An Explorative Analysis of SVM Classifier and ResNet50 Architecture on African Food Classification
Chinedu Emmanuel Mbonu, Kenechukwu Anigbogu, Doris Asogwa, Tochukwu Belonwu
TL;DR
The paper tackles the problem of African cuisine recognition by directly comparing a fine-tuned ResNet50 CNN with a traditional SVM classifier on a dataset of 1,658 images spanning six dishes. It implements a ResNet50-based transfer-learning pipeline with data augmentation and a separate SVM using flattened RGB features, both evaluated with accuracy, precision, recall, F1, and confusion matrices under a 5-fold cross-validation framework. The results show both approaches achieving 81% overall accuracy but with complementary strengths across classes: ResNet50 excels on Palm Nut Soup while SVM demonstrates more consistent performance across most categories, especially Ekwang and Eru. The study contributes actionable insights for African food classification, demonstrates the viability of both paradigms on limited/imbalanced data, and provides code for reproducibility, laying groundwork for ensembles and larger, balanced datasets in future work.
Abstract
Food recognition systems has advanced significantly for Western cuisines, yet its application to African foods remains underexplored. This study addresses this gap by evaluating both deep learning and traditional machine learning methods for African food classification. We compared the performance of a fine-tuned ResNet50 model with a Support Vector Machine (SVM) classifier. The dataset comprises 1,658 images across six selected food categories that are known in Africa. To assess model effectiveness, we utilize five key evaluation metrics: Confusion matrix, F1-score, accuracy, recall and precision. Our findings offer valuable insights into the strengths and limitations of both approaches, contributing to the advancement of food recognition for African cuisines.