几种计算图像/向量相似度的指标（实现）

from skimage.metrics import peak_signal_noise_ratio as cal_psnr  
from skimage.metrics import structural_similarity as cal_ssim

from skimage.io import imread  
import numpy as np

img1_path = r'imgpath\1.jpg' # RGB图像
img2_path = r'imgpath\2.jpg' # RGB图像

PSNR 峰值信噪比：越大越相似

img1 = imread(img1_path)  
img2 = imread(img2_path)
psnr = cal_psnr(img1, img2) 
print(f'PSNR: {psnr:.4f} dB')

SSIM 结构相似性指数：越大越相似

img1 = imread(img1_path)  
img2 = imread(img2_path)
ssim = cal_ssim(img1, img2, multichannel=True, channel_axis=2) 
print(f'SSIM: {ssim:.4f}')

MSE 均方误差：越小越相似

img1 = imread(img1_path)
img2 = imread(img2_path)
mse = np.mean((img1-img2) ** 2)
print("mse:", mse)

cosine similarity 余弦相似度：越大越相似

img1 = imread(img1_path)
img2 = imread(img2_path)
v1 = img1.flatten().astype(np.float32) # imread的数据类型为uint8,在下面dot计算时会发生上溢
v2 = img2.flatten().astype(np.float32)
c_s = np.dot(v1, v2) / (np.linalg.norm(v1) * np.linalg.norm(v2))
print(f'cosine similarity: {c_s:.3f}')

卡方距离：越小越相似

img1 = imread(img1_path)
img2 = imread(img2_path)
# 计算直方图
hist1, _ = np.histogram(img1.flatten(), bins=256, range=[0,256])
hist2, _ = np.histogram(img2.flatten(), bins=256, range=[0,256])
chi_square_distance = np.sum((hist1 - hist2) ** 2 / (hist1 + hist2 + 1e-9)) / 2.0
print("Chi Square Distance:", chi_square_distance)

实际使用时需要转化为 tensor 进行GPU加速计算