YOLOv8改进算法在油茶果分拣中的应用  

Research on sorting task of Camellia oleifera fruit based on YOLOv8 improved algorithm

作　　者：刘姜毅 高自成[1] 刘怀粤 尹浇钦 罗媛尹 LIU Jiangyi;GAO Zicheng;LIU Huaiyue;YIN Jiaoqin;LUO Yuanyin(School of Mechanical and Electrical Engineering,Central South University of Forestry and Technology,Changsha 410004,China)

机构地区：[1]中南林业科技大学机电工程学院,长沙410004

出　　处：《林业工程学报》2025年第1期120-127,共8页Journal of Forestry Engineering

基　　金：国家重点研发计划(2022YFD2202103);国家林业和草原局应急科技项目(202202-2);井冈山农高区省级科技专项揭榜挂帅项目(20222-051247)。

摘　　要：现有的油茶果分拣系统所依赖的YOLO等算法的目标检测、实例分割在低尺寸及密集型样本中鲁棒性较差,存在机械臂常抓取到枝叶、抓取不牢固、易脱落等问题。大部分系统使用目标识别,无法准确识别油茶果具体轮廓信息,不能对油茶果进行大小分类。针对这一问题,研究提出了YOWNet模型应对油茶果分拣的小目标、高密度识别任务。首先,研究了自动化边缘标注脚本,脚本调用零样本Segment Anything框架对原有已标注的油茶果目标检测框提取兴趣区间,将其自动转化为边缘标注信息;其次,为了提高模型对小目标的识别能力,研究摒弃了现有的固定感受野的卷积模块,针对油茶果特性提出三维注意力动态卷积模块用于捕捉特征图中的关键信息;最后,研究通过使用Wise⁃IoU损失函数,基于动态非单调聚焦机制的边界框损失,提升边框回归精度。总体网络模型命名为YOWNet,通过与YOLOv8在油茶果上的消融实验对比,试验结果表明:YOWNet模型能够快速准确地识别油茶果实例,在私有数据集上,准确度、Box_loss可达89.90%和0.523。With the rapid development of artificial intelligence,it is of great significance to incorporate mechanisms such as neural vision into the consideration of automation in agriculture,forestry and other industries.The existing Camellia oleifera fruit sorting system relies on algorithms such as YOLO,and the target detection and instance segmentation of these algorithms have poor robustness in low⁃dimensional and dense samples,resulting in problems such as the robotic arm often grasping branches and leaves,loose grasping,and easy shedding.Most systems use target recognition and cannot accurately identify the specific contour information of C.oleifera fruits and cannot classify the size of C.oleifera fruits.To address these issues,this study proposed the YOWNet model to deal with the small⁃target and high⁃density recognition tasks of C.oleifera fruit sorting.Firstly,a C.oleifera fruit dataset was constructed.By visiting the C.oleifera fruit plantation for on⁃site shooting,the collected data was subjected to image screening,labeling,enhancement,and division.Based on this,an automated edge labeling script was studied.The script called the zero⁃sample Segment Anything framework,in which,firstly the original labeled C.oleifera fruit target detection box in the LabelImg application was extracted,and then the region of interest was obtained to automatically convert into the edge labeling information required by the Labelme application.Secondly,the latest YOLOv8 instance segmentation model was used and improvements were achieved on this basis.To improve the models recognition ability for small targets,this study abandoned the existing fixed receptive field convolution module of YOLOv8 and proposed a three⁃dimensional attention dynamic convolution module for capturing key information in the feature map based on the characteristics of C.oleifera fruits.Finally,the bounding box regression accuracy was improved using the Wise⁃IoU loss function and the bounding box loss based on the dynamic non⁃monotonic focusing

关 键 词：油茶果 三维动态卷积 实例分割 YOLOv8 Segment Anything Model Wise⁃IoU 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]