Balancing Accuracy and Efficiency: A Comparative Study of Deep Learning Architectures on a Limited Thai Food Dataset

Abstract

This study presents a structured benchmarking framework for evaluating five widely adopted convolutional neural network (CNN) architectures —namely DenseNet201, MobileNetV2, ResNet50, VGG19, and NASNetMobile, all initialized with ImageNet pre-trained weights and fine-tuned using transfer learning—under limited-data conditions for Thai food image classification using a curated dataset of 3,000 images across ten food categories. Rather than focusing solely on predictive accuracy, the proposed evaluation integrates computational efficiency metrics—training time, inference latency, and model size—to provide a balanced assessment of model suitability in resource-constrained contexts. Experimental results reveal a consistent trade-off between classification performance and computational demand, highlighting the importance of aligning architectural complexity with dataset scale and operational constraints. Notably, DenseNet201 achieved the highest test accuracy (0.94), while MobileNetV2 provided the most favorable efficiency profile with competitive accuracy (0.91) and the lowest computational overhead. The findings demonstrate that lightweight architectures can achieve competitive accuracy with substantially lower computational overhead, while high-capacity networks deliver marginal performance gains at increased resource cost. By introducing a normalized multi-criteria evaluation strategy, this work advances balanced benchmarking for culturally specific image classification tasks, particularly addressing the visual similarity among Thai dishes and limited availability of high-quality public datasets, and provides practical guidance for model selection in deployment-limited environments.