基于K-means聚类和改进MLP的苹果分级研究  

Grading of Apple Based on K‐means Clustering and Improved MLP

作　　者：王迎超 张婧婧[1] 贾东霖 周腾飞 WANG Yingchao;ZHANG Jingjing;JIA Donglin;ZHOU Tengfei(School of Computer and Information Engineering,Xinjiang Agricultural University,Urumqi 830052,China)

机构地区：[1]新疆农业大学计算机与信息工程学院,新疆乌鲁木齐830052

出　　处：《河南农业科学》2023年第1期161-171,共11页Journal of Henan Agricultural Sciences

基　　金：新疆维吾尔自治区自然科学基金资助项目(2022D01A202);新疆农业大学研究生教育教学改革研究项目(xjaualk-yjs-2021012);新疆农业大学2022年度大学生创业训练项目(XJCY202219)。

摘　　要：为准确实现多特征融合的苹果分级,提出了一种基于K-means聚类和改进MLP的苹果分级方法。该方法主要包括图像预处理、亮度均衡化、背景分割、特征加权以及改进的MLP分级网络训练。首先借助均值滤波算法和直方图均衡化操作改善苹果图像质量;接着借助K-means聚类算法进行背景分割;在果体与背景分割的基础上,依次提取苹果的果径、果形、颜色、缺陷、纹理5个特征;然后借助皮尔逊相关性分析和人工挑选偏好权重对特征数据集综合加权,模拟人工分级场景;最后将特征数据送入改进的MLP神经网络中完成苹果的分级定等。通过对400个定好等级的苹果进行分级测试,准确率达到94.25%,验证了分级方法的可行性与准确性。该方法与现行的苹果分级标准相结合,具备时效性强、检测指标完备等分级优势。In order to accurately realize apple classification with multi‐feature fusion,a classification method of Fuji apples based on K‐means clustering and improved MLP was proposed.The method mainly included image preprocessing,brightness equalization,background segmentation,feature weighting and improved MLP classification network.Firstly,the image quality of apple was improved by means of mean filtering algorithm and brightness equalization operation;then background segmentation was performed by means of K‐means clustering algorithm;on the basis of fruit body and background segmentation,the fruit diameter,fruit shape,color,defect and texture features of apple were extracted in turn;then Pearson correlation analysis and artificial preference weight were used to comprehensively weight the feature data set to simulate the artificial grading scene;finally,the feature data was sent to the improved MLP neural network to complete the apple grading.Through the grading test of 400 graded apples,the accuracy rate reached 94.25%,which verified the feasibility and accuracy of the grading method.This method corresponds to the systematic apple grading standard,and has the advantages of strong timeliness and complete detection indicators.

关 键 词：K-MEANS聚类 皮尔逊相关系数 多特征融合 改进MLP 苹果分级 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] S126[自动化与计算机技术—计算机科学与技术]