OpenCV algorithm analysis
- 1 Image Enhancement
-
- 1.1 Meaning
- 1.2 Method
-
- 1.2.1 Histogram equalization
- 1.2.2 gamma transformation
- 2 Noise Cancellation
-
- 2.1 Meaning
- 2.2 Method
-
- 2.2.1 Gaussian filter
- 2.2.2 Mean filtering
- 2.2.3 Median filtering
- 3 Edge detection
-
- 3.1 canny
- 4 HOG Feature Extraction
-
- 4.1 Meaning
- 4.2 Process
- 4.3 Case
- 6 two games
-
- 6.1 Integration of three functions
- 6.2 Object Detection
- 6.3 Detailed explanation of yolov5 code
1 image enhancement
1.1 Meaning
Highlight image details and make images clearer.
1.2 Method
Histogram equalization
1.2.1 Histogram equalization
Histogram equalization
import cv2
import numpy as np
import matplotlib.pyplot as plt
def histequal(img_gray):
# 1 Calculate the frequency, cumulative frequency and normalization of each pixel
h,w = img_gray.shape
cum_freq = np.array([(img_gray==i).sum() for i in range(255)]).cumsum()
cum_freq = (cum_freq/cum_freq[-1]*255).astype(np.uint8)
tmp = img_gray. copy()
for i in range(255):
tmp[img_gray == i] = cum_freq[i]
return tmp
img = cv2.imread('opencv/features/cat0.jpg', 0)
out = histequal(img)
plt. subplot(121)
plt. title('org')
plt.imshow(img,cmap='gray')
plt. subplot(122)
plt. title('histequal')
plt.imshow(out,cmap='gray')
plt. show()
1.2.2 gamma transformation
import cv2
import numpy as np
import matplotlib.pyplot as plt
def gamma(img,c=2,r=2):
return np.array(c*img**r,dtype=np.uint8).clip(0,255)
img = cv2.imread('/Users/liushuang/Desktop/LearnGit/Project combat/fry detection/yolov5-master/opencv/features/cat.jpg')
out = gamma(img)
plt. subplot(121)
plt. title('org')
plt.imshow(img,cmap='gray')
plt. subplot(122)
plt. title('gamma')
plt.imshow(out,cmap='gray')
plt. show()
2 Denoising
2.1 Meaning
remove noise from image
2.2 Method
Spatial filtering (mean filtering, Gaussian filtering, median filtering, bilateral filtering), transform filtering (Fourier, wavelet transform), morphology, etc.
2.2.1 Gaussian filter
Gaussian filter
Gaussian filtering is suitable for dealing with normally distributed noise. The characteristic of the filter is that the weights are normally distributed. The larger the variance, the more obvious the filtering effect; otherwise, the worse the effect.
import cv2
import numpy as np
import matplotlib.pyplot as plt
def guassian_filter(img, k_size=3, sigma=2):
# get channel
if len(img.shape) == 3:
h, w, c = img. shape
else:
img = np. expand_dims(img, axis=-1)
h, w, c = img. shape
#img padiing
pad = (k_size - 1) // 2
out_put = np.zeros((h + pad * 2, w + pad * 2,3), dtype=np.float32)
out_put[pad:pad + h, pad:pad + w] = img.copy().astype(np.float32)
# generate kernel
k = np.zeros((k_size, k_size), dtype = np.float32)
for i in range(-pad, k_size - pad):
for j in range(-pad, k_size - pad):
k[j + pad, i + pad] = np. exp(-(i ** 2 + j ** 2) / (2 * sigma ** 2))
k /= (2 * np.pi * sigma ** 2)
k /= k.sum()
tmp = out_put. copy()
# Gaussian filter
for i in range(h):
for j in range(w):
for cc in range(c):
out_put[pad + i, pad + j, cc] = np. sum(k * tmp[i:i + k_size, j:j + k_size, cc])
out_put = np. clip(out_put, 0, 255)
return out_put[pad:pad + h, pad:pad + w].astype(np.uint8)
#
img = cv2.imread('opencv/features/dogGauss.jpeg')
out = guassian_filter(img, k_size=3, sigma=2)
plt. subplot(121)
plt. title('org')
plt.imshow( img[...,::-1])
plt. subplot(122)
plt. title('out')
plt.imshow( out[...,::-1])
plt. show()
# sigma = 1, Gaussian kernel
'''
array([[0.07511362, 0.12384141, 0.07511362],
[0.12384141, 0.20417996, 0.12384141],
[0.07511362, 0.12384141, 0.07511362]], dtype=float32)
# sigma = 3
array([[0.10699731, 0.11310981, 0.10699731],
[0.11310981, 0.11957153, 0.11310981],
[0.10699731, 0.11310981, 0.10699731]], dtype=float32)
'''
2.2.2 Mean filter
### 2.2.2 mean + filter
import cv2
import numpy as np
import matplotlib.pyplot as plt
def mean_filter(img, k_size=3):
#shape
if len(img.shape) != 3:
img = img[..., None]
h, w, c = img. shape
# 1 kernel
k = np.ones((k_size, k_size), np.float32) / (k_size * k_size)
# 2 out_Put
pad = (k_size - 1) // 2
out_put = np.zeros((h + 2*pad, w + 2*pad, c), np.float32)
out_put[pad:pad + h, pad:pad + w] = img.copy().astype(np.float32)
# 3 filter
tmp = out_put. copy()
for i in range(h):
for j in range(w):
for cc in range(c):
out_put[pad + i:, pad + j, cc] = (tmp[i:i + k_size, j:j + k_size, cc] * k).sum()
out_put = np. clip(out_put, 0, 255)
return out_put[pad:pad + h, pad:pad + w].astype(np.uint8)
# filter
img = cv2.imread('opencv/features/dogGauss.jpeg')
out = mean_filter(img, k_size=3)
plt. subplot(121)
plt. title('org')
plt.imshow(img[..., ::-1])
plt. subplot(122)
plt. title('out_mean_filter')
plt.imshow(out[..., ::-1])
plt. show()
2.2.3 Median filter
Take the median of the window as the value of the center point. If the effect is not good, you can increase the size of the kernel, or cycle several times.
