基于高光谱图像与果蝇优化算法的马铃薯轻微碰伤检测  

Detection of Potato Slight Bruise Based on Hyperspectral Image and Fruit Fly Optimization Algorithm

作　　者：李小昱[1] 徐森淼[1] 冯耀泽[1] 黄涛[1] 丁崇毅 

机构地区：[1]华中农业大学工学院,武汉430070

出　　处：《农业机械学报》2016年第1期221-226,共6页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家自然科学基金项目(61275156);湖北省自然科学基金重点项目(2011CDA033)

摘　　要：针对通常采用的反射高光谱无法准确检测随机放置马铃薯表面轻微碰伤的问题,提出了一种用V型平面镜的高光谱并结合果蝇优化算法(FOA)检测马铃薯轻微碰伤的方法。试验搭建了V型平面镜反射高光谱图像采集系统,分别采集随机放置下的轻微碰伤和合格马铃薯的高光谱图像,每张高光谱图像包含平面镜1反射图像F1、相机直接采集图像F2、平面镜2反射图像F3,分别提取F1、F2、F3感兴趣区域的平均光谱拼接成马铃薯的属性矩阵。采用标准正态变量变换(SNV)预处理后的光谱矩阵进行全波段的支持向量分类机(SVC)建模,预测集的识别率仅为84.11%;为了提高模型的性能,采用蚁群算法(ACO)进行变量优选,优选出9个变量建立的SVC模型预测准确率为95.32%;分别用网格搜索法(Grid search)、遗传算法(GA)和FOA对SVC的惩罚参数c和核函数参数g进行寻优,通过比较分析,FOA-SVC对训练集和预测集的识别准确率均达到100%。试验结果表明,用V型平面镜的高光谱结合FOA-SVC能够准确检测马铃薯的轻微碰伤,可为马铃薯的轻微碰伤在线检测提供技术基础。Potato is an indispensable food crop for the people in the world. As a kind of light injury on the surface of potato,slight bruise of potato cannot be accurately tested when potato placed in random orientation. This paper proposed a method by combining hyperspectral image based on V-shaped plane mirror with fruit fly optimization algorithm（ FOA） to identify slight bruise of potato randomly placed. In this study,hyperspectral imaging system was built based on V-shaped plane mirror and 322 potato samples were bought as the research subjects. To meet with the practical production,within half an hour after bruise occurred,potatoes were placed in three positions： the damage part facing to camera,side to camera,and back to camera. Then hyperspectral images of all potatoes were collected including reflection image F1 in mirror 1,image F2 directly obtained by camera and reflection image F3 in mirror 2. Average spectrums from these three images were spliced into attribute matrix of sample. Support vector classifier（ SVC） model was established in full bands after utilizing standard normal variate（ SNV） and the recognition accuracy of prediction set was only 84. 11%. Variable selection was processed by ant colony optimization（ ACO）. Nine spectral variables（ 762 nm,879 nm in F1; 711 nm,957 nm,1 020 nm in F2;510 nm,746 nm,1 000 nm,1 007 nm in F3） were selected and the recognition rate reached 95. 32%.FOA,genetic algorithm（ GA） and grid search were respectively applied to search the best penalty parameter c and kernel function parameter g. By comparing results of those models,FOA obtainedoptimal parameters（ c = 11. 076 3,g = 9. 262 5）. FOA-SVC was proved to be the best model and the training set and prediction set recognition accuracy both reached 100%. The results show that the combination of hyperspectral image based on V-shaped plane mirror with FOA-SVC could accurately detect the slight bruise of potato.

关 键 词：马铃薯 轻微碰伤 V型平面镜 高光谱成像 果蝇优化算法 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]