MSE、PSNR 與資料範圍
選讀・IP 02・數位影像基礎 的補充資料(第 5/5 則)。主線不依賴這則補充;引用編號對應本頁底部的參考文獻。
導覽:本頁與本系列: 導覽:本頁與本系列
處理後的影像離原圖有多遠?最簡單的答案是把逐像素差的平方取平均。這個數字單獨看很難判讀(100 算大嗎?),所以我們拿它與像素能取到的最大值相比,並用對數尺度報告比值,就像聲音的分貝。
設 與 是兩張各有 個像素的影像。均方誤差(mean squared error,MSE)與峰值訊噪比(peak signal-to-noise ratio,PSNR)為
其中 與 是 處的像素值, 是資料範圍(data range):像素格式允許的最大值(8 位元 uint8 影像為 255,縮放到 的浮點影像為 1.0)。在 scikit-image 中,它就是 skimage.metrics.peak_signal_noise_ratio 的 data_range 參數。因為 ,它與 IP 08 使用的 是同一個量。
為什麼用分貝?誤差可以跨越很多個數量級,取對數後,相同的比值就是相同的步長。MSE 減半會增加 dB;雜訊標準差減半(MSE 變成四分之一)則增加約 6.02 dB。PSNR 越高,誤差越小。由於 在公式裡面,同樣的像素誤差在不同範圍下會得到不同的 PSNR,所以 PSNR 必須連同它的範圍一起才有意義。
程式碼 · Python · 21 行
import numpy as np
from skimage import color, data
from skimage.metrics import mean_squared_error, peak_signal_noise_ratio
ref = data.camera() # uint8, values 0..255
rng = np.random.default_rng(0)
for sigma in (10, 20):
noisy = np.clip(ref + rng.normal(0, sigma, ref.shape), 0, 255).astype(np.uint8)
mse = mean_squared_error(ref, noisy)
print(f"sigma={sigma}: MSE={mse:.1f}, by hand {10 * np.log10(255**2 / mse):.2f} dB, "
f"uint8 {peak_signal_noise_ratio(ref, noisy):.2f} dB")
ref_f, noisy_f = ref / 255.0, noisy / 255.0 # the same pair as floats in [0, 1]
print(f" float, data_range=1.0: {peak_signal_noise_ratio(ref_f, noisy_f, data_range=1.0):.2f} dB")
print(f" float, data_range=255: {peak_signal_noise_ratio(ref_f, noisy_f, data_range=255):.2f} dB (wrong)")
# Pitfall: the same noise on RGB and on luma Y gives different PSNR
rgb = data.astronaut()
noisy_rgb = np.clip(rgb + rng.normal(0, 10, rgb.shape), 0, 255).astype(np.uint8)
y = lambda im: color.rgb2gray(im) * 255 # luma in [0, 255]
print(f"RGB: {peak_signal_noise_ratio(rgb, noisy_rgb):.2f} dB, "
f"Y: {peak_signal_noise_ratio(y(rgb), y(noisy_rgb), data_range=255):.2f} dB")執行後印出:
程式碼 · 7 行
sigma=10: MSE=97.8, by hand 28.23 dB, uint8 28.23 dB
float, data_range=1.0: 28.23 dB
float, data_range=255: 76.36 dB (wrong)
sigma=20: MSE=374.0, by hand 22.40 dB, uint8 22.40 dB
float, data_range=1.0: 22.40 dB
float, data_range=255: 70.53 dB (wrong)
RGB: 28.61 dB, Y: 31.03 dB解讀這些數字:當範圍與數值尺度一致時,uint8 與浮點兩種形式給出相同的 28.23 dB;但若宣告錯誤的範圍(對 的浮點影像用 255),會多出 dB,同一張影像就被報告成驚人的 76.36 dB。把 加倍,名義上 MSE 變成四倍(下降 6.02 dB);實測下降 dB,略少一些,因為把數值截斷(clipping)到 會在飽和的像素上移除一部分加入的誤差。最後一行顯示第三個陷阱:同樣的雜訊,在 RGB 上是 28.61 dB,在亮度通道 Y 上卻是 31.03 dB,因為 Y 是三個獨立含雜訊通道的平均。比較 PSNR 時,影像、範圍與通道(RGB 或 Y)都必須相同;整數影像也要先轉成浮點再相減,否則 uint8 運算會溢位繞回(scikit-image 內部已處理這點)。
最後,PSNR 衡量的是像素誤差,而不是外觀。PSNR 相同的兩張影像看起來可能差很多,而考慮結構的指標如 SSIM [13],或學習得到的感知距離 [15],與人的判斷較為吻合;見上面的「現代觀點」。Huynh-Thu 與 Ghanbari [19] 討論了 PSNR 作為品質指標在什麼範圍內有效。
參考文獻
- R. C. Gonzalez and R. E. Woods, Digital Image Processing, 4th ed., Pearson, 2018, Ch. 2. publisher page
- M. Unser, “Sampling—50 years after Shannon,” Proceedings of the IEEE, vol. 88, no. 4, pp. 569–587, 2000. doi