import cv2
import numpy as np
import matplotlib.pyplot as plt
def median_filter(img, k_size=10):
#shape
if len(img.shape) != 3:
img = img[..., None]
h, w, c = img. shape
# 2 out_Put
pad = (k_size - 1) // 2
out_put = np. ones((h + 2 * pad, w + 2 * pad, c), np. float32) * (-1)
out_put[pad:pad + h, pad:pad + w] = img.copy().astype(np.float32)
# 3 filter
tmp = out_put. copy()
for i in range(h):
for j in range(w):
for cc in range(c):
t = tmp[i:i + k_size, j:j + k_size, cc]
out_put[pad + i:, pad + j, cc] = np.median(t[t > 0])
return out_put[pad:pad + h, pad:pad + w].astype(np.uint8)
# filter
img = cv2.imread('opencv/features/letteropen.png')
out = median_filter(img, k_size=20)
plt. subplot(121)
plt. title('org')
plt.imshow(img[..., ::-1])
plt. subplot(122)
plt. title('median_filter')
plt.imshow(out[..., ::-1])
plt. show()
3 Edge detection
3.1 canny
canny edge detection process
- smoothing
- gradient detection
- non-maximum suppression
- Hysteresis Thresholding
import cv2 as cv
import numpy as np
import matplotlib.pyplot as plt
def canny(img):
# 1. Smooth
img_blur = cv.GaussianBlur(img, (5, 5), 2)
# 2. Gradient
gradx = cv.Sobel(img_blur, cv.CV_64F, 1, 0) # x
grady = cv. Sobel(img_blur, cv. CV_64F, 0, 1) # y
R = np.abs(gradx) + np.abs(grady)
T = np.arctan(grady / (gradx + 1e-3))
# 3. Non-maximum suppression, refine the edge
h, w = R.shape
img_thin = np.zeros_like(R)
for i in range(1, h - 1):
for j in range(1, w - 1):
theta = T[i, j]
if -np.pi / 8 <= theta < np.pi / 8: # (-22,22)
if R[i, j] == max([R[i, j], R[i, j - 1], R[i, j + 1]]): # up and down
img_thin[i, j] = R[i, j]
elif -3 * np.pi / 8 <= theta < -np.pi / 8: # (-66,-22)
if R[i, j] == max([R[i, j], R[i - 1, j + 1], R[i + 1, j - 1]]): # slash
img_thin[i, j] = R[i, j]
elif np.pi / 8 <= theta < 3 * np.pi / 8: # (22,66)
if R[i, j] == max([R[i, j], R[i - 1, j - 1], R[i + 1, j + 1]]): # backslash
img_thin[i, j] = R[i, j]
else: # (66,110)
if R[i, j] == max([R[i, j], R[i - 1, j], R[i + 1, j]]): # around
img_thin[i, j] = R[i, j]
# 4 Dual Threshold Suppression
th1 = 5
th2 = 30
maxv = 255
img_edge = np.zeros_like(img_thin)
h, w = img_thin.shape
for i in range(1, h - 1):
for j in range(1, w - 1):
if img_thin[i, j] >= th2:
img_edge[i, j] = maxv
elif img_thin[i, j] > th1:
around = img_thin[i - 1:i + 2, j - 1:j + 2]
if around.max() >= th2:
img_edge[i, j] = maxv
return img_edge
img = cv2.imread('opencv/features/letteropen.png', 0)
out = canny(img)
plt. subplot(121)
plt. title('org')
plt.imshow(img,cmap='gray')
plt. subplot(122)
plt. title('canny')
plt.imshow(out,cmap='gray')
plt. show()
4 HOG feature extraction
4.1 Meaning
By extracting the features of the gradient histogram in the local area of the image, the features of the entire image are obtained. HOG remains invariant to both geometric and optical deformations. Through local normalization, the change of details is ignored, and the common feature is extracted.
4.2 Process
img –> cells,blocks –> Calculate the gradient histogram of each cell