COMPARATIVE ANALYSIS OF VISION TRANSFORMER AND EFFICIENTNET FOR FOOD RECOGNITION WITH BMI-AWARE DIETARY MONITORING

Abstract

Dietary-related diseases and obesity impose a substantial global health burden, yet accurate and continuous dietary monitoring remains largely manual and inaccessible. This paper presents a comparative analysis of two lightweight deep learning architectures—Vision Transformer (ViT-tiny) and EfficientNet-B0—for automated food image classification and personalized calorie estimation. Both models are trained under identical conditions on 65 food categories derived from the Food101 dataset, comprising 48,750 training and 16,250 validation images pre-processed to 128×128 resolution. Phase I (ViT-tiny) achieves 76.17% top-1 and 93.49% top-5 validation accuracy with 5.5M parameters, while Phase II (EfficientNet-B0) achieves 77.56% top-1 and 93.91% top-5 accuracy with 4.1M parameters. Phase III benchmarks GPU inference throughput on an NVIDIA T4, with ViT-tiny achieving 2,918 img/s at 5.06 ms single-image latency and EfficientNet-B0 achieving 2,480 img/s at 8.65 ms. A personalized nutritional engine integrates the Mifflin–St Jeor basal metabolic rate equation with WHO BMI-category-adjusted gram-level serving recommendations, delivering user-specific caloric guidance across four weight categories. Experimental results demonstrate that EfficientNet-B0 achieves superior classification accuracy with fewer parameters, while ViT-tiny offers a 1.7× latency advantage suited for real-time edge deployment, establishing complementary efficiency-accuracy trade-offs for practical food recognition and dietary monitoring applications.