A Transfer Learning Approach to Mango Image Classification

Côte d'Ivoire is one of West Africa’s leading mango exporters, significantly contributing to its economy. With an annual production of around 180,000 tons, fresh mango is the third most exported fruit from Côte d'Ivoire. The country is the third biggest supplier to the European market, with 32,400 tons a year. However, to maximize profits, farmers must accurately assess mangoes’ ripening stage. This paper proposes a CNN-based system using transfer learning to accurately detect mango ripeness. Our study focuses on developing a system using renowned CNN algorithms such as VGG19, ResNet101, and DenseNet121. The model was first trained on a dataset of 996 mango images. Next, it was refined on a smaller dataset specific to the task. Finally, it was evaluated on a test dataset. The DenseNet121 model performed best, with an accuracy of 97.50%. These machine learning techniques were chosen because they can improve the accurate identification of mango maturity. The contributions of our research are manifold: not only have we proposed an accurate solution for detecting mango ripeness, but we have also evaluated the performance of different models of CNN algorithms. This system will strengthen Côte d’Ivoire’s position as a leader in the African mango industry by reducing product losses associated with mango harvesting and increasing export volumes.