HYBRID DUAL-STREAM DEEP LEARNING FRAMEWORK FOR CONTEXT-BASED SENTIMENT ANALYSIS USING AGE AND BODY FEATURE EXTRACTION WITH NLP-DRIVEN IMAGE DESCRIPTION GENERATION

Abstract

Sentiment analysis has evolved into a crucial component of affective computing, enabling systems to infer emotional states from visual and textual data. Traditional sentiment analysis frameworks predominantly relied on text-based approaches, which inherently failed to capture the rich affective cues embedded in visual content such as human facial expressions, scene semantics, and object-level context. The rise of multi-modal deep learning architectures has bridged this gap, enabling joint understanding of vision and language. However, existing methods still suffer from several significant challenges including inadequate feature representation, inability to model contextual sentiment holistically, and limited scalability in generating emotionally coherent natural language descriptions for images. This paper presents a novel Hybrid Deep Learning Framework for Context-Based Sentiment Analysis that fuses age-discriminative and body-feature-aware convolutional pathways with a Natural Language Processing (NLP)-based image description generation module. The proposed architecture employs dual-stream Convolutional Neural Networks (CNNs) to extract age-related facial features and holistic body posture cues simultaneously, which are subsequently fused through an attention-guided mechanism to derive a unified affective feature representation. A transformer-based language generation module then produces contextually rich descriptions of the analysed scenes, enabling dynamic sentiment articulation at both the object and scene levels. The proposed framework is evaluated on two benchmark datasets: the Microsoft Common Objects in Context (MS-COCO) dataset and the FER2013 facial expression benchmark. Experimental results demonstrate that the hybrid model achieves a BLEU-4 score of 38.6, a METEOR score of 29.4, a CIDEr score of 121.8, and a sentiment classification accuracy of 94.7%, significantly outperforming state-of-the-art methods including BLIP, OFA, and Oscar. These results validate the efficacy of the proposed dual-stream feature extraction strategy and the NLP language generation pipeline in producing sentiment-aware image descriptions that are semantically coherent and emotionally expressive. The proposed framework holds promising implications for applications in human-computer interaction, assistive technologies, and intelligent surveillance systems.