摘要
将图像的像素特征与矩特征结合,构建了神经网络分类器,利用提取的特征向量对分类器进行了训练和测试。将图像二值化,并归一化为16*16大小,提取了其每个像素点的0、1特征共16*16=256维,图像的网格特征13维,及Hu矩特征7维,一共276维特征。建立了BP神经网络分类器,分别使用最速下降BP算法、动量BP算法、学习率可变BP算法对BP神经网络分类器进行了训练,得出了在相同条件下学习率可变BP算法训练时间短,收敛快的结论。建立了PNN神经网络分类器,与BP神经网络分类器性能进行比较,实验结果表明,PNN神经网络分类器性能更好。
Combining the image pixel feature and torque feature, this paper built neural network classifier, and extracts feature classifier for training and testing. The image which was binarization, was normalized to a size of 16 ＊ 16, and extracts its each pixel 0 1 feature, total of 16 ＊ 16 = 256d, 13 image grid features 13d, 7d Hu moment feature, total of 276d. Established the BP neural network classifier, the paper gets the results that vector under the condition of the same variable BP algorithm training time is short, fast convergent, using the steepest descent of BP algorithm and momentum BP algorithm, BP algorithm of vector variable, respectively, to train the BP neural network classifier. Then, the paper established PNN neural network classifier, and it was compared with BP neural network classifier performance. At last, the paper concluded that PNN neural network classifier in the process of the experiment show the better performance.
出处
《软件导刊》
2014年第2期58-60,共3页
Software Guide
关键词
神经网络
数字识别
特征提取
Feature Extraction
Neural Network
Identify
