Recursive-YOLOv5 Network for Edible Mushroom Detection in Scenes With Vertical Stick Placement

IEEE Access

Yiliang Sun¹  Bohan Wei¹  Shihan Zhang¹  Kejun Wang¹  Yufan Pu¹  Yahui Zhao¹  Hongyu Lin¹  Changfan Zeng¹  Yao Zhang²  Zhiyong Chen² 

[1] Department of Electronic and Information Engineering, Tongji Zhejiang College, Jiaxing, China;Department of Physical Education, Tongji Zhejiang College, Jiaxing, China;

关键词: Edible mushroom;    object detection;    ASPP;    YOLOv5;    deep learning;   

DOI  :  10.1109/ACCESS.2022.3165160

来源: DOAJ

【 摘 要 】

In many rural edible mushroom growing sites, due to their growing environments and economic factors, growers are often unable to use industrial automation to pick edible mushrooms, resulting in losses of hundreds of millions of dollars each year due to missing the optimal picking cycles of edible mushrooms. For this reason, this project aimed to find an object detection algorithm that can be used with industrial cameras to help detect the growth statuses of edible mushrooms in real time and can be applied to future automatic picking machines. In this paper, based on the You Only Look Once version 5 (YOLOv5) network, by introducing the idea of recursion, remerging the first output with the convolutional layer of the backbone part of the network, introducing atrous spatial pyramid pooling (ASPP) instead of SPP (as in the original algorithm), utilizing the complete intersection over union (CIOU) and distance IOU (DIOU) instead of the generalized IOU (GIOU), and employing DIOU_nonmaximum suppression (NMS) instead of the original NMS algorithm, we propose a new deep learning network called Recursive-YOLOv5. This network can effectively identify 98% of edible mushrooms when dealing with large-resolution, small-target situations; this is a 12.87% improvement over the accuracy of YOLOv5X. Although the algorithm proposed in this paper utilizes almost double the parameters of the original network, with today’s computing power and the rapid development of cloud computing, this sacrifice of computing power for accuracy is very cost-effective.

【 授权许可】

Unknown