Advanced digital image processing and feature extraction using deep CNN architectures for automated image captioning

Image caption generation is an emerging research area that integrates Computer Vision (CV) and Natural Language Processing (NLP) to enable machines to understand visual content and generate meaningful textual descriptions. This technology has significant applications in assistive systems for visually impaired individuals, autonomous navigation, image retrieval, and intelligent multimedia analysis. The primary objective of image captioning is not only to recognize objects within an image but also to understand their relationships, activities, and surrounding context to produce human-like descriptions. This study investigates the impact of different deep learning-based feature extraction architectures on automated image caption generation. A CNN–LSTM Encoder–Decoder framework is implemented, where Convolutional Neural Networks (CNNs) are used as image encoders to extract high-level visual features, while Long Short-Term Memory (LSTM) networks are employed as decoders to generate sequential textual captions. Five widely used CNN architecturesVGG16, ResNet50, InceptionV3, Xception, and DenseNet201, are comparatively evaluated using the Flickr8k dataset. The performance of these models is assessed using BLEU evaluation metrics to identify the most effective encoder architecture for caption generation. The experimental analysis provides insights into how different feature extraction strategies influence caption accuracy and demonstrates the effectiveness of deep learning approaches for bridging the semantic gap between visual information and natural language descriptions. Keywords: Image Captioning; Deep Learning; Encoder–Decoder Architecture; Convolutional Neural Networks; Long Short-Term Memory Networks; Feature Extraction; BLEU Score.