Progressive Image Deraining Networks: A Better and Simpler Baseline (1901.09221v3)

Abstract: Along with the deraining performance improvement of deep networks, their structures and learning become more and more complicated and diverse, making it difficult to analyze the contribution of various network modules when developing new deraining networks. To handle this issue, this paper provides a better and simpler baseline deraining network by considering network architecture, input and output, and loss functions. Specifically, by repeatedly unfolding a shallow ResNet, progressive ResNet (PRN) is proposed to take advantage of recursive computation. A recurrent layer is further introduced to exploit the dependencies of deep features across stages, forming our progressive recurrent network (PReNet). Furthermore, intra-stage recursive computation of ResNet can be adopted in PRN and PReNet to notably reduce network parameters with graceful degradation in deraining performance. For network input and output, we take both stage-wise result and original rainy image as input to each ResNet and finally output the prediction of {residual image}. As for loss functions, single MSE or negative SSIM losses are sufficient to train PRN and PReNet. Experiments show that PRN and PReNet perform favorably on both synthetic and real rainy images. Considering its simplicity, efficiency and effectiveness, our models are expected to serve as a suitable baseline in future deraining research. The source codes are available at https://github.com/csdwren/PReNet.

• The paper presents PRN and PReNet, two progressive networks that use intra-stage recursion to reduce parameters while effectively removing rain streaks.
• It validates the models on synthetic and real-world datasets, achieving significant improvements in PSNR and SSIM metrics.
• The simplified architectures and efficient training methods offer practical benefits for real-world applications and future image restoration research.

Progressive Image Deraining Networks: A Synopsis

The paper "Progressive Image Deraining Networks: A Better and Simpler Baseline" presents a methodological advancement in the field of single image deraining using deep convolutional neural networks. The key innovation lies in the development of Progressive Residual Network (PRN) and Progressive Recurrent Network (PReNet), both designed to address the complexities of existing deep learning models for image deraining by leveraging simpler architectures and recursive computations.

Key Contributions

1. Progressive Residual Network (PRN): The PRN model unfolds a shallow ResNet into multiple stages, allowing recursive computation without increasing the number of model parameters. This recursion enables efficient feature extraction and rain streak removal over successive stages.
2. Progressive Recurrent Network (PReNet): Extending PRN, PReNet introduces a recurrent layer to capture dependencies of features across different stages using a convolutional LSTM. This addition capitalizes on temporal feature propagation, enhancing the model's capability to progressively remove rain streaks.
3. Intra-stage Recursion: The paper suggests leveraging intra-stage recursive computation wherein a single ResBlock is recursively unfolded multiple times within each stage. This approach significantly reduces network parameters while maintaining robust deraining performance.
4. Training Objectives: The paper posits that effective deraining performance can be achieved using simple loss functions such as single mean squared error (MSE) or negative structural similarity index measure (SSIM) loss. This deviation from complex hybrid losses facilitates easier and more efficient training of the networks.

Experimental Validation

The models were rigorously evaluated on synthetic datasets, namely Rain100H, Rain100L, and Rain12, as well as a real-world dataset, Rain1400. Here are the notable outcomes:

• Performance Metrics: The PReNet model consistently outperforms state-of-the-art methods in both quantitative (PSNR and SSIM) and qualitative assessments on synthetic datasets. Significant PSNR and SSIM gains were observed, particularly on the challenging Rain100H dataset.
• Efficiency: The computational efficiency of PRN and PReNet models is underscored by their relatively low parameter counts and reduced running times compared to other deep learning-based methods.
• Real-world Applicability: The application of PReNet to real rainy images and videos demonstrates its superior performance in practical scenarios. PReNet exhibits robust rain streak removal and generates visually appealing results without being constrained by the temporal continuity of video frames.

Implications and Future Directions

The simplicity and efficacy of PRN and PReNet suggest they provide a solid baseline for future research in image deraining. Here are the implications and prospective research directions:

• Simplified Network Design: The proposed models show that effective deraining does not necessarily require deep or complex architectures. Future research might focus on exploring similar progressive or recursive designs in other low-level vision tasks.
• Enhanced Training Efficiency: With the capability to perform well using simple loss functions, there is the potential for broader adoption in applications where computational resources or large-scale data availability is limited.
• Real-world Deployment: The models' demonstrated competence on real-world data indicates a pathway towards practical implementations in surveillance systems, autonomous driving, and other computer vision applications impacted by adverse weather conditions.
• Extensions to Other Degradation Problems: The progressive methodologies introduced might be adapted to related problems such as image dehazing, denoising, or other restoration tasks, where gradual refinement is beneficial.

In summary, this paper lays the groundwork for a new direction in image deraining research, focusing on progressive refinement and efficiency. Its contributions extend beyond the immediate task of deraining to potentially influence a wider array of image restoration and enhancement challenges.