Development of Convolutional Neural Network (CNN)-Based Automatic Food Classification (AFC) Model for Visual Impairments of Nigerians

Abstract

This study investigates the efficacy of a transfer learning approach utilizing LeNet-5 and a CNN-based AFC model for the classification of Nigerian swallow foods. Leveraging the lightweight LeNet-5 architecture, originally designed for 32×32×3 image inputs, we adapted our dataset with 100×100×3 images to meet its requirements. Meanwhile, the CNN-based AFC model was constrained to accept 100×100×3 images and classify them into 20 categories. Training, validation, and testing were conducted on both models, yielding distinct results. The CNN-based AFC model demonstrated superior performance, achieving an accuracy of 0.9494 and a validation accuracy of 0.9524, with corresponding losses of 0.4778 and 0.4735, respectively. Conversely, LeNet-5 achieved a training accuracy of 0.9418 and a validation accuracy of 0.9524, with losses of 0.0808 and 0.0421. Notably, the CNN-based AFC model exhibited significantly faster training times, outperforming LeNet-5 in this regard. This study lays the groundwork for further research, with plans to integrate the CNN-based AFC model into a mobile app to assist visually impaired individuals in identifying safe foods. The deployment of such a system has the potential to enhance the autonomy and quality of life for visually impaired individuals, reducing reliance on caregivers and mitigating health risks associated with improper food consumption. Future work will focus on expanding the model to encompass a wider range of Nigerian foods and finalizing its integration into the mobile app.