Optimizing Tomato Classification in Industrial Processing: A Comparative Study of YOLOv4, YOLOv8, and YOLOv11 with Rotation-Based Image Acquisition

Abstract

This study evaluates the effectiveness of YOLOv4, YOLOv8, and YOLOv11 models in classifying tomato quality during industrial processing. Using a dataset of 780 images, categorized as “viable” and “non-viable,” the models were assessed based on detection accuracy, processing speed, and resource efficiency. YOLOv4 achieved perfect accuracy ($100 \%$) with a precision of 0.99, recall of 1.00, and an F1-score of 0.99, but it had a slower processing time, averaging 64.8 milliseconds per image. YOLOv8, with $99 \%$ accuracy, processed images faster (11.7 milliseconds per image) and achieved an F1-score of $\mathbf{0. 9 7}$. YOLOv11 demonstrated balanced performance with $98 \%$ accuracy, the fastest processing time of 2.5 milliseconds, but a lower $F 1$-score of 0.93. A comprehensive evaluation of these models reveals that YOLOv8 emerges as the optimal choice for industrial tomato applications due to its ability to compromise on processing time accuracy for industrial-scale ketchup. This study’s practical contribution lies in identifying YOLOv8 as the most balanced model for automated tomato sorting in industrial food processing. YOLOv8’s superior speed and competitive accuracy enable real-time application, significantly improving efficiency without sacrificing quality control.

Access to Document

Other files and links

• Link to publication in Scopus

Fingerprint