要求
对一副图像加噪声,进行平滑,锐化作用。
待处理图像:
加噪
生成椒盐噪声:
def sp_noisy(image, s_vs_p=0.5, amount=0.08):
out = np.copy(image)
num_salt = np.ceil(amount * image.size * s_vs_p)
coords = [np.random.randint(0, i - 1, int(num_salt)) for i in image.shape]
out[tuple(coords)] = 255
num_pepper = np.ceil(amount * image.size * (1. - s_vs_p))
coords = [np.random.randint(0, i - 1, int(num_pepper)) for i in image.shape]
out[tuple(coords)] = 0
return out
结果
胡椒和盐各占0.5,总密度0.08的椒盐噪声:
平滑空间滤波(线性)
均值滤波
均值滤波过程:
g
(
x
,
y
)
=
∑
s
=
−
a
a
∑
t
=
−
b
b
w
(
s
,
t
)
f
(
x
+
s
,
y
+
t
)
∑
s
=
−
a
a
∑
t
=
−
b
b
w
(
s
,
t
)
g(x, y) = \frac{\sum_{s=-a}^a \sum_{t=-b}^b{ w(s, t) f(x+s, y+t)} }{\sum_{s=-a}^a \sum_{t=-b}^b w(s, t)}
g(x,y)=∑s=−aa∑t=−bbw(s,t)∑s=−aa∑t=−bbw(s,t)f(x+s,y+t)
a
=
(
m
−
1
)
/
2
a = (m -1)/2
a=(m−1)/2
b
=
(
n
−
1
)
/
2
b = (n - 1)/2
b=(n−1)/2
m=n=3方形卷积模板:
kernel = np.array([[1, 1, 1],
[1, 1, 1],
[1, 1, 1]], np.float32)/9
外围补0的线性滤波器:
def linear_filter(image, x, y, kernel, out):
sum_wf = 0
m = kernel.shape[0]
n = kernel.shape[1]
a = int((m - 1) / 2)
b = int((n - 1) / 2)
for s in range(-a, a + 1):
for t in range(-b, b + 1):
# convolution rotation 180
x_s = (x - s) if (x - s) in range(0, image.shape[0] - 1) else 0
y_t = (y - t) if (y - t) in range(0, image.shape[1] - 1) else 0
sum_wf += kernel[a + s][b + t] * image[x_s][y_t]
out[x][y] = sum_wf
空间滤波函数实现:
def spatial_filtering(image, kernel, filter_):
out = np.copy(image)
h = image.shape[0]
w = image.shape[1]
for x in range(h):
print(str(int(x/h * 100)) + "%")
for y in range(w):
filter_(image, x, y, kernel, out)
return out
调用
leaf_smooth = sp_convolution(leaf_sp_nose, k, linear_filter)
3 * 3均值滤波后:
另一个3 * 3 的均值滤波模板结果:
kernel = np.array([[1, 2, 1],
[2, 4, 2],
[1, 2, 1]], np.float32)/16
opencv 实现:
opencv速度要快很多,最后的效果是一样
5 * 5均值滤波:
kernel = np.ones((5, 5), np.float32)/(5**2)
15 * 15均值滤波:
kernel = np.ones((15, 15), np.float32)/(15**2)
图像太过模糊,因为对外围取了0,可以明显看到周围有暗边
opencv:
opencv外围不是补0
Embossment算子
kernel = np.array([[2, 0, 0],
[0, 0, 0],
[0, 0, 2]], np.float32)/4
对去椒盐燥没什么效果
统计排序滤波(非线性)
中值滤波
过程为求领域内像素值的中值,窗口由kernel给出,置1为需要统计的像素
中值滤波器:
def nonlinear_median_filter(image, x, y, kernel, out):
sp = []
m = kernel.shape[0]
n = kernel.shape[1]
a = int((m - 1) / 2)
b = int((n - 1) / 2)
for s in range(-a, a + 1):
for t in range(-b, b + 1):
x_s = (x + s) if (x + s) in range(0, image.shape[0] - 1) else 0
y_t = (y + t) if (y + t) in range(0, image.shape[1] - 1) else 0
if kernel[a + s][b + t]:
sp.append(image[x_s][y_t])
out[x][y] = np.median(sp)
3*3中值滤波结果
模板:
k = np.ones((3, 3), np.float32)/(3**2)
最大值最小值同理,下面是去取椒盐噪声对比图(均值,中值,最大值,最小值):
可以看到中值效果最好,最大值和最小值不适用于去除椒盐噪声
不同模板对比
k1 = np.array([
[1, 1, 1],
[1, 1, 1],
[1, 1, 1]
], np.float32)
k2 = np.ones((5, 5))
k3 = np.array([
[0, 0, 1, 0, 0],
[0, 0, 1, 0, 0],
[1, 1, 1, 1, 1],
[0, 0, 1, 0, 0],
[0, 0, 1, 0, 0],
])
使用k2图像整体颜色偏暗,个人感觉k3效果最好
空间锐化滤波器
待处理图:
二阶微分-拉普拉斯算子(线性)
∇
2
f
=
f
(
x
+
1
,
y
)
+
f
(
x
−
1
,
y
)
+
f
(
x
,
y
+
1
)
+
f
(
x
,
y
−
1
)
−
4
f
(
x
,
y
)
\nabla^2f =f(x+1,y) + f(x-1,y) + f(x,y+1) + f(x,y-1) - 4f(x,y)
∇2f=f(x+1,y)+f(x−1,y)+f(x,y+1)+f(x,y−1)−4f(x,y)
g
(
x
,
y
)
=
f
(
x
,
y
)
+
c
[
∇
2
f
(
x
,
y
)
]
g(x,y) = f(x,y) +c[\nabla^2f(x,y)]
