过滤 :是信号和图像处理中基本的任务。其目的是根据应用环境的不同,选择性的提取图像中某些认为是重要的信息。过滤可以移除图像中的噪音、提取感兴趣的可视特征、允许图像重采样等等。
频域分析 :将图像分成从低频到高频的不同部分。低频对应图像强度变化小的区域,而高频是图像强度变化非常大的区域。
在频率分析领域的框架中,滤波器是一个用来增强图像中某个波段或频率并阻塞(或降低)其他频率波段的操作。低通滤波器是消除图像中高频部分,但保留低频部分。高通滤波器消除低频部分。
滤波(高通、低通、带通、带阻) 、模糊、去噪、平滑等。
图像在频域里面,频率低的地方说明它是比较平滑的,因为平滑的地方灰度值变化比较小
,而频率高的地方通常是边缘或者噪声,因为这些地方往往是灰度值突变的。
高通滤波
就是保留频率比较高的部分,即突出边缘;低通滤波
就是保留频率比较低的地方,即平滑图像,弱化边缘,消除噪声。在pytorch中实现将sobel算子和卷积层结合来提取图像中物体的边缘轮廓图,如下代码是卷积执行soble边缘检测算子的实现:
import torch
import numpy as np
from torch import nn
from PIL import Image
from torch.autograd import Variable
import torch.nn.functional as F
# https://blog.csdn.net/weicao1990/article/details/100521530
def nn_conv2d(im):
# 用nn.Conv2d定义卷积操作
conv_op = nn.Conv2d(1, 1, 3, bias=False)
# 定义sobel算子参数
sobel_kernel = np.array([[-1, -1, -1], [-1, 8, -1], [-1, -1, -1]], dtype='float32')
# 将sobel算子转换为适配卷积操作的卷积核
sobel_kernel = sobel_kernel.reshape((1, 1, 3, 3))
# 给卷积操作的卷积核赋值
conv_op.weight.data = torch.from_numpy(sobel_kernel)
# 对图像进行卷积操作
edge_detect = conv_op(Variable(im))
# 将输出转换为图片格式
edge_detect = edge_detect.squeeze().detach().numpy()
return edge_detect
def functional_conv2d(im):
sobel_kernel = np.array([[-1, -1, -1], [-1, 8, -1], [-1, -1, -1]], dtype='float32') #
sobel_kernel = sobel_kernel.reshape((1, 1, 3, 3))
weight = Variable(torch.from_numpy(sobel_kernel))
edge_detect = F.conv2d(Variable(im), weight)
edge_detect = edge_detect.squeeze().detach().numpy()
return edge_detect
def main():
# 读入一张图片,并转换为灰度图
im = Image.open('./cat.jpg').convert('L')
# 将图片数据转换为矩阵
im = np.array(im, dtype='float32')
# 将图片矩阵转换为pytorch tensor,并适配卷积输入的要求
im = torch.from_numpy(im.reshape((1, 1, im.shape[0], im.shape[1])))
# 边缘检测操作
# edge_detect = nn_conv2d(im)
edge_detect = functional_conv2d(im)
# 将array数据转换为image
im = Image.fromarray(edge_detect)
# image数据转换为灰度模式
im = im.convert('L')
# 保存图片
im.save('edge.jpg', quality=95)
if __name__ == "__main__":
main()
效果展示:
四种算子:
'''
滤波与卷积
'''
# https://blog.csdn.net/m0_43609475/article/details/112447397
import cv2
import numpy as np
import matplotlib.pyplot as plt
def Padding(image,kernels_size,stride = [1,1],padding = "same"):
'''
对图像进行padding
:param image: 要padding的图像矩阵
:param kernels_size: list 卷积核大小[h,w]
:param stride: 卷积步长 [左右步长,上下步长]
:param padding: padding方式
:return: padding后的图像
'''
if padding == "same":
h,w = image.shape
p_h =max((stride[0]*(h-1)-h+kernels_size[0]),0) # 高度方向要补的0
p_w =max((stride[1]*(w-1)-w+kernels_size[1]),0) # 宽度方向要补的0
p_h_top = p_h//2 # 上边要补的0
p_h_bottom = p_h-p_h_top # 下边要补的0
p_w_left = p_w//2 # 左边要补的0
p_w_right = p_w-p_w_left # 右边要补的0
# print(p_h_top,p_h_bottom,p_w_left,p_w_right) # 输出padding方式
padding_image = np.zeros((h+p_h, w+p_w), dtype=np.uint8)
