You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

232 lines
6.6 KiB
Python

#!/usr/bin/env python
# vim:tabstop=4 shiftwidth=4 tw=79:
'''
SVM, Random forest and KNearest digit recognition.
Modified from the OpenCV example.
Sample loads a dataset of handwritten digits from '../data/digits.png'. Then
it trains a Random Forest, SVM and KNearest classifiers on it and evaluates
their accuracy.
Following preprocessing is applied to the dataset:
- Moment-based image deskew (see deskew())
- Digit images are split into 4 10x10 cells and 16-bin histogram of oriented
gradients is computed for each cell
- Transform histograms to space with Hellinger metric (see [1] (RootSIFT))
Or in the "simple setting":
- Only moment-based image deskew (see deskew())
[1] R. Arandjelovic, A. Zisserman
"Three things everyone should know to improve object retrieval"
http://www.robots.ox.ac.uk/~vgg/publications/2012/Arandjelovic12/arandjelovic12.pdf
Usage:
digits.py
'''
import cv2
import numpy as np
from numpy.linalg import norm
# local modules
from common import mosaic
SZ = 20 # size of each digit is SZ x SZ
CLASS_N = 10
DIGITS_FN = 'digits.png'
SIMPLE = True # Use simple preprocessing or HOG features (for SVM)
def split2d(img, cell_size, flatten=True):
h, w = img.shape[:2]
sx, sy = cell_size
cells = [np.hsplit(row, w//sx) for row in np.vsplit(img, h//sy)]
cells = np.array(cells)
if flatten:
cells = cells.reshape(-1, sy, sx)
return cells
def load_digits(fn):
print 'loading "%s" ...' % fn
digits_img = cv2.imread(fn, 0)
digits = split2d(digits_img, (SZ, SZ))
labels = np.repeat(np.arange(CLASS_N), len(digits)/CLASS_N)
return digits, labels
def deskew(img):
m = cv2.moments(img)
if abs(m['mu02']) < 1e-2:
return img.copy()
skew = m['mu11']/m['mu02']
M = np.float32([[1, skew, -0.5*SZ*skew], [0, 1, 0]])
img = cv2.warpAffine(img, M, (SZ, SZ),
flags=cv2.WARP_INVERSE_MAP | cv2.INTER_LINEAR)
return img
class StatModel(object):
def load(self, fn):
self.model.load(fn)
def save(self, fn):
self.model.save(fn)
class KNearest(StatModel):
def __init__(self, k=3):
self.k = k
self.model = cv2.KNearest()
def train(self, samples, responses):
self.model = cv2.KNearest()
self.model.train(samples, responses)
def predict(self, samples):
retval, results, neigh_resp, dists = self.model.find_nearest(samples,
self.k)
return results.ravel()
class SVM(StatModel):
def __init__(self, kernel_type=cv2.SVM_RBF, C=1, gamma=0.5):
self.params = dict(kernel_type=kernel_type,
svm_type=cv2.SVM_C_SVC,
C=C,
gamma=gamma)
self.model = cv2.SVM()
def train(self, samples, responses):
self.model = cv2.SVM()
self.model.train(samples, responses, params=self.params)
def train_auto(self, samples, responses):
self.model = cv2.SVM()
self.model.train_auto(samples, responses, None, None,
params=self.params)
def predict(self, samples):
return self.model.predict_all(samples).ravel()
class RForest(StatModel):
def __init__(self):
self.params = dict(max_depth=20)
self.model = cv2.RTrees()
def train(self, samples, responses):
self.model = cv2.RTrees()
self.model.train(samples, cv2.CV_ROW_SAMPLE, responses,
params=self.params)
def predict(self, samples):
predictions = map(self.model.predict, samples)
return predictions
def evaluate_model(model, digits, samples, labels):
resp = model.predict(samples)
err = (labels != resp).mean()
print 'error: %.2f %%' % (err*100)
confusion = np.zeros((10, 10), np.int32)
for i, j in zip(labels, resp):
confusion[i, j] += 1
print 'confusion matrix:'
print confusion
print
vis = []
for img, flag in zip(digits, resp == labels):
img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
if not flag:
img[..., :2] = 0
vis.append(img)
return mosaic(25, vis)
def preprocess_simple(digits):
return np.float32(digits).reshape(-1, SZ*SZ) / 255.0
def preprocess_hog(digits):
samples = []
for img in digits:
gx = cv2.Sobel(img, cv2.CV_32F, 1, 0)
gy = cv2.Sobel(img, cv2.CV_32F, 0, 1)
mag, ang = cv2.cartToPolar(gx, gy)
bin_n = 16
bin = np.int32(bin_n*ang/(2*np.pi))
bin_cells = bin[:10, :10], bin[10:, :10], bin[:10, 10:], bin[10:, 10:]
mag_cells = mag[:10, :10], mag[10:, :10], mag[:10, 10:], mag[10:, 10:]
hists = [np.bincount(b.ravel(), m.ravel(), bin_n)
for b, m in zip(bin_cells, mag_cells)]
hist = np.hstack(hists)
# transform to Hellinger kernel
eps = 1e-7
hist /= hist.sum() + eps
hist = np.sqrt(hist)
hist /= norm(hist) + eps
samples.append(hist)
return np.float32(samples)
if __name__ == '__main__':
print __doc__
digits, labels = load_digits(DIGITS_FN)
print 'preprocessing...'
# shuffle digits
rand = np.random.RandomState(321)
shuffle = rand.permutation(len(digits))
digits, labels = digits[shuffle], labels[shuffle]
digits2 = map(deskew, digits)
if SIMPLE:
samples = preprocess_simple(digits2)
else:
samples = preprocess_hog(digits2)
train_n = int(0.9*len(samples))
cv2.imshow('test set', mosaic(25, digits[train_n:]))
digits_train, digits_test = np.split(digits2, [train_n])
samples_train, samples_test = np.split(samples, [train_n])
labels_train, labels_test = np.split(labels, [train_n])
print 'training Random Forest...'
model = RForest()
model.train(samples_train, labels_train)
vis = evaluate_model(model, digits_test, samples_test, labels_test)
cv2.imshow('Random Forest test', vis)
print 'training KNearest...'
model = KNearest(k=4)
model.train(samples_train, labels_train)
vis = evaluate_model(model, digits_test, samples_test, labels_test)
cv2.imshow('KNearest test', vis)
print 'training SVM...'
if SIMPLE:
model = SVM(kernel_type=cv2.SVM_LINEAR, C=0.1)
model.train(samples_train, labels_train)
else:
model = SVM(kernel_type=cv2.SVM_RBF, C=2.67, gamma=5.383)
model.train(samples_train, labels_train)
vis = evaluate_model(model, digits_test, samples_test, labels_test)
cv2.imshow('SVM test', vis)
print 'saving SVM as "digits_svm.dat"...'
model.save('digits_svm.dat')
cv2.waitKey(0)