A surface defect detection system for golden diamond pineapple based on CycleGAN and YOLOv4

The golden diamond pineapple produced in Taiwan has a large phyllotaxy and is prone to cracking between phyllotaxy. In early days, skilled warehouse staff checked the surface condition and picked out defective pineapples by manual visual inspection. However, the manual inspection method was very subjective and prone to misreading such as underkill or overkill. In order to improve the quality of surface defect detection for golden diamond pineapples, and to reduce reliance on manpower, this research pro -poses an automatic defect detection method based on the you only look once version four (YOLOv4). Its strategic advantage is that we only need a small training model in which the Taguchi method was used to determine the initiation of mosaic image augmentation, and the initiation of cycle-consistent generative adversarial network (CycleGAN) image augmentation, to obtain 84.86% final average precision (AP) under 6.41 frame per second of inference. The experiment showed that CycleGAN contributes the most among the model training strategies, not only because the pseudo defects generated by CycleGAN enrich the defect variety, but also the patches fit the texture after being pasted onto original position, and the patches could conduct the process of auto-annotation. To sum up, CycleGAN helps improve the perfor-mance of YOLOv4 defect detection.(c) 2022 The Authors. Published by Elsevier B.V. on behalf of King Saud University. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Automated summary

This research proposes an automatic defect detection method based on YOLOv4 to improve the quality of surface defect detection for golden diamond pineapples. By using the Taguchi method and CycleGAN image augmentation, a high precision rate and frame per second of inference were achieved. The experiment showed that CycleGAN contributes the most among the model training strategies and helps improve the performance of YOLOv4 defect detection.