Desarrollo de un Sistema de Visión Artificial para la Detección Precisa de Subllenado, Llenado Correcto y Sobrellenado en Botellas

The study aimed to develop an intelligent system based on computer vision techniques and YOLOv8 for the detection and control of fill level in bottles, in order to achieve an accuracy greater than 97% and reduce errors due to overfilling or underfilling. To achieve this, an optimized YOLOv8-seg model was trained for multiclass classification (full, under full, and overfull) and the effectiveness of a computer vision system to correctly identify whether a bottle is full or not was evaluated using metrics such as precision, recall, F1-score and [email protected]. The methodology consisted of training the model with a set of 4384 images, divided into 80% for training and 20% for validation, for 100 epochs. Metric tracking showed that the training loss progressively decreased to values close to 0, while the validation loss remained stable at -0.04. The top-1 accuracy reached a final value of 98.5%, while the top-5 accuracy remained at 100%. Following evaluation on test data, a normalized confusion matrix was constructed, which showed a final accuracy of 99%. The precision-recall curve complemented the analysis, revealing a favorable balance between recall and precision in both classes ("full”, “under full”, and “overfull"), with no evidence of overfitting. Furthermore, an F1 score of 96% was achieved, indicating a solid balance between precision and sensitivity, and an [email protected] of 98.6%, confirming the model's high ability to correctly locate and classify relevant objects.