| 改进YOLOv5的香菇子实体生育期识别方法 |
| |
| 引用本文： | 杨林,曾大鑫,边银丙,陈红,宗望远,龚钰华.改进YOLOv5的香菇子实体生育期识别方法[J].农业工程学报,2024,40(9):182-189. |
| |
| 作者姓名： | 杨林 曾大鑫 边银丙 陈红 宗望远 龚钰华 |
| |
| 作者单位： | 华中农业大学工学院，武汉 430700;农业农村部长江中下游农业装备重点实验室，武汉 430700;华中农业大学植物科技学院，武汉 430700 |
| |
| 基金项目： | 湖北省重大科技攻关项目“食用菌精细栽培智能化装备研发与应用”（2020BBA040） |
| |
| 摘 要： | 在香菇栽培中，需要评估其生长发育状态，以便调控栽培环境和采取适当的栽培措施。针对香菇生育期子实体外观特征变化不显著，机器自动采收时部分成熟期香菇子实体易误检和漏检的问题，该研究提出了一种基于改进YOLOv5的香菇子实体生育期识别方法。首先替换YOLOv5模型中上采样模块，采用一种包含上采样预测模块和特征重组模块的轻量级上采样模块；其次在YOLOv5l模型中添加小目标检测层，增加模型对香菇子实体生育期特征信息的提取，提高模型区分香菇生育期和识别小香菇的能力。试验结果表明，改进的 YOLOv5l 模型具有较好的检测能力，平均帧率为 45.25 帧/s，平均精确度均值为92.70%，与SSD、Faster-RCNN、Mushroom-YOLO和YOLOv5相比平均精确度均值分别提升22.6、28.38、6.8和2.5个百分点。该研究方法能够满足对香菇子实体不同生育期识别的精度与速度要求，为香菇子实体生育期识别提供了一种方法参考。
|
| 关 键 词： | 图像处理 智慧农业 深度学习 香菇生育期 目标识别 YOLOv5 |
| 收稿时间： | 2023/3/4 0:00:00 |
| 修稿时间： | 2024/4/11 0:00:00 |
Recognizing fruiting body growth period of Lentinus edodes using improved YOLOv5 |
| |
| YANG Lin,ZENG Daxin,BIAN Yinbing,CHEN Hong,ZONG Wangyuan,GONG Yuhua.Recognizing fruiting body growth period of Lentinus edodes using improved YOLOv5[J].Transactions of the Chinese Society of Agricultural Engineering,2024,40(9):182-189. |
| |
| Authors: | YANG Lin ZENG Daxin BIAN Yinbing CHEN Hong ZONG Wangyuan GONG Yuhua |
| |
| Institution: | Huazhong Agriculture University, College of Engineering, Wuhan 430700, China;Key Laboratory of Agricultural Equipment in Mid-lower Yangtze River, Ministry of Agriculture and Rural Affairs, Wuhan 430070, China;Huazhong Agriculture University, College of Plant Science, Wuhan 430700, China |
| |
| Abstract: | Lentinus edodes can benefit from artificial Intelligence (AI) and Internet of Things (IoT) technologies in indoor farming. The mushroom growth can be cultivated by hand-free robots in greenhouses. It is necessary to evaluate the growth status in such next-generation scenarios of agricultural production. The cultivation environment can be regulated to take the optimal management. Since the fruiting bodies of Lentinus edodes cannot change significantly in the growth period, some mature ones cannot be detected by machine vision in the automatic machine harvesting. Specifically, Lentinus edodes bodies are distributed randomly and crowded. Moreover, the target detection of such mushrooms is also interfered from the image background in the texture of nutrient sticks on planting stands. In this study, a rapid and accurate detection was proposed to identify the growth period of Lentinus edodes using improved YOLOv5l. Firstly, the lightweight upsampling prediction and feature reorganization modules were replaced in the YOLOv5l model. Secondly, the detection layer of small target was added into the YOLOv5l model. The characteristic information was enhanced to distinguish the growth period and small of fruiting Lentinus edodes. Finally, the experimental results show that the better detection was achieved in the improved YOLOv5l model. A data set was collected using edge devices with inexpensive camera probes and an IoT platform. 3470 pictures were selected to train the improved model. The average precision rate of the proposed algorithm can be 92.7%. The improved YOLOv5l model was 0.6% higher than the highest average accuracy of mushroom recognition in the young mushroom stage, 2.9% higher than the highest average accuracy of mushroom recognition in the mature stage, 0.2% higher than the highest average precision rate, and 0.2% higher than the highest average recall rate, compared with the original. The rate also increased by 1.7 percentage points. However, the average frame rate of the improved YOLOv5l model decreased to 45.25 frames/s, compared with the original (57.47 frames/s). There was the increase in the inference time that caused by the increase of small target layers. In addition, the performance of the improved model increased the average accuracy by 22.6%, 28.38%, 6.8%, and 2.5%, respectively, compared with the SSD, Faster-RCNN, Mushroom-YOLO, and YOLOv5l. This improved model can meet the accuracy and speed requirements for the identification of different growth stages of Lentinus edodes fruiting bodies. Real-time growth assessment can be used to predict the mushroom yield in intelligent production. The existing data set of the growth period of mushroom fruiting bodies should be expanded to construct the national standards for mushroom classification at the mature grades. Maturity identification and yield prediction are planned to further explore the collaborative control of environmental facilities, robot picking, and harvest logistics. |
| |
| Keywords: | image processing smart agriculture deep learning the growth period of lentinus edodes target detection YOLOv5 |
|
| 点击此处可从《农业工程学报》浏览原始摘要信息 |
| 点击此处可从《农业工程学报》下载免费的PDF全文 |
|
