基于卷积神经网络和支持向量机的水稻种子图像分类识别  

Rice Seed Image Classification and Recognition Based on CNN_SVM

作　　者：杨红云[1] 黄琼 孙爱珍 王映龙[2] 肖小梅 罗建军 Yang Hongyun;Huang Qiong;Sun Aizhen;Wang Yinglong;Xiao Xiaomei;Luo Jianjun(College of Software,Jiangxi Agricultural University,Nanchang 330045;College of Computer and Information Engineering,Jiangxi Agricultural University,Nanchang 330045)

机构地区：[1]江西农业大学软件学院,南昌330045 [2]江西农业大学计算机与信息工程学院,南昌330045

出　　处：《中国粮油学报》2021年第12期144-150,共7页Journal of the Chinese Cereals and Oils Association

基　　金：国家自然科学基金(61562039)。

摘　　要：针对外形相似的水稻种子间分类难、识别正确率低等问题,提出一种卷积神经网络与支持向量机相结合的方法(CNN_SVM)对8类水稻种子进行分类识别。首先对图像进行切割、旋转等预处理后建立水稻种子图像数据库,其次通过提取图像的方向梯度直方图(HOG)、LBP纹理、SIFT描述子和CNN特征,分别建立SVM、KNN和Softmax分类模型对水稻种子图像进行分类识别比较。最后采用随机加入噪声点方法模拟噪声干扰稻种和调整色彩饱和度方法模拟不同年份稻种后进行分类识别。结果表明CNN_SVM模型对正常、噪声干扰和不同年份的水稻种子图像识别正确率分别为96.2%、95.8%和96.1%,识别单张图像时间为4.57 ms,明显优于CNN、SVM的传统模型。模型的抗噪和泛化能力强,能满足实际生活中水稻种子分类识别需求。Aiming at the problems of difficulty in classifying rice seeds with similar appearance and low recognition accuracy,a method combining convolutional neural network and support vector machine(CNN_SVM)is proposed to classify and recognize 8 types of rice seeds.First,the image is preprocessed by cutting,rotating,etc.,to establish a rice seed image database.Second,by extracting the image's orientation gradient histogram(HOG),LBP texture,SIFT descriptor and CNN features,the SVM,KNN and Softmax classification models are established respectively.The rice seed images are classified,recognized and compared.Finally,the method of randomly adding noise points to simulate the seeds of noise interference and the method of adjusting the color saturation to simulate the seeds of different years are used for classification and identification.The results show that the CNN_SVM model has a correct recognition rate of 96.2%,95.8%and 96.1%for normal,noisy,and different years of rice seed images,and the recognition time for a single image is 4.57 ms,which is significantly better than the traditional models of CNN and SVM.The model has strong anti-noise and generalization ability,which can meet the needs of rice seed classification and identification in real life.

关 键 词：水稻种子 卷积神经网络 分类识别 支持向量机 

分 类 号：S126[农业科学—农业基础科学]