Implementación de un sistema de clasificación de manzanas mediante visión artificial para control de calidad

Abstract

This degree project describes the implementation of an automatic apple classification system using artificial vision, designed to optimize quality control in the agri-food industry. Traditional manual inspection presents significant limitations in terms of subjectivity, operator fatigue, and inconsistency in selection criteria, which affects the competitiveness of the final product. For the implementation of the mechatronic prototype, a controlled image acquisition station was developed, equipped with a Logitech C922 camera and a diffuse lighting system. Data processing was carried out using Deep Learning algorithms, comparatively evaluating three convolutional neural network architectures: MobileNetV2, EfficientNet-B0, and ConvNeXt-Nano. A hybrid dataset was constructed by combining local images with external repositories, applying data augmentation techniques and the Focal Loss function to mitigate class imbalance. Experimental results determined that the ConvNeXt-Nano model offered superior performance, achieving an accuracy of 99.27 % and an F1-Score of 0.99, outperforming the other architectures in detecting subtle defects. The final system, deployed on a processing unit with an NVIDIA GTX 1650 GPU, operates in real- time with an average latency of 35 ms per frame, validating its technical feasibility for the automation of selection processes.