An Evaluation of YOLOv7 and YOLOv7-Tiny Models for Spool Counting Tasks on the Edge for Industrial Automation

This study introduces an automated system for estimating the net weight of thread in baskets within a fishing net manufacturing facility, addressing the critical need for accurate weight measurement of surplus threads produced during manufacturing for inventory management. The precision of this estimation relies on the accuracy of the object detection model employed, as it directly impacts predefined weight calculations for each spool and basket. Currently, manual counting and weighing methods involve significant labour for lifting and measurement, leading to inefficiencies, errors and reliance on human effort. To address these challenges and align with the factory’s policy of adopting artificial intelligence to enhance efficiency and reduce human error, YOLOv7, an object detection model, is employed to count spools and estimate their weight by subtracting the calculated spool and basket weights from the total weight. A programmable logic controller (PLC) based system ensures seamless automation, enabling efficient material handling throughout the detection and weighing process. Experimental results show that YOLOv7-Tiny achieves a detection accuracy of $1.0 \mathrm{mAP} \text{@} 0.5$, while YOLOv7 achieves $0.98 \mathrm{mAP} \text{@} 0.5$ on the test dataset. Additionally, the edge-device implementation on the NVIDIA Jetson Nano delivers precise weight estimation and real-time performance with low energy consumption, operating in 10 W mode.