用python实现基于概率分布的超声图像与自然图像性质差异分析,不用greycomatrix和greycoprops,使用opencv,两种图像的文件类型都是jpg,超声图像的文件路径是D:\zzz\us2,自然图像的文件路径是D:\zzz\na2,两种图像都有84张,图片形状都不同,图像名称是随机的,需要将图像归一化,需要对比两种图像的特征必须有颜色,纹理,形状,差异性分析方法也需要多种,包括完整详细最新的代码
时间: 2023-10-03 14:02:18 浏览: 89
以下是基于概率分布的超声图像与自然图像性质差异分析的Python代码,使用了OpenCV库和一些其他的科学计算库:
```python
import cv2
import numpy as np
from scipy.stats import entropy
from scipy.spatial.distance import jensenshannon
# 设置超声图像和自然图像文件路径
us_path = "D:/zzz/us2"
na_path = "D:/zzz/na2"
# 定义一个函数用于计算图像的颜色直方图
def calc_color_hist(img):
# 将图像转换为HSV颜色空间
hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)
# 计算H、S、V三个通道的直方图
hist_h = cv2.calcHist([hsv], [0], None, [180], [0, 180])
hist_s = cv2.calcHist([hsv], [1], None, [256], [0, 256])
hist_v = cv2.calcHist([hsv], [2], None, [256], [0, 256])
# 将三个通道的直方图拼接成一个特征向量
hist = np.concatenate((hist_h, hist_s, hist_v)).ravel()
# 归一化特征向量
hist = hist / np.sum(hist)
return hist
# 定义一个函数用于计算图像的灰度共生矩阵
def calc_glcm(img, d=1, theta=0):
# 将图像转换为灰度图像
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# 计算灰度共生矩阵
glcm = cv2.calcGLCM(gray, [d], [theta], 256, symmetric=True)
# 归一化灰度共生矩阵
glcm = glcm / np.sum(glcm)
return glcm
# 定义一个函数用于计算图像的纹理特征
def calc_texture_features(img):
# 计算四个方向上的灰度共生矩阵
glcm_0 = calc_glcm(img, 1, 0)
glcm_45 = calc_glcm(img, 1, 45)
glcm_90 = calc_glcm(img, 1, 90)
glcm_135 = calc_glcm(img, 1, 135)
# 计算四个方向上的对比度、能量、熵和相关度
contrast_0 = np.sum(np.square(glcm_0 - np.mean(glcm_0)))
energy_0 = np.sum(np.square(glcm_0))
entropy_0 = entropy(glcm_0.ravel())
correlation_0 = np.sum((glcm_0 - np.mean(glcm_0)) * (np.arange(256) - np.mean(np.arange(256)))) / (np.std(glcm_0) * np.std(np.arange(256)))
contrast_45 = np.sum(np.square(glcm_45 - np.mean(glcm_45)))
energy_45 = np.sum(np.square(glcm_45))
entropy_45 = entropy(glcm_45.ravel())
correlation_45 = np.sum((glcm_45 - np.mean(glcm_45)) * (np.arange(256) - np.mean(np.arange(256)))) / (np.std(glcm_45) * np.std(np.arange(256)))
contrast_90 = np.sum(np.square(glcm_90 - np.mean(glcm_90)))
energy_90 = np.sum(np.square(glcm_90))
entropy_90 = entropy(glcm_90.ravel())
correlation_90 = np.sum((glcm_90 - np.mean(glcm_90)) * (np.arange(256) - np.mean(np.arange(256)))) / (np.std(glcm_90) * np.std(np.arange(256)))
contrast_135 = np.sum(np.square(glcm_135 - np.mean(glcm_135)))
energy_135 = np.sum(np.square(glcm_135))
entropy_135 = entropy(glcm_135.ravel())
correlation_135 = np.sum((glcm_135 - np.mean(glcm_135)) * (np.arange(256) - np.mean(np.arange(256)))) / (np.std(glcm_135) * np.std(np.arange(256)))
# 将四个方向上的特征拼接成一个特征向量
features = np.array([contrast_0, energy_0, entropy_0, correlation_0,
contrast_45, energy_45, entropy_45, correlation_45,
contrast_90, energy_90, entropy_90, correlation_90,
contrast_135, energy_135, entropy_135, correlation_135])
# 归一化特征向量
features = features / np.sum(features)
return features
# 定义一个函数用于计算图像的形状特征
def calc_shape_features(img):
# 将图像转换为灰度图像
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# 二值化图像
_, binary = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)
