DCSFN: Deep Cross-scale Fusion Network for Single Image Rain Removal

• Rain removal is ill-posed because many background-rain pairs can explain an input, while prior assumptions may fail for varied rain.The cited priors assume particular rain shapes, directions, and sizes.Source: §1 · p. 1
• Single-network training without cross-scale relationships may cause information drop-out despite strong results from existing methods.Source: §1 · p. 2
• DCSFN combines inner-scale feature connections, densely connected encoder-decoder blocks with skip connections, and GRU-based cross-scale fusion.The design aims to improve rain-feature learning and use information across scales.Source: §1 · p. 2
• The proposed method reports state-of-the-art results on synthetic and real-world datasets.Source: §1 · p. 2

2.1 Single Image Deraining

• Single-image deraining work includes prior-based separation and deep-learning methods using residual, multi-scale, attention, density-guided, and contextual designs.The review also describes a benchmark spanning synthetic and real-world rainy images.Source: §2.1 · p. 2

2.2 Multi-scale Learning

• Multi-scale architectures provide features at different scales, but existing work still leaves cross-scale information insufficiently explored.The proposed network uses three scale-specific subnetworks, GRU fusion at encoder ends, and a final fusion to estimate rain.Source: §2.2 · p. 2, Figure 1 · p. 3

3.1 Proposed Network (DCSFN)

• Three subnetworks learn features at different scales, which GRUs fuse at each encoder’s end before the network estimates rain and produces a rain-free image.The network uses the fused multiscale features to estimate rain.Source: §3.1 · p. 3

3.2 Encoder and Decoder with Dense Connection

• The densely connected encoder and decoder use skip connections to support long-range spatial dependencies and reuse proceeding-layer activations.Dense connections are described as maximizing information and gradient flow, strengthening feature propagation, and encouraging feature reuse.Source: §3.2 · p. 3

3.3 Inner-scale Connection Block

• The Inner-scale Connection Block pools multiscale features, connects them through convolutions to explore cross-scale correlations, then fuses them with residual learning.The paper presents the block as a way to explore inner correlations between scales.Source: §3.3 · p. 4

3.4 Loss Function

• The network uses SSIM loss, comparing the de-rained image with its corresponding ground-truth background.The stated loss is the negative SSIM of those images.Source: §3.4 · p. 4

4 EXPERIMENTS RESULTS

• Evaluation covers four widely used synthetic datasets and real-world images.The paper describes the experiments as extensive.Source: §4 · p. 4
• The comparisons include six methods: DDN, RESCAN, NLEDN, REHEN, PreNet, and SSIR.The passage identifies these as state-of-the-art methods and gives their publication venues and years.Source: §4 · p. 4
• The experimental section includes results on synthetic datasets, results on real-world datasets, and an ablation study.It also introduces datasets, measurements, and implementation details.Source: §4 · p. 4

4.1.1 Synthetic Datasets.

• Rain100L, Rain100H, and Rain1200 are the three synthetic benchmarks, and Table 1 reports quantitative experiments on them.Rain100L and Rain100H each have 1,800 training pairs and 200 test pairs; Rain1200 has 12,000 training images and 1,200 test images.Source: §4.1.1 · pp. 4–5, Table 1 · p. 5
• Rain100L has light rain, while Rain100H contains heavy rain with varied shapes, directions, and sizes and is described as the most challenging dataset.Source: §4.1.1 · p. 4

• Real-world rainy images were collected from the Internet and selected from.Source: §4.1.2 · p. 5

4.1.3 Measurements.

• Synthetic results are evaluated against ground truth using PSNR and SSIM; real-world results are assessed visually and with NIQE because ground truth is unavailable.The paper states that higher NIQE corresponds to lower image quality.Source: §4.1.3 · pp. 5–6
• Figure 5 compares real-world restorations, highlighting texture preservation in the first two examples and residual-rain removal in the last two.Source: Figure 5 · p. 6

• Training uses randomly selected 64 × 64 image-pair patches, with 32 channels and LeakyReLU activation at α = 0.2.Source: §4.2 · p. 6
• The network is implemented in PyTorch and trained with Adam on two NVIDIA 1080ti GPUs.Source: §4.2 · p. 6
• Training starts at learning rate 0.0005, divides it by 10 at epochs 480 and 640, and stops at epoch 800.Source: §4.2 · p. 6

• The method achieves the newest state-of-the-art results on Rain100H, Rain100L, and Rain1200.Source: §4.3 · p. 6
• Synthetic examples show clearer texture and cleaner backgrounds, especially in cropped regions, than the compared methods.Source: §4.3 · p. 6
• SSIR fails to restore rain-free images on synthetic datasets, which the authors say illustrates that its semi-supervised method cannot work on them.Source: §4.3 · p. 7

• In the first two examples, the method preserves clearer texture while other methods blur it, particularly in cropped boxes.Source: §4.4 · p. 7
• In the last two examples, other methods leave rain while the proposed method restores cleaner images; SSIR is noted as retaining substantial rain.Source: §4.4 · p. 7
• The method has lower NIQE values than other methods on the collected real-world images, which the authors interpret as more natural results.Source: §4.4 · p. 7

4.5.1 Analysis on Network Designment.

• The cross-scale network significantly outperforms a single-scale model, while skip and dense connections also improve deraining results.The tested variants include removing skip connections or dense connections and using a single network.Source: §4.5.1 · p. 7

4.5.2 Discussion on Inner-scale Connection.

• The designed inner-scale connection improves deraining performance by strengthening correlation between scales.The authors describe this connection as meaningful for deraining.Source: §4.5.2 · p. 7

4.5.3 The analysis of the number of scales.

• K = 3 gives the best reported PSNR and SSIM, and the multiscale setting improves deraining over a single scale.The authors report a significant improvement at K = 3 and select it as the scale count.Source: Table 5 · p. 7, Figure 6 · p. 7, §4.5.3 · pp. 7–8

4.5.4 Analysis of recurrent unit.

• ConvGRU performs better than ConvLSTM and ConvRNN in the recurrent-unit comparison and is selected as the default.The comparison is reported in Table 6 and Figure 7.Source: Table 6 · p. 8, Figure 7 · p. 8, §4.5.4 · p. 8

4.5.5 Failure Case.

• The method fails on complex rainy images with heavy streaks; the authors suggest synthetic training samples may not represent real-world rain conditions.They identify training on real-world images as a possible direction for a more robust derainer.Source: Figure 8 · p. 8, §4.5.5 · p. 8

5 CONCLUSION

• DCSFN combines inner-scale connections with cross-scale fusion between subnetworks at different scales.The cross-scale manner is intended to use information across subnetworks, while inner-scale connections improve deraining performance.Source: §5 · p. 8
• Densely connected encoder and decoder blocks with skip connections are developed to maximize information flow and support long-range spatial dependencies.The design also uses feature activations from proceeding layers.Source: §5 · p. 8
• Quantitative and qualitative experiments report that DCSFN outperforms several state-of-the-art deraining methods on Rain100H, Rain100L, and Rain1200.Source: §5 · p. 8