for i in range(h):
for j in range(w):
padding_image[i+p_h_top][j+p_w_left] = image[i][j] # 将原来的图像放入新图中做padding
return padding_image
else:
return image
def filtering_and_convolution(image,kernels,stride,padding = "same"):
'''
:param image: 要卷积的图像
:param kernels: 卷积核 列表
:param stride: 卷积步长 [左右步长,上下步长]
:param padding: padding方式 “same”or“valid”
:return:
'''
image_h,image_w = image.shape
kernels_h,kernels_w = np.array(kernels).shape
# 获取卷积核的中心点
kernels_h_core = int(kernels_h/2+0.5)-1
kernels_w_core = int(kernels_w/2+0.5)-1
if padding == "valid":
# 计算卷积后的图像大小
h = int((image_h-kernels_h)/stride[0]+1)
w = int((image_w-kernels_w)/stride[1]+1)
# 生成卷积后的图像
conv_image = np.zeros((h,w),dtype=np.uint8)
# 计算遍历起始点
h1_start = kernels_h//2
w1_start = kernels_w//2
ii=-1
for i in range(h1_start,image_h - h1_start,stride[0]):
ii += 1
jj = 0
for j in range(w1_start,image_w - w1_start,stride[1]):
sum = 0
for x in range(kernels_h):
for y in range(kernels_w):
# print(i,j,int((i/image_h)*h),int((j/image_w)*w), i-kernels_h_core + x, j-kernels_w_core+y,x,y)
sum += int(image[i-kernels_h_core+x][j-kernels_w_core+y]*kernels[x][y])
conv_image[ii][jj] = sum
jj += 1
return conv_image
if padding == "same":
# 对原图进行padding
kernels_size = [kernels_h, kernels_w]
pad_image = Padding(image,kernels_size,stride,padding="same")
# 计算卷积后的图像大小
h = image_h
w = image_w
# 生成卷积后的图像
conv_image = np.zeros((h,w),dtype=np.uint8)
# # 计算遍历起始点
h1_start = kernels_h//2
w1_start = kernels_w//2
ii=-1
for i in range(h1_start,image_h - h1_start,stride[0]):
ii +=1
jj = 0
for j in range(w1_start,image_w - w1_start,stride[1]):
sum = 0
for x in range(kernels_h):
for y in range(kernels_w):
sum += int(image[i-kernels_h_core+x][j-kernels_w_core+y]*kernels[x][y])
conv_image[ii][jj] = sum
jj += 1
return conv_image
def sobel_filter(image):
h = image.shape[0]
w = image.shape[1]
image_new = np.zeros(image.shape, np.uint8)
for i in range(1, h - 1):
for j in range(1, w - 1):
sx = (image[i + 1][j - 1] + 2 * image[i + 1][j] + image[i + 1][j + 1]) - \
(image[i - 1][j - 1] + 2 * image[i - 1][j] + image[i - 1][j + 1])
sy = (image[i - 1][j + 1] + 2 * image[i][j + 1] + image[i + 1][j + 1]) - \
(image[i - 1][j - 1] + 2 * image[i][j - 1] + image[i + 1][j - 1])
image_new[i][j] = np.sqrt(np.square(sx) + np.square(sy))
# image_new[i][j] = sy
return image_new
# 设置matplotlib正常显示中文和负号
plt.rcParams['font.sans-serif']=['SimHei'] # 用黑体显示中文
plt.rcParams['axes.unicode_minus']=False # 正常显示负号
img = cv2.imread('lenna.png',1)
img_gray = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
plt.subplot(331)
plt.imshow(img_gray,cmap="gray")
plt.title("原图")
sobel_Gy = [[-1,0,1],[-2,0,2],[-1,0,1]]