# 计算轮廓
contours, _ = cv2.findContours(binary, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
# 提取最大轮廓
max_contour = max(contours, key=cv2.contourArea)
# 计算最大轮廓的面积、周长、宽高比和圆形度
area = cv2.contourArea(max_contour)
perimeter = cv2.arcLength(max_contour, True)
x, y, w, h = cv2.boundingRect(max_contour)
aspect_ratio = float(w) / h
circularity = (4 * np.pi * area) / (perimeter ** 2)
# 将四个形状特征拼接成一个特征向量
features = np.array([area, perimeter, aspect_ratio, circularity])
# 归一化特征向量
features = features / np.sum(features)
return features
# 定义一个函数用于计算两个图像的相似度
def calc_similarity(img1, img2):
# 计算两个图像的颜色直方图
hist1 = calc_color_hist(img1)
hist2 = calc_color_hist(img2)
# 计算两个颜色直方图的Jensen-Shannon距离
color_distance = jensenshannon(hist1, hist2)
# 计算两个图像的纹理特征
texture1 = calc_texture_features(img1)
texture2 = calc_texture_features(img2)
# 计算两个纹理特征的欧氏距离
texture_distance = np.linalg.norm(texture1 - texture2)
# 计算两个图像的形状特征
shape1 = calc_shape_features(img1)
shape2 = calc_shape_features(img2)
# 计算两个形状特征的欧氏距离
shape_distance = np.linalg.norm(shape1 - shape2)
# 计算三个距离的加权平均作为相似度
similarity = 1 / (1 + color_distance + texture_distance + shape_distance)
return similarity
# 加载所有的超声图像和自然图像
us_images = []
na_images = []
for i in range(1, 85):
us_filename = us_path + "/{}.jpg".format(i)
na_filename = na_path + "/{}.jpg".format(i)
us_img = cv2.imread(us_filename)
na_img = cv2.imread(na_filename)
# 将图像缩放到统一的大小
us_img = cv2.resize(us_img, (300, 300))
na_img = cv2.resize(na_img, (300, 300))
# 将图像归一化到[0, 1]范围
us_img = us_img / 255.0
na_img = na_img / 255.0
us_images.append(us_img)
na_images.append(na_img)
# 计算所有超声图像两两之间的相似度,得到一个84x84的相似度矩阵
us_similarity_matrix = np.zeros((84, 84))
for i in range(84):
for j in range(i, 84):
similarity = calc_similarity(us_images[i], us_images[j])
us_similarity_matrix[i, j] = similarity
us_similarity_matrix[j, i] = similarity
# 计算所有自然图像两两之间的相似度,得到一个84x84的相似度矩阵
na_similarity_matrix = np.zeros((84, 84))
for i in range(84):
for j in range(i, 84):
similarity = calc_similarity(na_images[i], na_images[j])
na_similarity_matrix[i, j] = similarity
na_similarity_matrix[j, i] = similarity
# 计算所有超声图像和自然图像之间的相似度,得到一个84x84的相似度矩阵
all_similarity_matrix = np.zeros((84, 84))
for i in range(84):
for j in range(84):
similarity = calc_similarity(us_images[i], na_images[j])
all_similarity_matrix[i, j] = similarity
# 打印所有相似度矩阵的形状和前5x5的子矩阵
print("US similarity matrix shape:", us_similarity_matrix.shape)
print(us_similarity_matrix[:5, :5])
print("NA similarity matrix shape:", na_similarity_matrix.shape)
print(na_similarity_matrix[:5, :5])
print("All similarity matrix shape:", all_similarity_matrix.shape)
print(all_similarity_matrix[:5, :5])
```
这个代码中,我们定义了三个函数:`calc_color_hist`用于计算图像的颜色直方图,`calc_glcm`用于计算图像的灰度共生矩阵,`calc_texture_features`用于计算图像的纹理特征,`calc_shape_features`用于计算图像的形状特征,`calc_similarity`用于计算两个图像之间的相似度。然后,我们加载了所有的超声图像和自然图像,计算了所有图像之间的相似度,得到了三个相似度矩阵。最后,我们打印了所有相似度矩阵的形状和前5x5的子矩阵。
这个代码中使用的相似度计算方法比较简单,只计算了颜色、纹理和形状三个方面的特征,并且使用了Jensen-Shannon距离和欧氏距离来计算相似度。如果需要更加准确的相似度计算方法,可以考虑使用更多的特征和更复杂的距离度量方法。
阅读全文