Low-quality watermarked face inpainting with discriminative residual learning

摘要: Most existing image inpainting methods assume that the location of repair area (watermark) is known, but this assumption does not always hold. In addition, actual watermarked face in a compressed low-quality form, which very disadvantageous to due compression distortion effects. To address these issues, paper proposes method based on joint residual learning with cooperative discriminant network. We first employ global and facial features local render clean clear faces under unknown watermark positions. Because process may distort genuine face, we further propose discriminative constraint network maintain fidelity repaired faces. Experimentally, average PSNR inpainted images increased by 4.16dB, SSIM 0.08. TPR improved 16.96% when FPR 10% verification.