簡(jiǎn)單的SVM支持向量機(jī)源代碼注釋
支持向量機(jī)
CV2庫：
https://www.lfd.uci.edu/~gohlke/pythonlibs/
# coding: utf-8
"""
支持向量機(jī)
用于區(qū)分手寫字母abc
"""
from scipy.misc import imread
import CV2
import numpy as np
from sklearn import svm
import matplotlib.pyplot as plt
def boundaries(binarized, axis): # 2值化后的大數(shù)組 坐標(biāo)軸
"""
輸入二值化的數(shù)組和坐標(biāo)
輸出分割的邊界
"""
# 在separate時(shí)用到
# variables named assuming axis = 0; algorithm valid for axis=1
# [1,0][axis] effectively swaps axes for summing
rows = np.sum(binarized, axis=[1, 0][axis]) > 0 # 在這個(gè)坐標(biāo)軸上求和 1*3013
rows[1:] = np.logical_xor(rows[1:], rows[:-1]) # 邊緣檢測(cè)？當(dāng)前和前一個(gè)
change = np.nonzero(rows)[0] # 記錄第幾列[28,60,169,202,297,326……] 有1 3 5 7 沒2 4 6 8
ymin = change[::2]
ymax = change[1::2]
height = ymax-ymin # 兩個(gè)邊緣的距離
too_small = 10 # 10像素以下不考慮了
ymin = ymin[height > too_small] # 正常數(shù)據(jù)
ymax = ymax[height > too_small] # 正常數(shù)據(jù)
return zip(ymin, ymax)
def separate(abc_list):
orig_img = abc_list.copy()
pure_white = 255.
white = np.max(abc_list)
black = np.min(abc_list)
thresh = (white+black)/2.0 # （最大值+最小值）/2.0 ？？？
binarized = abc_list < thresh # 用中值區(qū)分 二值化的大數(shù)組
row_bounds = boundaries(binarized, axis=0)
cropped = []
for top, bottom in row_bounds: # 打包的ymin ymax每次取一對(duì)
abc_list = binarized[top:bottom, :] # 上下夾
left, right = zip(*boundaries(abc_list, axis=1)) # 左右夾
rects = [top, bottom, left[0], right[0]] # 上下左右 在大數(shù)組中的位置
# print('上下左右=', top, bottom, left[0], right[0])
cropped.append(np.array(orig_img[rects[0]:rects[1], rects[2]:rects[3]] / pure_white))
return cropped
def partition(data_set, label_set, training_ratio):
"""
輸入 全部數(shù)據(jù) 全部標(biāo)號(hào) 訓(xùn)練比例（0~1）
輸出 訓(xùn)練數(shù)據(jù)集 訓(xùn)練標(biāo)號(hào) 測(cè)試數(shù)據(jù)集 測(cè)試標(biāo)號(hào)
"""
label_num = len(np.unique(label_set)) # 這里3
sum_num = int(len(data_set)/label_num)
train_num = int(sum_num * training_ratio)
test_num = int(sum_num - train_num)
# _train_label = np.chararray(int(label_num*train_num))
# _test_label = np.chararray(int(label_num*test_num))
_train_label = np.empty(int(label_num * train_num), dtype=label_set.dtype)
_test_label = np.empty(int(label_num*test_num), dtype=label_set.dtype)
_train_data = np.empty([len(_train_label), len(data_set[1])])
_test_data = np.empty([len(_test_label), len(data_set[1])])
for _target in np.arange(label_num): # 對(duì)每組標(biāo)號(hào)
for _i in np.arange(sum_num): # 對(duì)每組中的每個(gè)數(shù)據(jù)
if _i < train_num: # 在約定的訓(xùn)練集大小之內(nèi)
_train_label[_i+(_target*train_num-1)] = label_set[_target * sum_num]
_train_data[_i+(_target*train_num-1)] = data_set[_i+(_target*sum_num-1)]
else: # 在約定的測(cè)試集大小之內(nèi)
_test_label[(_target*test_num)+sum_num-_i-1] = label_set[_target * sum_num]
