基于改进YOLOv5算法的马铃薯表皮缺陷程度检测方法研究

Research on Potato Defect Detection Based on Improved YOLOv5

马铃薯作为一种产量可观、营养丰富的农作物,已经成为全球不可或缺的食物之一。但恰恰因为其体量庞大的特点在对马铃薯进行分类出售时需要耗费大量的人力和物力以及时间。为了实现对马铃薯品质的自动分类,解放人力物力,提升效率。我们提出了一种基于计算机视觉及改进特征融合YOLOv5s算法的马铃薯表皮缺陷程度检测方法,我们把YOLOv5s的颈部网络中的特征金字塔网络结构替换为加权特征金字塔网络结构,采用这种双向加权特征网络能够更好的提取特征信息,更好的融合特征。并且我们加入了二分K均值聚类算法,该算法的加入极大提升了检测时的收敛速度和精度,并且有效避免了K均值聚类算法因初始聚类点质心选取不适所带来的影响。经过我们的实验表明,本项技术能够对马铃薯表皮检测的正确率达到98%。由此可见,本项基于改进YOLOv5算法的马铃薯表皮缺陷程度检测方法可行性较强,可以用于市场对马铃薯检测分类。

Potato,as a crop with considerable yield and rich nutrition,has become one of the world's indispensable foods.But precisely because of its large size,it takes a lot of manpower and material resources and time to classify potatoes for sale.In order to realize the automatic classification of potato quality,the liberation of manpower and material resources,improve efficiency.We proposed a potato skin defect detection method based on computer vision and improved feature fusion YOLOv5s algorithm.We replaced the feature pyramid network structure in the neck network of YOLOv5s with a weighted feature pyramid network structure.This two-way weighted feature network can better extract feature information and better integrate features.In addition,we added the binary K-means clustering algorithm,which greatly improved the convergence speed and accuracy of detection,and effectively avoided the influence of K-means clustering algorithm caused by the inappropriate selection of centroid of initial clustering points.Our experiments show that this technology can detect the potato skin correctly up to 98%.It can be seen that the potato skin defect degree detection method based on the improved YOLOv5 algorithm is highly feasible and can be used in the market to detect and classify potatoes.