g(x,y)=f(x,y)+c[∇2f(x,y)]
当
c
=
1
时
当c =1时
当c=1时
g
(
x
,
y
)
=
f
(
x
+
1
,
y
)
+
f
(
x
−
1
,
y
)
+
f
(
x
,
y
+
1
)
+
f
(
x
,
y
−
1
)
−
3
f
(
x
,
y
)
g(x,y) = f(x+1,y) + f(x-1,y) + f(x,y+1) + f(x,y-1) - 3f(x,y)
g(x,y)=f(x+1,y)+f(x−1,y)+f(x,y+1)+f(x,y−1)−3f(x,y)
线性滤波实现函数基本和平滑的一样,但是锐化运算时会出现小于0或大于255的情况,所以需要对其处理
也是进行卷积运算:
def spatial_filtering(image, kernel, filter_):
out = np.copy(image)
h = image.shape[0]
w = image.shape[1]
for x in range(h):
# print(str(int(x/h * 100)) + "%")
for y in range(w):
filter_(image, x, y, kernel, out)
return out
def linear_filter(image, x, y, kernel, out):
sum_wf = 0
m = kernel.shape[0]
n = kernel.shape[1]
a = int((m - 1) / 2)
b = int((n - 1) / 2)
for s in range(-a, a + 1):
for t in range(-b, b + 1):
# convolution rotation 180
x_s = (x - s) if (x - s) in range(0, image.shape[0] - 1) else 0
y_t = (y - t) if (y - t) in range(0, image.shape[1] - 1) else 0
sum_wf += kernel[a + s][b + t] * image[x_s][y_t]
if sum_wf < 0:
sum_wf = 0
if sum_wf > 255:
sum_wf = 255
out[x][y] = int(sum_wf)
模板:
laplacian_mask1 = np.array([
[0, 1, 0],
[1, -4, 1],
[0, 1, 0],
])
laplacian_mask2 = np.array([
[1, 1, 1],
[1, -8, 1],
[1, 1, 1],
])
laplacian_mask3 = np.array([
[-1, -1, -1],
[-1, 9, -1],
[-1, -1, -1],
])
模板2考虑了对角项,模板3对原图像进锐化,由于是线性操作,直接调用线性滤波
调用:
image_laplacian_mask_ = spatial_filtering(leaf, laplacian_mask_ , linear_filter)
结果
可以看到模板2的滤波效果要好与模板1,模板3实现了对原图像的锐化
一阶微分-梯度(非线性)
虽然是非线性的操作,但是求
g
x
,
g
y
g_x, g_y
gx,gy是线性操作,因此可以分开求解,最后做非线性的操作,如求开方和绝对值:
简单起见,直接修改原来的线性滤波函数,改成求绝对值:
def linear_filter(image, x, y, kernel, out):
sum_wf = 0
m = kernel.shape[0]
n = kernel.shape[1]
a = int((m - 1) / 2)
b = int((n - 1) / 2)
for s in range(-a, a + 1):
for t in range(-b, b + 1):
# convolution rotation 180
x_s = (x - s) if (x - s) in range(0, image.shape[0] - 1) else 0
y_t = (y - t) if (y - t) in range(0, image.shape[1] - 1) else 0
sum_wf += kernel[a + s][b + t] * image[x_s][y_t]
sum_wf = abs(sum_wf)
if sum_wf > 255:
sum_wf = 255
out[x][y] = int(sum_wf)
然后分别求 ∣ g x ∣ , ∣ g y ∣ |g_x|,|g_y| ∣gx∣,∣gy∣:
gradient_mask_1 = np.array([
[0, 0, 0],
[0, -1, 1],
[0, 0, 0],
])
gradient_mask_2 = np.array([
[0, 0, 0],
[0, -1, 0],
[0, 1, 0],
])
image_gradient_mask_1 = spatial_filtering(image, gradient_mask_1, linear_filter)
image_gradient_mask_2 = spatial_filtering(image, gradient_mask_1, linear_filter)
image_gradient_mask = image_gradient_mask_1 + image_gradient_mask_2
结果:
Roberts 算法 交叉差分
调用
roberts_mask_1 = np.array([
[0, 0, 0],
[0, -1, 0],
[0, 0, 1],
])
roberts_mask_2 = np.array([
[0, 0, 0],
[0, 0, -1],
[0, 1, 0],
])
image_soble_mask_1 = spatial_filtering(image, gradient_mask_1, linear_filter)
image_soble_mask_2 = spatial_filtering(image, gradient_mask_1, linear_filter)
image_soble_mask = image_soble_mask_1 + image_soble_mask_2
结果
Soble算子
调用:
soble_mask_1 = np.array([
[-1, -2, -1],
[0, 0, 0],
[1, 2, 1],
])
soble_mask_2 = np.array([
[-1, 0, 1],
[-2, 0, 2],
[-1, 0, 1],
])
image_soble_mask_1 = spatial_filtering(image, soble_mask_1, linear_filter)
image_soble_mask_2 = spatial_filtering(image, soble_mask_2, linear_filter)
image_soble_mask = image_soble_mask_1 + image_soble_mask_2
结果
最后都增加到原图中的效果:
可以看到:从梯度算子、Roberts 算子、Soble算子,效果依次增强