| 基于迁移学习与YOLOv8n的田间油茶果分类识别 |
| |
| 引用本文： | 周宏平,金寿祥,周磊,郭自良,孙梦梦,施明宏.基于迁移学习与YOLOv8n的田间油茶果分类识别[J].农业工程学报,2023,39(20):159-166. |
| |
| 作者姓名： | 周宏平 金寿祥 周磊 郭自良 孙梦梦 施明宏 |
| |
| 作者单位： | 南京林业大学机械电子工程学院，南京 210037 |
| |
| 基金项目： | 国家林业和草原局应急科技项目（202202-3） |
| |
| 摘 要： | 为降低视觉引导式油茶果采摘机器人采摘被遮挡油茶果时造成的果树和抓取装置损伤，该研究提出了一种基于迁移学习和YOLOv8n算法的油茶果分类识别方法，将油茶果分成无遮挡和遮挡两类。首先，采用COCO128目标检测数据集作为源域，苹果数据集为辅助域的迁移学习方法训练模型。其次，将学习方法、训练数据量、学习率和训练轮数这4种因素组合，共进行了52组YOLOv8n检测性能的消融试验。最后，将YOLOv8n模型与YOLOv3-tiny、YOLOv5n和YOLOv7-tiny等模型进行比较。试验结果表明，随机权重初始化方式受训练数据量和学习率影响较大，学习率为0.01时模型检测效果最好；而迁移学习方法仅用随机权重初始化1/2的数据量即可达到与其相当的平均精度均值；迁移学习方式下，YOLOv8n模型的平均精度均值最高达到92.7%，比随机权重初始化方式提升1.4个百分点。与YOLOv3-tiny、YOLOv5n和YOLOv7-tiny等模型相比，YOLOv8n模型的平均精度均值分别提高24.0、1.7和0.4个百分点，研究结果可为YOLOv8n模型训练参数优化和油茶果分类识别提供参考。
|
| 关 键 词： | 图像识别 深度学习 油茶果 YOLOv8n 迁移学习 目标检测 |
| 收稿时间： | 2023/7/26 0:00:00 |
| 修稿时间： | 2023/10/9 0:00:00 |
Classification and recognition of camellia oleifera fruit in the field based on transfer learning and YOLOv8n |
| |
| ZHOU Hongping,JIN Shouxiang,ZHOU Lei,GUO Ziliang,SUN Mengmeng,SHI Minghong.Classification and recognition of camellia oleifera fruit in the field based on transfer learning and YOLOv8n[J].Transactions of the Chinese Society of Agricultural Engineering,2023,39(20):159-166. |
| |
| Authors: | ZHOU Hongping JIN Shouxiang ZHOU Lei GUO Ziliang SUN Mengmeng SHI Minghong |
| |
| Institution: | College of Mechanical and Electronic Engineering, Nanjing Forestry University, Nanjing 210037, China |
| |
| Abstract: | Visually guided grabbing has been widely used for picking camellia oleifera fruits at present. However, picking robots cannot directly reach the camellia oleifera fruits that are severely obstructed by the stems, even damaging the fruit tree or clamping devices. In this study, the classification and recognition were proposed for the camellia oleifera fruits using transfer learning and YOLOv8n, in order to reduce the manual annotation workload for the detection accuracy of the model. Camellia oleifera fruits were also divided into two categories: no occlusion and occlusion. Firstly, transfer learning was designed using the COCO128 dataset as the source domain, the publicly available apple dataset as the auxiliary domain, and the self-made camellia oleifera fruit dataset as the target domain. The YOLOv8n model was then trained to recognize the camellia oleifera fruits. Secondly, a total of 52 ablation experiments were performed using two learnings (transfer learning and random weight initialization training), three training datasets (consisting of 200, 400, and 732 training images), three learning rates (0.001, 0.005, and 0.01), and three training rounds (50, 100, and 200 rounds). A systematic investigation was then made to explore the impact of these four factors on the detection performance of the YOLOv8n model. Finally, YOLOv8n was compared with the models, such as YOLOv3 tiny, YOLOv5n, and YOLOv7 tiny. The experimental results show that the training mode initialized with random weights was very sensitive to the amount of training data and the size of the learning rate. The larger the learning rate was, the faster the convergence speed of the model was, with an optimal value of 0.01. The larger the dataset was, the higher the average accuracy of the model was, but there were marginal effects. The training of transfer learning significantly accelerated the convergence speed of the model. The peak detection accuracy was achieved within the first 50 rounds of training. The mean average accuracy was improved by 2-24 percentage points, compared with the random weight initialization. The transfer learning was achieved in the mean average accuracy equivalent to the random weight initialization with only 1/2 of the training data, which greatly reduced the manual annotation time. The mean average accuracy of the YOLOv8n model reached the maximum of 92.7% under the transfer learning, which was 1.4 percentage points higher than that with random weight initialization. Compared with the lightweight YOLO series models, such as YOLOv3 tiny, YOLOv5n, and YOLOv7 tiny, the mean average accuracy of the YOLOv8n model was 24, 1.7, and 0.4 percentage points higher, respectively, indicating the best detection performance. The experiments also verified the performance of the YOLOv8n model and the effectiveness of the transfer learning. The finding can also provide a strong reference to optimize the training parameters of the YOLOv8n model for better classification and recognition of camellia oleifera fruits. |
| |
| Keywords: | image recognition deep learning camellia oleifera fruit YOLOv8n transfer learning target detection |
|
| 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 |
|
