Total Variation Regularization Algorithms for Images Corrupted with Different Noise Models: A Review

Figure 1

Deconvolution with the ℓ²-TV and the nonnegative ℓ²-TV (via [81, 87] resp.) methods for the grayscale “Goldhill” (720×576) image and the color “Lena” (512×512) image.

(a) Blurred grayscale “Goldhill” image corrupted with 𝜎=0.2 Gaussian noise

(b) ℓ²-TV reconstruction with regularization parameter 𝜆=0.1 (SNR: 11.59 dB)

(c) Nonnegative ℓ²-TV deconvolution with regularization parameter 𝜆=0.1 (SNR: 12.12 dB)

(d) Blurred color “Lena” image corrupted with 𝜎=0.2 Gaussian noise

(e) ℓ²-TV reconstruction with regularization parameter 𝜆=0.1 (SNR: 13.02 dB)

(f) Nonnegative ℓ²-TV deconvolution with regularization parameter 𝜆=0.1 (SNR: 13.15 dB)