Mobile Application Deployment of CNN-based Food Classification System for Visually Impaired Individuals: A Real-Life Application for Nigerian Swallow Foods

The study addresses the significant challenge of food identification faced by visually impaired individuals, particularly in culturally specific contexts like Nigerian cuisine. A convolutional neural network (CNN) classification system was developed to categorize Nigerian swallow foods and assess their freshness (Fresh, 4 Hours Old, 1 Day Old) across 60 fine-grained classes using a dataset of 22,823 items. Six light CNN architectures were trained and evaluated based on metrics such as accuracy, macro-averaged precision, recall, F1 score, and AUC-ROC, alongside their parameter count and inference latency for deployment suitability. Among these, MobileNetV2 emerged as the top performer, achieving a test accuracy of 99.78%, a macro F1 score of 0.9979, and a perfect AUC-ROC of 1.0000, while maintaining a moderate parameter count of 2.33 million, making it suitable for mobile applications. LeNet-5 and ShuffleNetV2 also demonstrated strong performances with accuracies above 98.9%. In contrast, CCFNN exhibited poor performance with only 23.54% accuracy due to capacity issues. The best-performing model, MobileNetV2, was integrated into an application named Naija Food Eye, which facilitates food identification and freshness assessment for the visually impaired, providing confidence scores, top-5 predictions, and quick response times. The findings confirm that lightweight CNN models can effectively assist visually impaired individuals in recognizing foods within a culturally rich culinary landscape, offering a viable, deployable solution.