时间戳

首页 > 解决方案 > 比使用 np.where() 在 Python 中的二维数组中查找指定值的所有遇到的索引更快的方法

python - 比使用 np.where() 在 Python 中的二维数组中查找指定值的所有遇到的索引更快的方法

问题描述

我正在开发一个 OpenCV 程序来计算视图中红色对象的中心。

这是我的代码片段，它计算视图中所有红色对象的“质心”。

red_only_array = np.array(red_only)
locations = np.where(red_only==255)
x_avg = np.mean(locations[1])
y_avg = np.mean(locations[0])

我在 cv2.ConnectedComponents() 标记的 10 个不同对象矩阵上重复这个过程。这次我这样做是为了获得单个红色物体的质心。

以下是我正在使用的代码

_,labels = cv2.connectedComponents(red_only_array, connectivity = 8)
b = np.matrix(labels)

Obj1= b==1
Obj1 = np.uint8(Obj1)
Obj1[Obj1>0] =255
c1_max = np.where(Obj1 ==255)
centroid1 = np.array([np.mean(c1_max[1]),np.mean(c1_max[0])])

Obj2= b==2
Obj2 = np.uint8(Obj2)
Obj2[Obj2>0] =255
c2_max = np.where(Obj2 ==255)
centroid2 = np.array([np.mean(c2_max[1]),np.mean(c2_max[0])])

上述代码重复直到 b ==10

现在，我在 8GB RAM 的 Raspberry Pi 4 上延迟了大约 160 毫秒。我的同事认为 np.where() 是我代码中的瓶颈。有没有办法进一步优化呢？我的目标循环时间是 50 毫秒。

谢谢

标签： pythonnumpyopencvmultidimensional-arrayindexing

解决方案

剖析1：

当有 100 个连通分量且图像大小为 2000x2000 时，找到质心是最慢的一步。整个程序在笔记本电脑上运行需要 28 秒。

质心很慢

from skimage import measure
from skimage import filters
import numpy as np
import cProfile


def make_blobs(size=256, n_blobs=12):
    np.random.seed(1)
    im = np.zeros((size, size))
    points = size * np.random.random((2, n_blobs ** 2))
    im[(points[0]).astype(np.int), (points[1]).astype(np.int)] = 1
    im = filters.gaussian(im, sigma=size / (4. * n_blobs))
    blobs = im > 0.7 * im.mean()
    return blobs


def faster_centroid(img):
    s = 1 / np.mean(img)
    shape = img.shape
    x_coords = np.arange(shape[0])
    y_coords = np.arange(shape[1])
    x_mean = np.mean(img * x_coords[:, np.newaxis]) * s
    y_mean = np.mean(img * y_coords[np.newaxis, :]) * s
    return x_mean, y_mean


def label_blobs(blobs):
    all_labels = measure.label(blobs)
    blobs_labels = measure.label(blobs, background=0)
    return all_labels, blobs_labels


def find_all_centroids(all_labels):
    max_ix = np.max(all_labels)
    centroid_list = []
    for i in range(max_ix + 1):
        centroid = faster_centroid(all_labels == i)
        centroid_list.append(centroid)
    return centroid_list


def main():
    blobs = make_blobs(2000, n_blobs=100)
    # Label connected regions of an integer array.
    all_labels, blobs_labels = label_blobs(blobs)
    print(all_labels)
    all_centroids = find_all_centroids(all_labels)
    print(all_centroids)


cProfile.run("main()", "results.cprofile")

剖析 2：

[['<function get_centroids1 at 0x7f0027ba6280>', 1.3774937389971456], 
['<function get_centroids2 at 0x7f0027ba6310>', 2.308947408993845], 
['<function get_centroids3 at 0x7f0027ba63a0>', 0.695534451995627]]

4.262 main  red3.py:61
├─ 2.245 get_centroids2  red3.py:36
│  ├─ 1.258 [self]  
│  └─ 0.954 mean  <__array_function__ internals>:2
│        [5 frames hidden]  <__array_function__ internals>, numpy...
│           0.954 ufunc.reduce  <built-in>:0
├─ 1.334 get_centroids1  red3.py:25
│  ├─ 1.031 where  <__array_function__ internals>:2
│  │     [3 frames hidden]  <__array_function__ internals>, <buil...
│  │        1.031 implement_array_function  <built-in>:0
│  ├─ 0.188 [self]  
│  └─ 0.080 mean  <__array_function__ internals>:2
│        [5 frames hidden]  <__array_function__ internals>, numpy...
└─ 0.683 get_centroids3  red3.py:51
   ├─ 0.333 <dictcomp>  red3.py:57
   ├─ 0.233 nonzero  <__array_function__ internals>:2
   │     [5 frames hidden]  <__array_function__ internals>, numpy...
   └─ 0.048 [self]  

from skimage import measure
from skimage import filters
import numpy as np
#import cProfile
from pyinstrument import Profiler
import timeit


def make_blobs(size=256, n_blobs=12):
    np.random.seed(1)
    im = np.zeros((size, size))
    points = size * np.random.random((2, n_blobs ** 2))
    im[(points[0]).astype(np.int), (points[1]).astype(np.int)] = 1
    im = filters.gaussian(im, sigma=size / (4. * n_blobs))
    blobs = im > 0.7 * im.mean()
    return blobs


def label_blobs(blobs):
    all_labels = measure.label(blobs)
    blobs_labels = measure.label(blobs, background=0)
    return all_labels, blobs_labels


def get_centroids1(all_labels):
    n_blobs = np.max(all_labels) + 1
    centroid_list = []
    for i in range(n_blobs):
        locations = np.where(all_labels == i)
        x_avg = np.mean(locations[1])
        y_avg = np.mean(locations[0])
        centroid_list.append([x_avg, y_avg])
    return centroid_list


def get_centroids2(all_labels):
    n_blobs = np.max(all_labels) + 1
    centroid_list = []
    for i in range(n_blobs):
        img = (all_labels == i)
        s = 1 / np.mean(img)
        shape = img.shape
        x_coords = np.arange(shape[0])
        y_coords = np.arange(shape[1])
        x_mean = np.mean(img * x_coords[:, np.newaxis]) * s
        y_mean = np.mean(img * y_coords[np.newaxis, :]) * s
        centroid_list.append([x_mean, y_mean])
    return centroid_list


def get_centroids3(x):
    # https://stackoverflow.com/questions/32748950/
    n_blobs = np.max(x) + 1
    nz = np.nonzero(x)
    coords = np.column_stack(nz)
    nzvals = x[nz[0], nz[1]]
    res = {k: coords[nzvals == k] for k in range(1, n_blobs + 1)}
    return res


def main():
    f_list = [get_centroids1, get_centroids2, get_centroids3]
    blobs = make_blobs(2000, n_blobs=5)
    # Label connected regions of an integer array.
    all_labels, blobs_labels = label_blobs(blobs)

    profiler = Profiler()
    profiler.start()
    timings = []
    for f in f_list:
        s = timeit.default_timer()
        for i in range(10):
            r = f(all_labels)
        e = timeit.default_timer()
        print(r)
        timings.append([str(f), e - s])

    print(timings)
    profiler.stop()
    print(profiler.output_text(unicode=True, color=True))


main()

推荐阅读