Question

我想为具有两个约束的非二进制用例实现扩展操作：1）结果应该是Opencv的cv2.dilate方法的副本，2）计算时间应该比cv2快10倍。扩张法。下面是我的相同代码段：

import cv2
from numba import jit
from scipy.ndimage.morphology import (generate_binary_structure,
                                  iterate_structure)
import numpy as np
import time

struct = generate_binary_structure(2, 1) #default structuring element
neighborhood = iterate_structure(struct, 20) #create a structuring element of shape(41,41)
neighborhood = neighborhood.astype(np.uint8) #convert the dtype from bool to uint8

n_h,n_w = neighborhood.shape #get the shape of the structuring kernel

test_arr = np.random.random((320,240)) #create a test data to perform dilation

app_len = (n_h - 3)//2 + 1 #compute the requirements for zero padding

@jit(nopython=True)
def compute_dilation(test_arr,app_len,neighborhood,n_h):

    #create an array to hold the values of the dilation operation
    dil_arr = np.zeros((test_arr.shape)) 
    #get the height and width of the test_arr
    test_arr_h,test_arr_w = test_arr.shape
    #create an array for zero padding the test array
    #zero padding will help to avoid multiple if else checks while performing the dilation operation on the test_arr
    zero_padded_array = np.zeros((test_arr_h+2*app_len,test_arr_w+2*app_len))
    #append the test array to the zero padded array
    zero_padded_array[app_len:app_len+test_arr_h,app_len:app_len+test_arr_w] = test_arr

    #iterate over each of the pixel values and populate them with the result of dilation operation
    for i in range(test_arr_h):
        for j in range(test_arr_w):
            #perform an element wise multiplication followed by extracting the maximum value
            dil_arr[i,j] = np.max(np.multiply(zero_padded_array[i:i+n_h,j:j+n_h],neighborhood))  

    #return the output 
    return dil_arr

st_numba = time.time() #log the start time
numba_out = compute_dilation(test_arr,app_len,neighborhood,n_h)
print(time.time()-st_numba) #compute the elapsed difference

st_cv2 = time.time() #log the start time
cv2_out = cv2.dilate(test_arr,neighborhood)
print(time.time()-st_cv2) #compute the elapsed time

#validate the results of the dilation outputs from opencv vs the custom implementation

print(np.array_equal(numba_out,cv2_out))

尽管两种方法的输出都匹配，但是cv2.dilate方法所花费的时间（40毫秒）比我的自定义实现（900毫秒）好得多。我应该做些什么改变才能达到预期的效果？

比较OpenCV的膨胀实现与使用numba的膨胀函数的自定义实现的性能

0 个答案: