基于生成对抗网络的深度图像去噪

DEPTH IMAGE DENOISING BASED ON GENERATIVE ADVERSARIAL NETWORK

深度图被广泛应用于三维重建等领域,然而,由深度相机捕获的深度图会产生各种类型的失真,这使得从深度图中准确估计深度信息变得困难。针对低质量深度图中存在的各种类型的噪声,提出一种基于生成对抗网络的深度图像去噪算法。生成对抗网络由生成网络和判别网络组成。在生成网络中引入残差网络,避免模型退化问题,使用跳跃连接,加快网络训练速度同时保证图像细节的有效传递;在判别网络中使用步幅卷积代替池化层,减少模型的计算量;通过优化模型的训练,使得生成的深度图像更加清晰。实验结果表明,该算法能够生成效果更好的深度图,在主观视觉和客观评价方面均优于其他算法。

Depth image is widely used in 3D reconstruction and other fields.However,depth image captured by depth camera will produce various types of distortion,which makes it difficult to accurately estimate depth information from depth image.Aimed at various types of noise in low-quality depth image,a depth image denoising algorithm based on generative adversarial network is proposed.The generative adversarial network consisted of a generative network and a discriminative network.The residual network was introduced in the generative network to avoid the problem of model degradation.The skip connection was used to speed up the network training while ensuring the effective transmission of image details.The step convolution was used instead of pooling layer in the discriminative network to reduce the computational burden of the model.Through the training of optimization model,the depth image generated was more clear.Experimental results show that the proposed algorithm can generate better depth image and is superior to other algorithms in subjective vision and objective evaluation.