研究目的
To solve the problem of image denoising when the noise level is unknown, handling spatially varying Gaussian noise and preserving fine texture details.
研究成果
RDS-Denoiser effectively handles unknown and spatially varying noise levels, outperforming existing methods in PSNR and detail preservation. RDS-GAN further improves visual quality, especially in recovering sharp edges and handling high noise cases. The hierarchical dense connections and two-stage approach contribute to robust denoising, with potential for real-time applications.
研究不足
Discriminatively trained models may overfit to training data, oversmooth edges at high noise levels, and fail on natural noisy images that do not satisfy Gaussian distribution. GANs can introduce artifacts. The method is sensitive to noise types and may not generalize well to all real-world scenarios.
1：Experimental Design and Method Selection:
The methodology involves a two-stage stacked network: Stage-I uses RDS-Denoiser (a densely connected convolutional neural network) to predict the noise map from a noisy image, and Stage-II uses RDS-GAN (a conditional generative adversarial network) to enhance denoising results. Theoretical models include convolutional layers, ReLU activations, RDS blocks with dense connections, scaling layers, and residual connections.
2：Sample Selection and Data Sources:
Training uses the DIV2K dataset with 800 color images, split into 2231x128 patches, and adds Gaussian noise with standard deviation from 0 to
3：Testing uses public datasets:
BSD68, CBSD68, Kodak24, McMaster, Set14, and RNI
4：List of Experimental Equipment and Materials:
A computer with GPU (1080Ti mentioned) for implementation using TensorFlow.
5：Experimental Procedures and Operational Workflow:
Patches of size 55x55x3 are processed; Adam optimizer is used for training with weight decay and learning rate adjustments. For RDS-GAN, the generator input is the concatenation of noisy image and predicted noise map, and perceptual loss is applied.
6：Data Analysis Methods:
Performance is evaluated using PSNR (Peak Signal-to-Noise Ratio) and qualitative comparisons with state-of-the-art methods.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
