Advances in Bangladeshi Cuisine Recognition: A Review of Deep Learning, Vision–Language Models, Fine-Tuning and Parameter-Efficient Adaptation

Shafiul Islam Khokon, Tonmoy Barua, Sajid Ibne Alam, Ishmam Ahmed Solaiman, Nahiyan Bin Noor

The automated analysis of food through computational methods has emerged as a significant field of research, driven by applications in health, gastronomy, and cultural preservation. A key task within this domain is food recognition, a challenging form of fine-grained visual classification (FGVC) complicated by subtle inter-class differences and high intra-class variance. This challenge is exacerbated by a well-documented geographic and cultural bias in the large-scale datasets used to train modern artificial intelligence (AI) models, which results in poor performance on non-Western cuisines. This review examines the intersection of these challenges in the context of classifying traditional Bangladeshi food. We survey the evolution of FGVC, from early part-based models to the current paradigm of vision transformers and vision–language models (VLMs), which are better suited to the non-rigid and variable nature of food. We then explore the critical role of parameter-efficient fine-tuning methods, such as low-rank adaptation and its variant, quantized low-rank adaptation, which make the adaptation of massive foundation models computationally feasible for niche domains. An analysis of existing food datasets reveals significant representational gaps, underscoring the necessity of developing and utilizing culturally specific datasets. By synthesizing these technical and socio-cultural threads, we argue that fine-tuning a VLM on a Bangladeshi food dataset using parameter-efficient fine-tuning is a methodologically sound and culturally significant approach. This work not only addresses a technical gap in computer vision but also contributes to the digital preservation of culinary heritage, laying the groundwork for future research in multimodal food analysis, data augmentation, and the development of equitable AI systems.