Diseño e implementación de un algoritmo basado en aprendizaje profundo para la extracción de propiedades y atributos discriminantes de frutos cítricos desde imágenes espectrales tipo VNIR.

Abstract

This study presents the development and implementation of a system based on hyperspectral imaging in the visible and near-infrared range (VNIR) and deep neural networks, designed to classify and evaluate the ripening stages of citrus fruits, specifically mandarins. Using a Specim FX10e hyperspectral camera operating between 400 and 1000 nm, hyperspectral datacubes were captured to analyze both internal and external properties of the fruits. The images were processed using dimensionality reduction techniques, such as Principal Component Analysis (PCA), to extract the most relevant spectral bands, retaining 97% of the useful information in the spectral signatures. For classification, a deep neural network was developed using the Levenberg-Marquardt backpropagation algorithm, configured to optimize the identification of ripening stages (green, partially ripe, and fully ripe). This approach achieved 100% precision in the confusion matrix, demonstrating the system’s effectiveness under controlled conditions. Segmentation and selection of specific bands (562.4 nm, 750 nm, and 951.7 nm) were key to noise reduction and improved class discrimination. Beyond its practical application in citrus fruit classification, this study significantly contributes to the field of hyperspectral spectroscopy and image processing by integrating deep learning and dimensionality reduction techniques with complex spectral data. The results demonstrate that hyperspectral analysis combined with neural networks can overcome the limitations of traditional methods, providing a non-destructive, precise, and scalable solution for quality assessment in agricultural products. This work lays the foundation for future research in automated classification systems, not only for citrus fruits but also for other agricultural products. Additionally, it offers a robust methodological framework that can be adapted to various industrial environments, driving the modernization of quality control in the agroindustry.