python - 性能:扩大二值图像(形态膨胀)

Question

我编写了一个 Cython 函数，它将二进制图像(numpy 数组)扩展 1 个像素。所以我只想扩大数组值为1的区域。这是我的天真的方法:

def expand_1px (numpy.ndarray[numpy.uint8_t, ndim = 2] A):
    cdef int  h = A.shape[0]
    cdef int  w = A.shape[1]
    cdef numpy.ndarray[numpy.uint8_t, ndim = 2] RES = numpy.zeros([h, w], dtype = numpy.uint8)
    # These Two lines below were originally missing 
    cdef int  y, x
    cdef unsigned char  prev, cur

    for x in range (0, w):
        for y in range (1, h):
            prev = A[y-1,x]
            cur = A[y,x]
            if cur > prev:
                RES[y-1, x] = 1
            if cur < prev:
                RES[y,x] = 1
    for y in range (0, h):
        for x in range (1, w):
            prev = A[y,x-1]
            cur = A[y,x]
            if cur > prev:
                RES[y, x-1] = 1
            if cur < prev:
                RES[y,x] = 1
    return numpy.bitwise_or(A,RES)

这可以正常工作，但是非常慢。OpenCV 函数 dilate() 比我的 Cython 变体快 1000 ~30 倍，并给出相同的结果。我这样使用它:

kernel = numpy.ones((3,3), dtype="uint8")
kernel[0,0] = 0
kernel[2,2] = 0
kernel[0,2] = 0
kernel[2,0] = 0
...
IMG = cv2.dilate(IMG,kernel,iterations = 1)

<小时/>

问:

• OpenCV 的变种怎么可以这么快？它实际上有什么作用？
• 如何让我的 Cython 函数运行得这么快？

<小时/>

更新:

如此糟糕的性能是由于缺少“cdef”声明，我的错。将其添加到函数中会产生不同:

cdef int  y, x
cdef unsigned char  prev, cur

性能差异仍然是大约 30 倍，这也有点令人失望。有进一步改进的建议吗？

Answer 1

您的函数存在一个大问题:您正在执行所需的更多循环。

import cython

@wraparound(false)
def expand_1px(numpy.ndarray[numpy.uint8_t, ndim = 2] A):
    cdef int  h = A.shape[0]
    cdef int  w = A.shape[1]
    cdef numpy.ndarray[numpy.uint8_t, ndim = 2] RES = numpy.zeros([h, w], dtype = numpy.uint8)
    cdef int  y, x
    cdef unsigned char  prev, cur

    for x in range(1, w):
        for y in range(1, h):
            cur = A[y,x]

            prev = A[y-1,x]
            if prev < cur:
                RES[y-1,x] = 1
            elif cur < prev:
                RES[y,x] = 1

            prev = A[y, x-1]
            if prev < cur:
                RES[y,x-1] = 1
            elif cur < prev:
                RES[y,x] = 1

    for x in range(1, w):
        cur = A[0,x]
        prev = A[0,x-1]
        if prev < cur:
            RES[0,x-1] = 1
        elif cur < prev:
            RES[0,x] = 1

    for y in range(1, h):
        cur = A[y,0]
        prev = A[y-1,0]
        if prev < cur:
            RES[y-1,0] = 1
        elif cur < prev:
            RES[y,0] = 1

    return numpy.bitwise_or(A,RES)

最初，您进行 2wh 次迭代，通过此更改，您将只进行 wh + w + h 次迭代。