结合复频域多特征融合的轮钟花、金钱豹、大花金钱豹植物种子智能鉴别研究  

Intelligent identification of the seeds of Cyclocodon lancifolius,Campanumoea javanica and Campanumoea javanica subsp.javanica based on multi-characteristic fusion in complex frequency domain

作　　者：徐雪 郭文凯 罗祥敏 许成艳 吴芳芳 XU Xue;GUO Wenkai;LUO Xiangmin;XU Chengyan;WU Fangfang(Shizhen College of Guizhou University of Traditional Chinese Medicine,Guiyang 550200,China)

机构地区：[1]贵州中医药大学时珍学院,贵州贵阳550200

出　　处：《贵州科学》2024年第4期12-18,共7页Guizhou Science

基　　金：贵州省高等学校教学内容和课程体系改革项目(GZJG20220835)。

摘　　要：目的:为了提高目级别轮钟花、金钱豹、大花金钱豹种子分类的精度,本研究提出了一种结合复频域和空间域多特征融合的识别算法。方法:首先,使用体视镜拍摄三种种子显微图像;其次,在空间域中提取种子图像的颜色特征和方向梯度直方图特征,使用多方向对偶树复小波变换(M-DTCWT)对图像进行多尺度多方向复频域变换,在低频子带图像中提取形状特征和纹理特征;最后,使用ReliefF算法将得到的空间域特征和复频域特征进行融合,并利用SVM(支持向量机)、BPNN(BP神经网络)、RF(随机森林)三个分类器实现三种种子目级别的分类识别,确定最佳分类器并进行交叉验证。结果:SVM最高识别率为98.67%,BPNN最高识别率为94.33%,RF最高识别率为99.67%,RF交叉验证的识别率为99.93%。结论:通过对三种目级别的种子图像识别的结果表明:提出的在复频域和空间域中分别提取图像特征,并进行特征融合的方法可以显著提高三种种子目级别识别的准确率。In order to improve the classification accuracy of the seeds of Cyclocodon lancifolius,Campanumoea javanica and Campanumoea javanica subsp.javanica,a recognition algorithm based on multi-characteristic fusion in complex frequency domain and spatial domain was proposed.Firstly,stereoscopic microscopic images of the three kinds of seeds were taken.Secondly,the color features and directional gradient histogram features of the seed images were extracted in the spatial domain.Multi-scale and multi-direction complex frequency domain transformation of the images was carried out by using multi-directional dual tree complex wavelet transformation(M-DTCWT),and the shape features and texture features are extracted from the low-frequency sub-band images.Finally,ReliefF algorithm was used to fuse the spatial domain features and complex frequency domain features,and three classifiers(SVM,BPNN and RF)were used to realize the classification and recognition of the three kinds of seeds.The best classifier was determined and cross validation was conducted.The highest recognition rate of SVM was 98.67%,BPNN was 94.33%,RF was 99.67%,and RF cross-validation was 99.93%.The method could significantly improve the accuracy in the recognition of the three kinds of seeds.

关 键 词：轮钟花 金钱豹 大花金钱豹 种子图像 SVM BPNN RF 鉴别 

分 类 号：R282.5[医药卫生—中药学]