Utilizing a YOLOv8 Segmentation-Based Model for Automated Defect Detection in Bread Images

Abstract
This paper evaluates the YOLOv8 model for its effectiveness in identifying defects in bread within the Mexican market, supporting the decision to use this state-of-the-art single-shot object detection convolutional neural network. The model demonstrates remarkable proficiency in detecting common defects in various types of bread, with a specific focus on identifying burned and impacted pieces. Achieving up to 51.7 frames-per-second (FPS) across various scenarios and image detection tasks, the model exhibits optimal recall and Mean Average Precision (MAP) metrics, with a score of 0.825 for the mAP50 metric. Rigorous testing and validation against relevant datasets highlight the model’s reliability and efficacy in quality control applications within the bread industry. Implementing the YOLOv8 system promises significant improvements in quality control processes for bread production facilities of all scales, from small artisan bakeries to large-scale industrial operations. By enabling quick and automated defect identification, the system can reduce product wastage, enhance production effectiveness, and increase financial gains. Additionally, transfer learning is identified as a promising approach to customize the model for detecting defects in sweet bread, considering the similarity in characteristics between salty and sweet varieties. This paper underscores the advanced capabilities of YOLOv8 and its potential use in quality control during bread production in the Mexican market.