Source-linked AI summary
Progressive Image Deraining Networks: A Better and Simpler Baseline
Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, Deyu Meng
TL;DR
Complex and diverse deraining networks make module contributions difficult to analyze. This paper develops simpler progressive and recurrent ResNet baselines with recursive computation, tailored inputs and outputs, and single losses; experiments find favorable performance on synthetic and real rainy images. The models are presented as efficient baselines for future deraining research.
Problem
Increasingly complex and diverse deep deraining networks make it difficult to analyze the contributions of their modules.
Method
The paper repeatedly unfolds a shallow ResNet into PRN, adds recurrent cross-stage features for PReNet, and uses stage-wise results plus the rainy image with residual prediction and single losses.
Results
PRN and PReNet are computationally efficient and perform favorably against state-of-the-art methods on synthetic and real rainy images.
Takeaways & Limitations
Their simplicity, effectiveness, and efficiency make PRN and PReNet suitable baseline networks for future image-deraining studies.
Takeaways & Limitations
Ablation studies exclude 546 of 1,800 training images because they share background contents with the Rain100H test images, leaving 1,254 training images.
Abstract
from arXiv · showhide
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.
1. Introduction
The paper proposes simple progressive deraining baselines to address increasingly complex network designs, combining recursive shallow ResNets with recurrent features, input/output choices, and single losses. PRN and PReNet achieve favorable, efficient performance on synthetic and real rainy images.
- Increasingly complicated and diverse deraining networks make it difficult to analyze individual module contributions.
- PRN recursively unfolds a shallow five-ResBlock ResNet into multiple stages without increasing model parameters.PReNet adds a recurrent layer to exploit dependencies across recursive stages.
- Taking both stage-wise results and the original rainy image as input to each ResNet improves deraining performance.The networks predict the rain-streak residual rather than directly outputting the clean image.
- PRN and PReNet can be trained with a single negative SSIM or MSE loss instead of carefully tuned hybrid losses.
- PReNet and PRN are computationally efficient and outperform state-of-the-art methods across four synthetic datasets, with visually pleasing results on real images and videos.Performance gains remain significant on the heavy-rain Rain100H dataset.
- Intra-stage recursive computation reduces parameters while maintaining state-of-the-art deraining performance with graceful degradation.
2. Related Work
Prior deraining research uses optimization-based layer separation and increasingly elaborate deep architectures. The paper positions PRN and PReNet as simpler alternatives that reduce parameters without obvious performance loss.
- Optimization-based methods model rainy images as compositions of clean background and rain layers, using regularizers and specific algorithms for separation.
- Deep deraining methods have explored detail-layer prediction, residual learning, joint rain detection, and rain-density estimation.
- Deep architectures combine modules such as residual, dilated, dense, squeeze-and-excitation, recurrent, multi-stream, and multi-stage components.
- Lightweight cascaded and Laplacian-pyramid networks improve computational efficiency but incur obvious deraining-performance degradation.
- PRNr and PReNetr use intra-stage recursive computation to reduce network parameters while maintaining state-of-the-art deraining performance.
3. Progressive Image Deraining Networks
The paper builds progressive deraining networks by recursively reusing a shallow ResNet across stages, adding recurrent feature propagation in PReNet, and simplifying inputs, outputs, and training losses. Recursive computation reduces parameters, while experiments support single-loss training and intermediate-stage supervision for earlier visual results.
- Progressive network framework: PRN recursively unfolds a shallow ResNet across multiple stages while reusing stage-invariant parameters.Each stage uses convolutional input and output layers around residual blocks; the shared architecture avoids increasing parameters across stages.
- Input and output: Including both the current stage estimate and the original rainy image as input improves deraining performance, while the network predicts either rain or clean-image layers.The paper adopts residual learning by predicting the rain streak layer, although background-image prediction is also possible.
- Progressive network framework: PReNet extends PRN with a recurrent state that propagates deep-feature dependencies across stages.Its recurrent layer receives the intermediate representation and previous recurrent state; the implementation uses an LSTM.
- Network architectures: Intra-stage recursive unfolding of one ResBlock five times produces PRNr and PReNetr, substantially reducing model size with an evaluated tradeoff between size and deraining performance.The conventional implementation instead uses five ResBlocks, while recursive ResBlocks share computation within a stage.
- Learning objective: A single MSE or negative SSIM loss is sufficient for training PRN and PReNet, and negative SSIM empirically outperforms MSE on both PSNR and SSIM.Recursive supervision can improve visual quality at early stages, but does not provide a performance gain when supervision is applied at the final stage.
4. Experimental Results
Experiments evaluate progressive networks through ablations, synthetic benchmarks, efficiency tests, and real rainy images and video. PReNet generally improves deraining quality, while recursive variants reduce parameters and retain strong visual performance.
- Loss Functions: A single negative SSIM loss outperforms MSE for PReNet on Rain100H in both SSIM and PSNR, with more visually plausible results.Recursive negative SSIM supervision performs moderately inferior to final-stage supervision, although their visual results are indistinguishable.
- Loss Functions: PReNet-RecSSIM produces much better intermediate results than PReNet-SSIM, enabling inference to stop at any stage in resource-constrained environments.
- Recurrent Layer: LSTM performs slightly better than GRU, and introducing a recurrent layer improves PReNet over PRN in PSNR and SSIM.
- Intra-stage Recursion: Intra-stage recursive ResBlocks significantly reduce parameters, with graceful degradation in average PSNR and SSIM and comparable visual quality.The resulting models are PRNr and PReNetr.
- Recursive Stage Number: Adding stages from T = 2 to T = 6 usually raises average PSNR and SSIM, whereas T = 7 performs slightly worse and is harder to train.The experiments therefore use T = 6 in subsequent evaluations.
- Network Input and Output: PReNet outperforms its input-restricted variant in PSNR and SSIM, while residual learning further contributes to performance gains.The full PReNet takes the original rainy image and the stage-wise result as inputs.
- Comparisons: Across synthetic datasets, progressive networks outperform competing methods, run efficiently, and produce visually improved results on real rainy images and videos.PReNet also avoids the frame-switching and accumulation errors observed for FastDerain in some video frames.
5. Conclusion
The paper concludes that simple recursive combinations of shallow ResNets can provide effective deraining baselines. Recurrent and input-design additions improve PReNet, while recursive computation can reduce parameters and support strong synthetic and real-image performance.
- Conclusion: PRN combines a shallow ResNet with multi-stage recursion to achieve favorable deraining performance without relying on deeper, more complex networks.
- Conclusion: PReNet further boosts performance by adding a recurrent layer and using the stage-wise result as input to each ResNet.
- Conclusion: Inter- and intra-stage recursive computation reduces network parameters in PRNr and PReNetr, with extensive experiments supporting their effectiveness on synthetic and real rainy images.
- Conclusion: The models' simplicity, effectiveness, and efficiency make them appealing baselines for developing new deraining networks.