1 简介
字符特征提取完成后，接下来的主要任务就是识别字符。传统的机器学习将这一任务转换为一个分类任务。针对该任务，一系列的分类方法模型出现，主要包括：支持向量机、近邻算法、多层感知器等。
2 支持向量机
SVM原理请移步该文：
SVM字符分类代码如下：
# -*- coding: UTF-8 -*- import cv2 import os import numpy as np from sklearn import svm from PIL import Image # 垂直投影 def verticle_projection(thresh1): (h, w) = thresh1.shape a = [0 for z in range(0, w)] # 记录每一列的波峰 # 遍历一列 for j in range(0, w): # 遍历一行 for i in range(0, h): # 如果改点为黑点 if thresh1[i, j] == 0: # 该列的计数器加一计数 a[j] += 1 # 记录完后将其变为白色 thresh1[i, j] = 255 # 遍历每一列 for j in range(0, w): # 从该列应该变黑的最顶部的点开始向最底部涂黑 for i in range((h - a[j]), h): # 涂黑 thresh1[i, j] = 0 # 存储所有分割出来的图片 roi_list = list() start_index = 0 end_index = 0 in_block = False for i in range(0, w): if in_block == False and a[i] != 0: in_block = True start_index = i elif a[i] == 0 and in_block: end_index = i in_block = False roiImg = thresh1[0:h, start_index:end_index + 1] roi_list.append(roiImg) return roi_list # 将二值化后的数组转化成网格特征统计图 def get_features(array): # 拿到数组的高度和宽度 h, w = array.shape data = [] for x in range(0, w / 4): offset_y = x * 4 temp = [] for y in range(0, h / 4): offset_x = y * 4 # 统计每个区域的1的值 sum_temp = array[0 + offset_y:4 + offset_y, 0 + offset_x:4 + offset_x] temp.append(sum(sum(sum_temp))) data.append(temp) return np.asarray(data) def train_main(): # 读取训练样本 train_path = "../dataset/train/" train_files = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9'] train_X = [] train_y = [] for train_file in train_files: pictures = os.listdir(train_path + train_file) for picture in pictures: img = cv2.imread(train_path + train_file + "/" + picture) img = cv2.resize(img, (32, 32)) gray_image = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) ret, thresh1 = cv2.threshold( gray_image, 130, 255, cv2.THRESH_BINARY) feature = get_features(thresh1) feature = feature.reshape(feature.shape[0] * feature.shape[1]) train_X.append(feature) train_y.append(train_file) train_X = np.array(train_X) train_y = np.array(train_y) linearsvc_clf = svm.LinearSVC() linearsvc_clf.fit(train_X, train_y) return linearsvc_clf def test_main(linearsvc_clf): # 原图 img = cv2.imread("../dataset/test/idcard1.jpg") # 灰度图 gray_image = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) # 二值图 ret, thresh1 = cv2.threshold(gray_image, 130, 255, cv2.THRESH_BINARY) # 垂直投影 roi_list = verticle_projection(thresh1) test_X = [] # 输入分类器 for single in roi_list: single = cv2.resize(single, (32, 32), interpolation=cv2.INTER_CUBIC) feature = get_features(single) feature = feature.reshape(feature.shape[0] * feature.shape[1]) test_X.append(feature) test_X = np.array(test_X) result = linearsvc_clf.predict(test_X) print(result) if __name__ == '__main__': linearsvc_clf = train_main() test_main(linearsvc_clf)
3 近邻算法
近邻算法原理请移步该文：
近邻算法字符分类代码如下：
# -*- coding: UTF-8 -*- import cv2 import os import numpy as np from sklearn import neighbors from PIL import Image # 垂直投影 def verticle_projection(thresh1): (h, w) = thresh1.shape a = [0 for z in range(0, w)] # 记录每一列的波峰 # 遍历一列 for j in range(0, w): # 遍历一行 for i in range(0, h): # 如果改点为黑点 if thresh1[i, j] == 0: # 该列的计数器加一计数 a[j] += 1 # 记录完后将其变为白色 thresh1[i, j] = 255 # 遍历每一列 for j in range(0, w): # 从该列应该变黑的最顶部的点开始向最底部涂黑 for i in range((h - a[j]), h): # 涂黑 thresh1[i, j] = 0 # 存储所有分割出来的图片 roi_list = list() start_index = 0 end_index = 0 in_block = False for i in range(0, w): if in_block == False and a[i] != 0: in_block = True start_index = i elif a[i] == 0 and in_block: end_index = i in_block = False roiImg = thresh1[0:h, start_index:end_index + 1] roi_list.append(roiImg) return roi_list # 将二值化后的数组转化成网格特征统计图 def get_features(array): # 拿到数组的高度和宽度 h, w = array.shape data = [] for x in range(0, w / 4): offset_y = x * 4 temp = [] for y in range(0, h / 4): offset_x = y * 4 # 统计每个区域的1的值 sum_temp = array[0 + offset_y:4 + offset_y, 0 + offset_x:4 + offset_x] temp.append(sum(sum(sum_temp))) data.append(temp) return np.asarray(data) def train_main(): # 读取训练样本 train_path = "../dataset/train/" train_files = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9'] train_X = [] train_y = [] for train_file in train_files: pictures = os.listdir(train_path + train_file) for picture in pictures: img = cv2.imread(train_path + train_file + "/" + picture) img = cv2.resize(img, (32, 32)) gray_image = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) ret, thresh1 = cv2.threshold( gray_image, 130, 255, cv2.THRESH_BINARY) feature = get_features(thresh1) feature = feature.reshape(feature.shape[0] * feature.shape[1]) train_X.append(feature) train_y.append(train_file) train_X = np.array(train_X) train_y = np.array(train_y) knn_clf = neighbors.KNeighborsClassifier() knn_clf.fit(train_X, train_y) return knn_clf def test_main(knn_clf): # 原图 img = cv2.imread("../dataset/test/idcard1.jpg") # 灰度图 gray_image = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) # 二值图 ret, thresh1 = cv2.threshold(gray_image, 130, 255, cv2.THRESH_BINARY) # 垂直投影 roi_list = verticle_projection(thresh1) test_X = [] # 输入分类器 for single in roi_list: single = cv2.resize(single, (32, 32), interpolation=cv2.INTER_CUBIC) feature = get_features(single) feature = feature.reshape(feature.shape[0] * feature.shape[1]) test_X.append(feature) test_X = np.array(test_X) result = knn_clf.predict(test_X) print(result) if __name__ == '__main__': knn_clf = train_main() test_main(knn_clf)
4 多层感知器
多层感知器是一种常见的人工神经网络模型，主要包括：输入层、隐含层和输出层三个部分。每个MLP模型均可包含一个或者多个隐含层。最简单的MLP是一种三层结构，如图所示：
可以看出，无论输入层和隐含层之间，还是隐含层和输出层之间都是全连接的。输入层，顾名思义，就是原始数据的输入，比如输入一个n维的向量，那么输入有n个神经元。隐含层是对输入层的数据进行一定的运算后得到的，具体公式如下：
其中，表示权重参数，表示偏置项，为激活函数，常用的有函数和函数。相应的公式分别为：
最后是输出层，输出层和隐含层之间可以看作是一个类似于多类别逻辑回归的映射关系，也就是回归。所以，输出层最终的输出可以表示为：
其中，为上面提到的隐含层的输出，为权重参数，为偏置项。将上述的公式整合，可以得到上述三层MLP最终的输出层输出公式：
近邻算法字符分类代码如下：
# -*- coding: UTF-8 -*- import cv2 import os import numpy as np from sklearn.neural_network import MLPClassifier from PIL import Image # 垂直投影 def verticle_projection(thresh1): (h, w) = thresh1.shape a = [0 for z in range(0, w)] # 记录每一列的波峰 # 遍历一列 for j in range(0, w): # 遍历一行 for i in range(0, h): # 如果改点为黑点 if thresh1[i, j] == 0: # 该列的计数器加一计数 a[j] += 1 # 记录完后将其变为白色 thresh1[i, j] = 255 # 遍历每一列 for j in range(0, w): # 从该列应该变黑的最顶部的点开始向最底部涂黑 for i in range((h - a[j]), h): # 涂黑 thresh1[i, j] = 0 # 存储所有分割出来的图片 roi_list = list() start_index = 0 end_index = 0 in_block = False for i in range(0, w): if in_block == False and a[i] != 0: in_block = True start_index = i elif a[i] == 0 and in_block: end_index = i in_block = False roiImg = thresh1[0:h, start_index:end_index + 1] roi_list.append(roiImg) return roi_list # 将二值化后的数组转化成网格特征统计图 def get_features(array): # 拿到数组的高度和宽度 h, w = array.shape data = [] for x in range(0, w / 4): offset_y = x * 4 temp = [] for y in range(0, h / 4): offset_x = y * 4 # 统计每个区域的1的值 sum_temp = array[0 + offset_y:4 + offset_y, 0 + offset_x:4 + offset_x] temp.append(sum(sum(sum_temp))) data.append(temp) return np.asarray(data) def train_main(): # 读取训练样本 train_path = "../dataset/train/" train_files = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9'] train_X = [] train_y = [] for train_file in train_files: pictures = os.listdir(train_path + train_file) for picture in pictures: img = cv2.imread(train_path + train_file + "/" + picture) img = cv2.resize(img, (32, 32)) gray_image = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) ret, thresh1 = cv2.threshold(gray_image, 130, 255, cv2.THRESH_BINARY) feature = get_features(thresh1) feature = feature.reshape(feature.shape[0] * feature.shape[1]) train_X.append(feature) train_y.append(train_file) train_X = np.array(train_X) train_y = np.array(train_y) mlp_clf = MLPClassifier(solver='lbfgs', alpha=1e-5, hidden_layer_sizes=(32, 32), random_state=1) mlp_clf.fit(train_X, train_y) return mlp_clf def test_main(mlp_clf): # 原图 img = cv2.imread("../dataset/test/idcard1.jpg") # 灰度图 gray_image = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) # 二值图 ret, thresh1 = cv2.threshold(gray_image, 130, 255, cv2.THRESH_BINARY) # 垂直投影 roi_list = verticle_projection(thresh1) test_X = [] # 输入分类器 for single in roi_list: single = cv2.resize(single, (32, 32), interpolation=cv2.INTER_CUBIC) feature = get_features(single) feature = feature.reshape(feature.shape[0] * feature.shape[1]) test_X.append(feature) test_X = np.array(test_X) result = mlp_clf.predict(test_X) print(result) if __name__ == '__main__': mlp_clf = train_main() test_main(mlp_clf)
![深度学习应用篇-计算机视觉-OCR光学字符识别[7]：OCR综述、常用CRNN识别方法、DBNet、CTPN检测方法等、评估指标、应用场景](https://ucc.alicdn.com/fnj5anauszhew_20230611_eba1a5baa8bf41b08e8306e0af20dbcd.png?x-oss-process=image/resize,h_160,m_lfit)