传统的茶叶茶梗分选方法在特征选取方面存在着样本颜色特征提取单一的问题，以及现有的茶叶茶梗分类器普遍存在分类精度低、耗费时间长等问题。针对CCD相机采集的茶叶茶梗的数字图像，首先经过二值化、开运算、闭运算、样本图像去噪、图像分割等预处理过程，再根据茶叶茶梗样本形态学特征的差异，提取出圆形度、矩形度、延伸率、Hu二阶不变矩、最大内切圆与其面积比等5类区分度大、独立性好的特征，作为BP神经网络分类器的输入向量，并采用L-M( Leven-berg-Marquardt)学习算法对传统的BP神经网络分类器进行优化，用于茶叶茶梗的分类。实验和仿真结果表明，经过L-M算法优化的BP网络分类器对茶叶茶梗样本的分类精度高达95%，且耗时相对较少，是一种有效的茶叶茶梗分类方法。%Traditional tea and tea-stalk sorting method exists problems that color feature extraction for sample is single in feature extrac-tion aspect and general classifier has low precision and large time consuming. In term of digital image of tea and tea stems collected by CCD camera,according to different shape features between them,firstly after binarization,open and close operation,sample image denois-ing,image segmentation and other pre-processing process,it extracts circularity,rectangularity,extensibility,Hu second-order moment invariants,and the ratio of maximum inscribed circle and its area,etc in this paper,which has great distinction and independence,as the input vector of BP ( Back-Propagation) neural network. It also applies L-M ( Levenberg-Marquardt) learning algorithm to optimize the traditional BP neural network for the classification of tea and tea stalk. Experiment and simulation results proves that the BP network clas-sifier optimized by L-M algorithm is as high as 98% on classification accuracy for tea and tea-stalk,and has relatively few time-consu-ming. It is an effective classification method of tea and tea-stalk.
展开▼
