Recurrent Attention Dense Network for Single Image De-Raining

Guoqiang Chai, Zhao-Ba Wang, Guodong Guo, Youxing Chen, Yong Jin, Wei Wang, Xia ZhaoIEEE Access 2020 · OpenReview

Abstract

The problem of single image rain removal has attracted tremendous attention as the blurry images caused by rain streaks can degrade the performance of many computer vision algorithms. Although deep learning based de-raining methods have achieved a significant success, there are still unresolved issues in terms of the performance. In this work, we propose a novel recurrent attention dense network (RADN) for single image de-raining. In RADN, a region-level attention module is first utilized to identify rain streaks regions for the subsequent removal task. As rain streaks have different sizes and shapes, a modified densely connected convolutional network (DenseNet) with dilation convolutions and reduced channels is developed for an effective feature representation. The rain streaks are removed stage by stage and a Gate Recurrent Unit (GRU) is incorporated to deliver useful information from previous stages to later stages for a better performance. Qualitative and quantitative evaluations on both synthetic and real-world datasets demonstrate that the proposed approach can achieve a remarkable performance in comparison with the state-of-the-art methods for single image rain removal.

Introduction

Rain streaks are more than just an aesthetic nuisance; they significantly degrade the performance of computer vision systems used in autonomous driving, surveillance, and outdoor object detection. The challenge of Single Image De-Raining (SIRR) is "ill-posed" because, unlike video-based de-raining, there is no temporal information to help distinguish moving rain from static background objects. Existing deep learning solutions often suffer from complexity overload or detail loss, where background textures are inadvertently blurred or "ghost" artifacts remain.

Proposed Method: Recurrent Attention Dense Network (RADN)

The proposed RADN architecture addresses these challenges through three innovative pillars:

Region-level Direction-aware Attention Module (RDAM)

Instead of treating every pixel uniformly, RADN employs a region-level attention mechanism. It uses a two-round, four-directional Recurrent Neural Network (RNN) to generate an attention map. This map identifies the specific locations of rain streaks, allowing the network to focus its restorative efforts on those regions while preserving clear background areas.

The RDAM utilizes a four-directional RNN to capture spatial context for identifying rain regions.

Modified DenseNet (MDeNet)

The authors adapted the DenseNet architecture for image restoration by removing Batch Normalization (BN) layers, which were found to hinder generalization in de-raining. To maintain efficiency, channel counts were reduced to 1/4 of the original. Furthermore, dilated convolutions were utilized to expand the "receptive field" without losing pixel-level resolution, enabling the network to capture the context of long rain streaks.

Recurrent Removal with GRUs

Rain removal is performed progressively in four stages. A Gate Recurrent Unit (GRU) acts as a memory bridge, carrying useful information from previous stages to subsequent ones. This iterative approach allows the network to peel away complex rain layers stage-by-stage.

The final output mapping of the attention module identifies regions for restoration.

Experiments and Results

The model was evaluated on standard synthetic benchmarks (Rain100H, Rain100L, Rain800, Rain12) and real-world datasets.

  • Quantitative Performance: On the challenging Rain100H dataset, RADN achieved a PSNR of 29.506 dB and an SSIM of 0.902, significantly outperforming state-of-the-art methods like JORDER (26.541 dB / 0.835).
  • Qualitative Performance: Visual comparisons demonstrate that RADN is particularly effective at removing rain from sky regions and preserving fine details, such as car logos and intricate background structures, which other models often smudge.
  • Efficiency: RADN processes a 480x640 image in approximately 0.58 seconds, making it competitive with other deep learning models and significantly faster than traditional optimization-based methods.

Conclusion

The Recurrent Attention Dense Network (RADN) provides a robust framework for single image de-raining. By combining region-level attention with a modified DenseNet and recurrent processing, it effectively removes rain streaks while preserving image integrity. Future work will focus on integrating temporal consistency for video applications and improving performance on diverse real-world rain conditions.

Summary updated

Abstract

The problem of single image rain removal has attracted tremendous attention as the blurry images caused by rain streaks can degrade the performance of many computer vision algorithms. Although deep learning based de-raining methods have achieved a significant success, there are still unresolved issues in terms of the performance. In this work, we propose a novel recurrent attention dense network (RADN) for single image de-raining. In RADN, a region-level attention module is first utilized to identify rain streaks regions for the subsequent removal task. As rain streaks have different sizes and shapes, a modified densely connected convolutional network (DenseNet) with dilation convolutions and reduced channels is developed for an effective feature representation. The rain streaks are removed stage by stage and a Gate Recurrent Unit (GRU) is incorporated to deliver useful information from previous stages to later stages for a better performance. Qualitative and quantitative evaluations on both synthetic and real-world datasets demonstrate that the proposed approach can achieve a remarkable performance in comparison with the state-of-the-art methods for single image rain removal.
BibTeX
@inproceedings{chai2020recurrent,
  author = {Guoqiang Chai and Zhao-Ba Wang and Guodong Guo and Youxing Chen and Yong Jin and Wei Wang and Xia Zhao},
  title = {{Recurrent Attention Dense Network for Single Image De-Raining}},
  booktitle = {IEEE Access},
  year = {2020},
  url = {https://openreview.net/forum?id=sPCUUVMvgd},
}
Figure