基于机器学习的字体辅助选择系统.zip

#!/usr/local/bin/python import cv2 import numpy from sklearn import cross_validation from sklearn import svm import matplotlib.pyplot as plt from sklearn import metrics from sklearn.externals import joblib MIN_DESCRIPTOR = 10 TRAINING_SIZE = 2 def findDescriptor(img): contour = [] _, contour, hierarchy = cv2.findContours( img, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE, contour) contour_array = contour[0][:, 0, :] contour_complex = numpy.empty(contour_array.shape[:-1], dtype=complex) contour_complex.real = contour_array[:, 0] contour_complex.imag = contour_array[:, 1] fourier_result = numpy.fft.fft(contour_complex) return fourier_result def truncate_descriptor(descriptors, degree): descriptors = numpy.fft.fftshift(descriptors) center_index = len(descriptors) / 2 descriptors = descriptors[ center_index - degree / 2:center_index + degree / 2] descriptors = numpy.fft.ifftshift(descriptors) return descriptors def reconstruct(descriptors, degree): descriptor_in_use = truncate_descriptor(descriptors, degree) contour_reconstruct = numpy.fft.ifft(descriptor_in_use) contour_reconstruct = numpy.array( [contour_reconstruct.real, contour_reconstruct.imag]) contour_reconstruct = numpy.transpose(contour_reconstruct) contour_reconstruct = numpy.expand_dims(contour_reconstruct, axis=1) if contour_reconstruct.min() < 0: contour_reconstruct -= contour_reconstruct.min() contour_reconstruct *= 800 / contour_reconstruct.max() contour_reconstruct = contour_reconstruct.astype(numpy.int32, copy=False) black = numpy.zeros((800, 800), numpy.uint8) cv2.drawContours(black, contour_reconstruct, -1, 255, thickness=5) return descriptor_in_use def sample_generater(sample1): response = numpy.array([0, 1]) response = numpy.tile(response, TRAINING_SIZE / 2) response = response.astype(numpy.float32) training_set = numpy.empty( [TRAINING_SIZE, MIN_DESCRIPTOR], dtype=numpy.float32) # assign descriptors with noise to our training_set for i in range(0, TRAINING_SIZE - 1, 2): descriptors_sample1 = findDescriptor(sample1) descriptors_sample1 = truncate_descriptor( descriptors_sample1, MIN_DESCRIPTOR) return training_set, response def getFourier(sample1): _, sample1 = cv2.threshold(sample1, 127, 255, cv2.THRESH_BINARY_INV) training_set, response = sample_generater() fourier_result = findDescriptor(sample1) contour_reconstruct = reconstruct(fourier_result, MIN_DESCRIPTOR) return contour_reconstruct def getFeatures (filename): img = cv2.imread(filename) gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) corners = cv2.goodFeaturesToTrack(gray, 3, 0.7, 40) corners = numpy.int0(corners) descriptor = numpy.zeros([3, 10]) j = 0 for i in corners: x, y = i.ravel() imgfeature = gray[x-20:x+20, y-20:y+20] descriptor[j] = getFourier(imgfeature) j += 1 return descriptor for i in range(1, 6): for j in range(1, 6): filename = "Bbs/" + str(i)+"_"+str(j)+".jpg" sample = cv2.imread(filename, 0) print(filename) if i == 1 and j == 1: array = getFourier(sample) print(array) else: print(getFourier(sample)) array = numpy.row_stack((array, getFourier(sample))) Y = [1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5 ] clf = svm.SVC(gamma=0.01) clf.fit(array, Y) joblib.dump(clf, "Bbs.m") # sample = cv2.imread('qw.jpg', 0) # X_test = getFourier(sample) # desc = getFeatures('wdj.jpg') # print desc[1] # predicted = clf.predict(X_test) # print(predicted)