning performance. The developed framework is further integrated with a Streamlit-based web application that enables dynamic image upload and automatic caption generation. Experimental results demonstrate that the proposed framework can generate contextually relevant and semantically meaningful captions for diverse image scenes. Future improvements may involve attention mechanisms, transformer-based architectures, and vision–language pre-trained models for enhanced caption generation performance.