Average = [[1/9,1/9,1/9],[1/9,1/9,1/9],[1/9,1/9,1/9]]
Gaussian = [[1/16,2/16,1/16],[2/16,4/16,2/16],[1/16,2/16,1/16]]
Laplace = [[-1,-1,-1],[-1,9,-1],[-1,-1,-1]]
stride=[1,1]
img_sobel_Gy = filtering_and_convolution(img_gray,sobel_Gy,stride,padding="same")
img_Average = filtering_and_convolution(img_gray,Average,stride,padding="same")
img_Gaussian = filtering_and_convolution(img_gray,Gaussian,stride,padding="same")
img_Laplace = filtering_and_convolution(img_gray,Laplace,stride,padding="same")
plt.subplot(332)
plt.imshow(img_sobel_Gy,cmap = "gray")
plt.title("sobel_Gy")
plt.subplot(333)
plt.imshow(img_Average,cmap = "gray")
plt.title("Average")
plt.subplot(334)
plt.imshow(img_Gaussian,cmap = "gray")
plt.title("Gaussian")
plt.subplot(335)
plt.imshow(img_Laplace,cmap = "gray")
plt.title("Laplace")
plt.show()
import cv2
import numpy as np
# https://wangsp.blog.csdn.net/article/details/82872838
def blur_demo(image):
"""
均值模糊 : 去随机噪声有很好的去噪效果
(1, 15)是垂直方向模糊,(15, 1)是水平方向模糊
"""
dst = cv2.blur(image, (1, 15))
cv2.imshow("avg_blur_demo", dst)
def median_blur_demo(image): # 中值模糊 对椒盐噪声有很好的去燥效果
dst = cv2.medianBlur(image, 5)
cv2.imshow("median_blur_demo", dst)
def custom_blur_demo(image):
"""
用户自定义模糊
下面除以25是防止数值溢出
"""
kernel = np.ones([5, 5], np.float32)/25
dst = cv2.filter2D(image, -1, kernel)
cv2.imshow("custom_blur_demo", dst)
src = cv2.imread("./fapiao.png")
img = cv2.resize(src,None,fx=0.8,fy=0.8,interpolation=cv2.INTER_CUBIC)
cv2.imshow('input_image', img)
blur_demo(img)
median_blur_demo(img)
custom_blur_demo(img)
cv2.waitKey(0)
cv2.destroyAllWindows()
进行边缘保留滤波通常用到两个方法:高斯双边滤波和均值迁移滤波。
"""
bilateralFilter(src, d, sigmaColor, sigmaSpace[, dst[, borderType]]) -> dst
- src: 输入图像。
- d: 在过滤期间使用的每个像素邻域的直径。如果输入d非0,则sigmaSpace由d计算得出,如果sigmaColor没输入,则sigmaColor由sigmaSpace计算得出。
- sigmaColor: 色彩空间的标准方差,一般尽可能大。
较大的参数值意味着像素邻域内较远的颜色会混合在一起,
从而产生更大面积的半相等颜色。
- sigmaSpace: 坐标空间的标准方差(像素单位),一般尽可能小。
参数值越大意味着只要它们的颜色足够接近,越远的像素都会相互影响。
当d > 0时,它指定邻域大小而不考虑sigmaSpace。
否则,d与sigmaSpace成正比。
"""
import cv2
def bi_demo(image): #双边滤波
dst = cv2.bilateralFilter(image, 0, 100, 5)
cv2.imshow("bi_demo", dst)
def shift_demo(image): #均值迁移
dst = cv2.pyrMeanShiftFiltering(image, 10, 50)
cv2.imshow("shift_demo", dst)
src = cv2.imread('./100.png')
img = cv2.resize(src,None,fx=0.8,fy=0.8,
interpolation=cv2.INTER_CUBIC)
cv2.imshow('input_image', img)
bi_demo(img)
shift_demo(img)
cv2.waitKey(0)
cv2.destroyAllWindows()
import cv2
import numpy as np
def salt(img, n):
for k in range(n):
i = int(np.random.random() * img.shape[1])
j = int(np.random.random() * img.shape[0])
if img.ndim == 2:
img[j,i] = 255
elif img.ndim == 3:
img[j,i,0]= 255
img[j,i,1]= 255
img[j,i,2]= 255
return img
