A Generative Adversarial Network to Denoise Depth Maps for Quality Improvement of DIBR-Synthesized Stereoscopic Images

Abstract

Depth map quality is an important factor that affects the quality of synthesized stereoscopic images in stereoscopic visual communication systems using the depth image-based rendering (DIBR) technique. This paper proposes a method using a generative adversarial network (GAN) to denoise depth maps corrupted by several types of distortion. The generative network of the proposed GAN builds on convolutional layers, residual layers, and transposed convolutional layers with symmetric skip connections. The discriminative network of the proposed GAN is designed as a convolutional neural network. The generative network for denoising depth maps is trained with cropped depth maps where distortion is applied. Objective and subjective assessment of denoised depth maps and DIBR-synthesized stereoscopic images demonstrate that the proposed GAN effectively reduces the distortion in the depth maps and improves the quality of DIBR-synthesized stereoscopic images.