_test_data[(_target*test_num)+sum_num-_i-1] = data_set[_i+(_target*sum_num-1)]
return _train_data, _train_label, _test_data, _test_label
def main():
"""主程序 參數(shù)"""
visualize = True
training_percent = .15
resized_w = 5
resized_h = 5
letters = list()
for i in range(ord('a'), ord('z')+1):
letters.append(chr(i))
""" 讀入手寫圖像 """
original_pic = list()
for i in range(3):
original_pic.append(imread(letters[i]+".png", flatten=True))
""" 格式化圖片 """
# 分割圖像 每個(gè)字符上下左右夾后重整到10*10
separated_pic = list()
resized_pic = list()
data = resized_pic
for i in range(len(original_pic)):
separated_pic.append(separate(original_pic[i]))
for img in separated_pic[i]:
resized_pic.append(CV2.resize(img, (resized_w, resized_h)))
""" 整理數(shù)據(jù)和標(biāo)號(hào) """
data = np.asarray(data)
data = data.reshape((-1, resized_w*resized_h)) # data[0:68]=一個(gè)10*10的數(shù)組，筆跡處數(shù)值較小0.5，空白處數(shù)值較大0.9
# 全數(shù)據(jù)數(shù)組 data[第幾個(gè)圖片, 10, 10]
label = np.empty(len(data), dtype=int)
for i in range(len(data)):
if i <= 22:
label[i] = ord(letters[0])
elif 23 <= i <= 45:
label[i] = ord(letters[1])
elif 46 <= i <= 68:
label[i] = ord(letters[2])
train_data, train_label, test_data, test_label = partition(data, label, training_percent)
""" 訓(xùn)練分類器 和 測(cè)試其正確率 """
# clf = svm.LinearSVC()
# http://scikit-learn.org/stable/modules/svm.html
clf = svm.SVC(C=1.0, kernel='rbf', gamma='auto') # C懲罰量 kernel核函數(shù)選擇(rbf=exp(-gamma|u-v|^2)) gamma核函數(shù)參數(shù)
clf.fit(train_data, train_label) # 訓(xùn)練集 訓(xùn)練
predict = clf.predict(test_data) # 用測(cè)試集 測(cè)試
correct = len(predict) # 測(cè)試集大小 全對(duì)初始化...
for i in range(len(predict)): # 對(duì)測(cè)試集長(zhǎng)度
if predict[i] != test_label[i]: # 預(yù)測(cè)結(jié)果和測(cè)試集標(biāo)號(hào)不同
correct -= 1
""" 打印結(jié)果 """
for i in range(len(predict)):
print(i + 1, chr(test_label[i]), '->', chr(predict[i]))
print("總數(shù)據(jù)量: %d類標(biāo)號(hào) * %d個(gè)樣本" % (len(np.unique(label)), int(len(data) / len(np.unique(label)))))
train_data_num = int(int(len(data) / len(np.unique(label))) * training_percent)
test_data_num = int(int(len(data) / len(np.unique(label))) - train_data_num)
print("訓(xùn)練數(shù)據(jù)集 %d, " % train_data_num, end='')
print("測(cè)試數(shù)據(jù)集 %d" % test_data_num, end='')
print("(訓(xùn)練數(shù)據(jù)占總數(shù)據(jù){0}%)".format(training_percent * 100))
print("測(cè)試集正確率{0}%".format(100 * correct / len(predict)))
""" 可視化 """
if visualize:
test_data = test_data.reshape(-1, resized_h, resized_w)
images_and_labels = list(zip(test_data, predict))
for index, (image, label) in enumerate(images_and_labels[::2]):
plt.subplot(6, 5, index + 1)
plt.axis('off')
plt.imshow(image, cmap='gray', interpolation='nearest')
plt.title('predict:%s' % chr(label))
plt.show()
if __name__ == '__main__':
main()
