『OpenCV3』Harris角點特征_API調用及python手動實現

一、OpenCV接口調用示意

介紹了OpenCV3中提取圖像角點特征的函數：

# coding=utf-8
import cv2
import numpy as np


'''Harris算法角點特征提取'''

img = cv2.imread('chess_board.png')
gray = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
gray = np.float32(gray)

# {標記點大小，敏感度（3~31,越小越敏感）}
# OpenCV函數cv2.cornerHarris() 有四個參數 其作用分別為 :
#img - Input image, it should be grayscale and float32 type.
#blockSize - It is the size of neighbourhood considered for corner detection
#ksize - Aperture parameter of Sobel derivative used.
#k - Harris detector free parameter in the equation,在0.04 到0.05之間
dst = cv2.cornerHarris(gray,2,23,0.04)
img[dst>0.01 * dst.max()] = [0,0,255]

cv2.imshow('corners',img)
cv2.waitKey()
cv2.destroyAllWindows()

 dst = cv2.cornerHarris(gray,2,23,0.04)中第3個參數(23)調整對結果影響如下：

取值為3時：

取值為23時：

二、使用Python實現harris膠墊檢測

計算機視覺課后作業，因為已經提交了一段時間了，之前也注意到網上很少有python版本的harris角點檢測代碼，所以開源出來，

# Author : hellcat
# Time   : 18-3-22

"""
import os
os.environ["CUDA_VISIBLE_DEVICES"]="-1"
 
import numpy as np
np.set_printoptions(threshold=np.inf)
 
import tensorflow as tf
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
"""

import numpy as np
from PIL import Image
from skimage import filters
from datetime import datetime
import matplotlib.pyplot as plt

IMAGE_PATH = '1653033843.jpg'     # 圖片路徑
WITH_NMS = False                  # 是否非極大值抑制，True/False
k = 0.04                          # 響應函數參數k
threshold = 0.01                  # 界定閾值

img = Image.open(IMAGE_PATH)
img_gray = img.convert('L')
img_num = np.pad(np.asarray(img_gray, dtype=np.float32), ((1, 1), (1, 1)), 'constant')
h, w = img_num.shape  # padding之后的圖像尺寸

# 計算Ix,Iy
grad = np.empty([h, w, 2], dtype=np.float)
grad[:, 1:-1, 0] = img_num[:, 2:] - img_num[:, :-2]  # Ix
grad[1:-1, :, 1] = img_num[2:, :] - img_num[:-2, :]  # Iy

# 計算Ixx,Iyy,Ixy
m = np.empty([h, w, 3], dtype=np.float)
# m[:, :, 0] = grad[:, :, 0]**2
# m[:, :, 1] = grad[:, :, 0]**2
# m[:, :, 2] = grad[:, :, 0]*grad[:, :, 1]
m[:, :, 0] = filters.gaussian(grad[:, :, 0]**2, sigma=2)  # Ixx
m[:, :, 1] = filters.gaussian(grad[:, :, 1]**2, sigma=2)  # Iyy
m[:, :, 2] = filters.gaussian(grad[:, :, 0]*grad[:, :, 1], sigma=2)  # Ixy
m = [np.array([[m[i, j, 0], m[i, j, 2]],
               [m[i, j, 2], m[i, j, 1]]]) for i in range(h) for j in range(w)]

# 記錄一下R計算時耗
start = datetime.now()
# 0:00:42.123384 迭代器策略：用時間換空間
# R = np.array([d-k*t**2 for d, t in zip(map(np.linalg.det, m), map(np.trace, m))])
# 0:00:35.846864
D, T = list(map(np.linalg.det, m)), list(map(np.trace, m))
R = np.array([d-k*t**2 for d, t in zip(D, T)])
end = datetime.now()
print(end-start)

R_max = np.max(R)
R = R.reshape(h, w)

# 標注角點
record = np.zeros_like(R, dtype=np.int)
img_row = np.pad(np.asarray(img, dtype=np.float32), ((1, 1), (1, 1), (0, 0)), 'constant')
for i in range(1, h-2):
    for j in range(1, w-2):
        if WITH_NMS:
            if R[i, j] > R_max*threshold and R[i, j] == np.max(R[i-1:i+2, j-1:j+2]):
                record[i, j] = 255
                img_row[i, j] = [255, 255, 255]
        else:
            if R[i, j] > R_max*0.01:
                record[i, j] = 255
                img_row[i, j] = [255, 255, 255]
# record[R > 0.01*R_max] = 255
# img_row[R > 0.01*R_max] = [255, 255, 255]

# 圖像展示與保存
res = Image.fromarray(np.uint8(record[1:-1, 1:-1]))
img_row = Image.fromarray(np.uint8(img_row[1:-1, 1:-1]))

plt.figure()
plt.subplot(1, 2, 1)
plt.imshow(res)
plt.subplot(1, 2, 2)
plt.imshow(img_row)
if WITH_NMS:
    plt.savefig('角點檢測_NMS.jpg')
    res.save('角點檢測_NMS_1.png')
    img_row.save('角點檢測_NMS_2.png')
else:
    plt.savefig('角點檢測_no_NMS.jpg')
    res.save('角點檢測_no_NMS_1.png')
    img_row.save('角點檢測_no_NMS_2.png')

實際上在計算Ixx,Ixy,Iyy時要進行高斯濾波，理論推導中都采用了最簡單的權重(全部為1)，這點注意，使用全1權重啥也檢測不出來。

不進行非極大值抑制結果：

進行非極大值抑制結果（實際上檢測出來的點很多，因為分辨率看不清）：