Procesamiento de imágenes para la clasificación de granos de café por forma y color

Abstract

The present project focuses on the development of an image processing algorithm based on Convolutional Neural Networks (CNN) for the classification of coffee beans, aiming to distinguish between beans in good condition and those with defects based on irregularities in shape and color. The chosen CNN model is MobileNet, which is optimized for convolutional operations to create deep neural networks with fewer parameters, resulting in low computational requirements and low latency. The manual visual inspection currently performed by many coffee producer associations for proper selection of exported coffee motivated the objective of this project. As coffee is highly valued both in Ecuadorian society and globally, ensuring high international prices for the exported product requires careful selection. To evaluate the performance of the coffee bean classification algorithm, a comparative analysis was conducted using the MobileNetV1 and MobileNetV2 architectures, comparing loss statistics. The results of the architecture comparison demonstrated a precision of 99.18% and validation accuracy of 99.20% for the MobileNetV2 model, while the MobileNetV1 model achieved a precision of 88.72% and validation accuracy of 81.25%. Consequently, the performance of the MobileNetV2-based model was considered the optimal choice for coffee bean classification.