[CV] Mnist手写数字分类

数据集

手写数字Mnist数据集Mnist(lecun.com)

• 每张手写数字图片包含28*28个灰度像素点
• 包含从0~9十个数字

任务

使用SVM对手写数字进行分类

步骤

1. 对图像提取特征
2. 划分数据集，分为训练集和测试集7：3
3. 使用SVM对图像进行分类
4. 输出混淆矩阵和精确度

方法

1. 由于原图片只有28*28，直接使用像素作为特征送入SVM分类器

   # 直接对28*28像素进行svm分类
   from sklearn.metrics import accuracy_score, confusion_matrix
   import numpy as np
   from sklearn.model_selection import train_test_split
   from sklearn import svm
   import cv2
   from sklearn.datasets import fetch_openml
   # 获取数据集
   X, y = fetch_openml('mnist_784', data_home="./data",
                       as_frame=False, return_X_y=True)
   
   # 数据集的格式是每行是一张图片（一维数组numpy.dnarray总长度是28*28）
   print(X[0])
   assert X[0].shape[0] == 28*28
   
   
   X_train, X_test, y_train, y_test = train_test_split(
       X, y, test_size=0.3, random_state=0)
   
   svc = svm.SVC()
   svc.fit(X_train, y_train)
   
   predicts = svc.predict(X_test)
   
   # 打印出混淆矩阵
   
   print(confusion_matrix(predicts, y_test))
   
   print("accuracy : {}".format(accuracy_score(predicts, y_test)))
   

   运行结果：

   

   P.S. 直接送入分类器进行分类准确度就已经可以达到0.976了

2. 由之前学习的Sobel等梯度算子（边缘信息），同时为了旋转不变性就联想到数据增强——多角度提取边缘特征。基于梯度方向直方图（HOG)进行特征提取，再送入SVM分类器进行分类

   # 使用HOG特征提取
   from sklearn.metrics import accuracy_score, confusion_matrix
   import numpy as np
   from sklearn.model_selection import train_test_split
   from sklearn import svm
   import cv2
   from sklearn.datasets import fetch_openml
   # 获取数据集
   X, y = fetch_openml('mnist_784', data_home="./data",
                       as_frame=False, return_X_y=True)
   
   winSize = (4, 4)
   blockSize = (4, 4)
   blockStride = (8, 8)
   cellSize = (4, 4)
   nbins = 9
   derivAperture = 1
   winSigma = 4.
   histogramNormType = 0
   L2HysThreshold = 2.0000000000000001e-01
   gammaCorrection = 0
   nlevels = 64
   
   hog = cv2.HOGDescriptor(winSize, blockSize, blockStride, cellSize, nbins, derivAperture,
                               winSigma, histogramNormType, L2HysThreshold, gammaCorrection, nlevels)
   
   print(hog.compute(X[0].reshape(28,28).astype(np.uint8)).flatten().shape)
   
   # 计算提取HOG特征
   new_X = np.asarray([hog.compute(x.reshape(28,28).astype(np.uint8)).flatten() for x in X])
   
   X_train, X_test, y_train, y_test = train_test_split(
       new_X, y, test_size=0.3, random_state=0)
   
   svc = svm.SVC()
   svc.fit(X_train, y_train)
   
   predicts = svc.predict(X_test)
   
   # 打印出混淆矩阵
   
   print(confusion_matrix(predicts, y_test))
   
   print("accuracy : {}".format(accuracy_score(predicts, y_test)))
   

   

   使用HOG提取特征后，特征由28*28维度降到441维度，同时准确率也提升接近0.01。