使用 OpenCV detectMultiScale 加载 Numpy 数组二进制图像

Question

我有最初来自我用 OpenCV 操作的 Mnist 数据集的数字序列。它们保存在 pickle 文件中。它们是形状为 (112, 112) 的 1 通道图像。

我想通过 OpenCV 级联分类器运行这些，类似于 the face detection tutorial，但使用这些 Google-Street-View-House-Numbers-Digit-Localization cascades。

这是我的尝试：

data = p.load_file('data/train_sequences00.pickle')
zero = cv2.CascadeClassifier('data/Google-Street-View-House-Numbers-Digit-Localization/cascades/cascade0/cascade.xml')
one = cv2.CascadeClassifier('data/Google-Street-View-House-Numbers-Digit-Localization/cascades/cascade1/cascade.xml')
two = cv2.CascadeClassifier('data/Google-Street-View-House-Numbers-Digit-Localization/cascades/cascade2/cascade.xml')
three = cv2.CascadeClassifier('data/Google-Street-View-House-Numbers-Digit-Localization/cascades/cascade3/cascade.xml')
four = cv2.CascadeClassifier('data/Google-Street-View-House-Numbers-Digit-Localization/cascades/cascade4/cascade.xml')
five = cv2.CascadeClassifier('data/Google-Street-View-House-Numbers-Digit-Localization/cascades/cascade5/cascade.xml')
six = cv2.CascadeClassifier('data/Google-Street-View-House-Numbers-Digit-Localization/cascades/cascade6/cascade.xml')
seven = cv2.CascadeClassifier('data/Google-Street-View-House-Numbers-Digit-Localization/cascades/cascade7/cascade.xml')
eight = cv2.CascadeClassifier('data/Google-Street-View-House-Numbers-Digit-Localization/cascades/cascade8/cascade.xml')
nine = cv2.CascadeClassifier('data/Google-Street-View-House-Numbers-Digit-Localization/cascades/cascade9/cascade.xml')
gray = np.array(data['sequences'][0]).astype(np.float32)
gray = cv2.cvtColor(gray, cv2.COLOR_GRAY2BGR)
#gray = cv2.cvtColor(img, cv2.CV_RGB2GRAY)

zeros = zero.detectMultiScale(gray, 1.3, 5, scaleFactor=0.6)
ones = one.detectMultiScale(gray, 1.3, 5, scaleFactor=0.6)
twos = two.detectMultiScale(gray, 1.3, 5, scaleFactor=0.6)
threes = three.detectMultiScale(gray, 1.3, 5, scaleFactor=0.6)
fours = four.detectMultiScale(gray, 1.3, 5, scaleFactor=0.6)
fives = five.detectMultiScale(gray, 1.3, 5, scaleFactor=0.6)
sixes = size.detectMultiScale(gray, 1.3, 5, scaleFactor=0.6)
sevens = seven.detectMultiScale(gray, 1.3, 5, scaleFactor=0.6)
eights = eight.detectMultiScale(gray, 1.3, 5, scaleFactor=0.6)
nines = nine.detectMultiScale(gray, 1.3, 5, scaleFactor=0.6)

不幸的是，这只会导致以下错误：

Traceback (most recent call last):
  File "digit_cascade.py", line 22, in <module>
    zeros = zero.detectMultiScale(gray, 1.3, 5, scaleFactor=0.6)
SystemError: error return without exception set

有人熟悉如何将原始 numpy 数组加载到 OpenCV detectMultiscale 中吗？

以下是我的一些不确定性：

• OpenCV docs 有两个不同的detectMultiscale 签名。第一个的目的是什么，带有rejectLevels 和levelWeights 的参数？
• 我的图像是 112x112。此图像中的每个数字大约为 11x11 像素。级联尺寸为 20x30。我应该/我需要设置什么参数才能使级联正常工作？

系统信息：

cv2.__version__ # '2.4.11'
sys.version #'2.7.12 |Anaconda 4.2.0 (x86_64)| (default, Jul  2 2016, 17:43:17) \n[GCC 4.2.1 (Based on Apple Inc. build 5658) (LLVM build 2336.11.00)]'

Answer 1

我可以使用相同的级联文件运行分类器，同时直接从图像加载数据：

import numpy as np
import cv2

workspace = '/path/to/directory'

# cascades
cascades = [
    (cv2.CascadeClassifier(workspace + '/Google-Street-View-House-Numbers-Digit-Localization/cascades/cascade1/cascade.xml'), (255, 0, 0)),
    (cv2.CascadeClassifier(workspace + '/Google-Street-View-House-Numbers-Digit-Localization/cascades/cascade6/cascade.xml'), (0, 255, 0))
]

# input
img = cv2.imread('/path/to/image/house-number-16.jpg')
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

# detect and mark
for cascade, color in cascades:
    digits = cascade.detectMultiScale(gray, 1.3, 5)
    for (x, y, w, h) in digits:
        cv2.rectangle(img, (x, y), (x + w, y + h), color, 2)
        roi_gray = gray[y:y + h, x:x + w]
        roi_color = img[y:y + h, x:x + w]

# show result
cv2.imshow('img', img)
cv2.waitKey()

结果（虽然检测不到“1”）：

我怀疑您的问题在于您从 pickle 保存图像的方式或加载它们的方式。

例如，默认情况下cv2.imread 将图像读取为整数numpy.ndarray。当我将其更改为浮点数（例如除以 255）时，出现错误：

OpenCV 错误：断言失败 (scaleFactor > 1 && image.depth() == CV_8U) 在detectMultiScale，文件 /home/yohanesgultom/opencv/modules/objdetect/src/cascadedetect.cpp， 第 1081 行 Traceback（最近一次调用最后一次）：文件“cascade.py”，行 21，在 数字 = cascade.detectMultiScale(gray, 1.3, 5) cv2.error: /home/yohanesgultom/opencv/modules/objdetect/src/cascadedetect.cpp:1081: 错误：(-215) scaleFactor > 1 && image.depth() == CV_8U in function 检测多尺度

我在 Ubuntu 上运行它：

>>> cv2.__version__
'2.4.13'
>>> sys.version
'2.7.10 (default, Oct 14 2015, 16:09:02) \n[GCC 5.2.1 20151010]'