img = cv2.imread("./original_img.png",cv2.IMREAD_GRAYSCALE)
result = salt(img, 500)
median = cv2.medianBlur(result, 5)
cv2.imshow("original_img", img)
cv2.imshow("Salt", result)
cv2.imshow("Median", median)
cv2.waitKey(0)
cv2.destroyWindow()
import cv2
import numpy as np
def clamp(pv):
if pv > 255:
return 255
if pv < 0:
return 0
else:
return pv
def gaussian_noise(image): # 加高斯噪声
h, w, c = image.shape
for row in range(h):
for col in range(w):
s = np.random.normal(0, 20, 3)
b = image[row, col, 0] # blue
g = image[row, col, 1] # green
r = image[row, col, 2] # red
image[row, col, 0] = clamp(b + s[0])
image[row, col, 1] = clamp(g + s[1])
image[row, col, 2] = clamp(r + s[2])
cv2.imshow("noise image", image)
src = cv2.imread('888.png')
cv2.imshow('input_image', src)
gaussian_noise(src)
dst = cv2.GaussianBlur(src, (15,15), 0) #高斯模糊
cv2.imshow("Gaussian_Blur2", dst)
cv2.waitKey(0)
cv2.destroyAllWindows()
使用的函数有:cv2.Sobel()
, cv2.Schar()
, cv2.Laplacian()
Sobel,scharr其实是求一阶或者二阶导数。scharr是对Sobel的优化。
Laplacian是求二阶导数。
"""
dst = cv2.Sobel(src, ddepth, dx, dy[, dst[, ksize[, scale[, delta[, borderType]]]]])
src: 需要处理的图像;
ddepth: 图像的深度,-1表示采用的是与原图像相同的深度。
目标图像的深度必须大于等于原图像的深度;
dx和dy: 求导的阶数,0表示这个方向上没有求导,一般为0、1、2。
dst 不用解释了;
ksize: Sobel算子的大小,必须为1、3、5、7。 ksize=-1时,会用3x3的Scharr滤波器,
它的效果要比3x3的Sobel滤波器要好
scale: 是缩放导数的比例常数,默认没有伸缩系数;
delta: 是一个可选的增量,将会加到最终的dst中, 默认情况下没有额外的值加到dst中
borderType: 是判断图像边界的模式。这个参数默认值为cv2.BORDER_DEFAULT。
"""
import cv2
img=cv2.imread('888.png',cv2.IMREAD_COLOR)
x=cv2.Sobel(img,cv2.CV_16S,1,0)
y=cv2.Sobel(img,cv2.CV_16S,0,1)
absx=cv2.convertScaleAbs(x)
absy=cv2.convertScaleAbs(y)
dist=cv2.addWeighted(absx,0.5,absy,0.5,0)
cv2.imshow('original_img',img)
cv2.imshow('y',absy)
cv2.imshow('x',absx)
cv2.imshow('dsit',dist)
cv2.waitKey(0)
cv2.destroyAllWindows()
import matplotlib.pyplot as plt
import numpy as np
import cv2
def log_filter(gray_img):
gaus_img = cv2.GaussianBlur(gray_img,(3,3),sigmaX=0) # 以核大小为3x3,方差为0
log_img = cv2.Laplacian(gaus_img,cv2.CV_16S,ksize=3) # laplace检测
log_img = cv2.convertScaleAbs(log_img)
return log_img
def filter_imgs(gray_img):
# 尝试一下不同的核的效果
Emboss = np.array([[ -2,-1, 0],
[ -1, 1, 1],
[ 0, 1, 2]])
Motion = np.array([[ 0.333, 0, 0],
[ 0, 0.333, 0],
[ 0, 0, 0.333]])
Emboss_img = cv2.filter2D(gray_img,cv2.CV_16S,Emboss)
Motion_img = cv2.filter2D(gray_img, cv2.CV_16S, Motion)
Emboss_img = cv2.convertScaleAbs(Emboss_img)
Motion_img = cv2.convertScaleAbs(Motion_img)
different_V = np.array([[ 0, -1, 0],
[ 0, 1, 0],
[ 0, 0, 0]])
different_H = np.array([[ 0, 0, 0],
[ -1, 1, 0],
[ 0, 0, 0]])
different_temp = cv2.filter2D(gray_img,cv2.CV_16S,different_V)
different_temp = cv2.filter2D(different_temp, cv2.CV_16S, different_H)