- R. Keys, “Cubic convolution interpolation for digital image processing,” IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. 29, no. 6, pp. 1153–1160, 1981. doi
- P. Thévenaz, T. Blu, and M. Unser, “Interpolation revisited,” IEEE Transactions on Medical Imaging, vol. 19, no. 7, pp. 739–758, 2000. doi
- A. Rosenfeld and J. L. Pfaltz, “Distance functions on digital pictures,” Pattern Recognition, vol. 1, pp. 33–61, 1968. doi
- B. Zitová and J. Flusser, “Image registration methods: a survey,” Image and Vision Computing, vol. 21, no. 11, pp. 977–1000, 2003. doi
- M. Delbracio, D. Kelly, M. S. Brown, and P. Milanfar, “Mobile computational photography: A tour,” arXiv:2102.09000, 2021. arXiv
- R. Ramanath, W. E. Snyder, Y. Yoo, and M. S. Drew, “Color image processing pipeline,” IEEE Signal Processing Magazine, vol. 22, no. 1, pp. 34–43, 2005. doi
- Z. Wang, J. Chen, and S. C. H. Hoi, “Deep learning for image super-resolution: A survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. arXiv
- B. Moser, F. Raue, S. Frolov, J. Hees, S. Palacio, and A. Dengel, “Hitchhiker’s guide to super-resolution: Introduction and recent advances,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. arXiv
- Y. Chen, S. Liu, and X. Wang, “Learning continuous image representation with local implicit image function,” CVPR, 2021. arXiv
- X. Wang, L. Xie, C. Dong, and Y. Shan, “Real-ESRGAN: Training real-world blind super-resolution with pure synthetic data,” arXiv:2107.10833, 2021. arXiv
- Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: From error visibility to structural similarity,” IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600–612, 2004. project page
- G. Zhai and X. Min, “Perceptual image quality assessment: A survey,” Science China Information Sciences, vol. 63, no. 11, 2020. doi
- R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” CVPR, 2018. arXiv
- K. Ding, K. Ma, S. Wang, and E. P. Simoncelli, “Comparison of full-reference image quality models for optimization of image processing systems,” International Journal of Computer Vision, 2021. arXiv
- C. Chen, Q. Chen, J. Xu, and V. Koltun, “Learning to see in the dark,” CVPR, 2018. arXiv
- A. Ignatov, L. Van Gool, and R. Timofte, “Replacing mobile camera ISP with a single deep learning model,” arXiv:2002.05509, 2020. arXiv
- Q. Huynh-Thu and M. Ghanbari, “Scope of validity of PSNR in image/video quality assessment,” Electronics Letters, vol. 44, no. 13, pp. 800–801, 2008. doi
- R. Hartley and A. Zisserman, Multiple View Geometry in Computer Vision, 2nd ed., Cambridge University Press, 2004. doi
- scikit-image contributors, “ProjectiveTransform.estimate,” scikit-image 0.25.2 documentation, accessed 5 November 2026. API 文件.