different_img = cv2.convertScaleAbs(different_temp)
Sobel_V = np.array([[ 1, 2, 1],
[ 0, 0, 0],
[ -1, -2, -1]])
Sobel_H = np.array([[ 1, 0, -1],
[ 2, 0, -2],
[ 1, 0, -1]])
Sobel_temp = cv2.filter2D(gray_img,cv2.CV_16S, Sobel_V)
Sobel_temp = cv2.filter2D(Sobel_temp, cv2.CV_16S, Sobel_H)
Sobel_img = cv2.convertScaleAbs(Sobel_temp)
Prewitt_V = np.array([[-1, -1, -1],
[ 0, 0, 0],
[ 1, 1, 1]])
Prewitt_H = np.array([[-1, 0, 1],
[-1, 0, 1],
[-1, 0, 1]])
Prewitt_temp = cv2.filter2D(gray_img, cv2.CV_16S, Prewitt_V)
Prewitt_temp = cv2.filter2D(Prewitt_temp, cv2.CV_16S, Prewitt_H)
Prewitt_img = cv2.convertScaleAbs(Prewitt_temp)
kernel_P = np.array([[0, 0, -1, 0, 0],
[0, -1, -2, -1, 0],
[-1,-2, 16, -2,-1],
[0, -1, -2, -1, 0],
[0, 0, -1, 0, 0]])
kernel_N = np.array([[0, 0, 1, 0, 0],
[0, 1, 2, 1, 0],
[1, 2, -16, 2, 1],
[0, 1, 2, 1, 0],
[0, 0, 1, 0, 0]])
lap4_filter = np.array([[0, 1, 0],
[1, -4, 1],
[0, 1, 0]]) # 4邻域laplacian算子
lap8_filter = np.array([[0, 1, 0],
[1, -8, 1],
[0, 1, 0]]) # 8邻域laplacian算子
lap_filter_P = cv2.filter2D(gray_img, cv2.CV_16S, kernel_P)
edge4_img_P = cv2.filter2D(lap_filter_P, cv2.CV_16S, lap4_filter)
edge4_img_P = cv2.convertScaleAbs(edge4_img_P)
edge8_img_P = cv2.filter2D(lap_filter_P, cv2.CV_16S, lap8_filter)
edge8_img_P = cv2.convertScaleAbs(edge8_img_P)
lap_filter_N = cv2.filter2D(gray_img, cv2.CV_16S, kernel_N)
edge4_img_N = cv2.filter2D(lap_filter_N, cv2.CV_16S, lap4_filter)
edge4_img_N = cv2.convertScaleAbs(edge4_img_N)
edge8_img_N = cv2.filter2D(lap_filter_N, cv2.CV_16S, lap8_filter)
edge8_img_N = cv2.convertScaleAbs(edge8_img_N)
return (Emboss_img,Motion_img,different_img,Sobel_img,Prewitt_img,edge4_img_P,edge8_img_P,edge4_img_N,edge8_img_N)
def show(Filter_imgs):
titles = [u'原图', u'Laplacian算子',\
u'Emboss滤波',u'Motion滤波',
u'diff(差分)滤波',u'Sobel滤波',u'Prewitt滤波',
u'Lap4算子-kernel_P', u'Lap8算子-kernel_P',
u'Lap4算子-kernel_N', u'Lap8算子-kernel_N']
plt.rcParams['font.sans-serif'] = ['SimHei']
plt.figure(figsize=(12, 8))
for i in range(len(titles)):
plt.subplot(3, 4, i + 1)
plt.imshow(Filter_imgs[i])
plt.title(titles[i])
plt.xticks([]), plt.yticks([])
plt.show()
if __name__ == '__main__':
img = cv2.imread('yinying3.png')
img_raw = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
gray_img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
LoG_img = log_filter(gray_img)
Filter_imgs = [img_raw,LoG_img]
Filter_imgs.extend(filter_imgs(gray_img))
show(Filter_imgs)
Pytorch 实现sobel算子的卷积操作_洪流之源的博客-CSDN博客_卷积实现sobel
(七)滤波与卷积_淡定的炮仗的博客-CSDN博客_卷积滤波
图像处理之高通滤波及低通滤波_ReWz的博客-CSDN博客_低通滤波和高通滤波对图像的影响
数字图像处理——图像梯度和空间滤波 - 知乎 (zhihu.com)
OpenCV—Python 图像滤波(均值、中值、高斯、高斯双边、高通等滤波)_SongpingWang的博客-CSDN博客