Residual-Guide Feature Fusion Network for Single Image Deraining

1 INTRODUCTION

• SSIM rises from 0.927 at block1 to 0.943 at block5, illustrating progressive reconstruction improvement with depth.Block3 reaches SSIM 0.935.Source: Figure 1 · p. 1, §1 · p. 1
• Shallower-block residuals guide deeper predictions, yielding finer negative-rain-residual estimates as the cascade progresses.The authors attribute this to preserving intermediate reconstruction features that a simple cascade may lose.Source: §1 · p. 1
• ResGuideNet supervises intermediate and final reconstructions, then merges intermediate outputs into the final result as an ensemble.The basic block uses recursive computation and a shortcut strategy to reduce parameters while retaining deraining performance.Source: §1 · p. 2
• The network is separable: users can detach blocks to meet different rainy conditions and test-time computational requirements.The paper contrasts light-rain cases suited to a simple model with heavy-rain cases that need more intensive processing.Source: §1 · pp. 1–2
• The authors discuss applying ResGuideNet to denoising and suggest reconstructed images could benefit downstream applications such as object detection.Rainy-image degradation is described as adversely affecting outdoor computer vision applications.Source: §1 · pp. 1–2

2 RELATED WORKS

• Video deraining uses information between adjacent frames to identify rainy regions and remove streaks.Single-image deraining is described as more challenging because less information is available.Source: §2 · p. 2
• Single-image deraining includes sparse-coding methods and, while learns background and rain using two Gaussian mixture models; CNN approaches include DDN, conditional GAN-based removal [33], JORDER [30], and DID-MDN [32].The first CNN-based single-image deraining method was introduced by [7].Source: §2 · p. 2
• Image-restoration studies reuse convolutional filter weights to learn hierarchical feature representations [6].The section frames good performance with a moderate parameter count as an important design goal.Source: §2 · p. 2
• Recursive computation with supervision and shortcuts was used for single-image super-resolution to reduce parameters while achieving state-of-the-art performance [16].The cited work is presented as a parameter-reduction strategy for deep models.Source: §2 · p. 2

3.1 Motivation

• The detachable design targets variable compute budgets: light-rain reconstruction changes little from blocks 2–5, while users can reserve computation for heavier cases.The paper motivates slicing the network to meet device constraints and different rainy conditions.Source: §3.1 · p. 2

3.2 Residual Feature Reuse

• ResGuideNet reuses negative residuals from shallow blocks to guide deeper blocks, addressing the loss of useful intermediate reconstruction information in a simple cascade.The blocks estimate negative residuals progressively, providing more guidance to deeper blocks.Source: §3.2 · p. 3, §3.6 · p. 5
• Baseline-RR gains progressively in SSIM as blocks deepen, unlike the native cascade, and yields more eye-pleasing reconstructions in the five-block comparison.The comparison matches parameter counts between Baseline and Baseline-RR.Source: §3.2 · p. 3

3.3 Loss Function

• ℓ2+SSIM loss supervises each block, combining per-pixel similarity with global structure; experiments report benefits from SSIM loss and intermediate-output supervision.The authors note that ℓ2 loss alone tends toward blurred reconstructions because per-pixel loss does not capture perceptual differences.Source: §3.3 · pp. 3–4

3.4 Recursive Computation

• Recursive computation shares a nonlinear mapping within each block to reduce parameters, while forwarding the first feature map through later recursive outputs to ease gradient flow.The five-recursion comparison reports that kernel reuse and forward propagation of that feature map benefit image-content restoration.Source: §3.4 · p. 4

3.5 Inter-block Ensemble

• The final reconstruction concatenates intermediate block outputs and compresses them with a 1×1 convolution; the paper reports an improved result for the merged output.Table 1 compares ResGuideNet5 and ResGuideNet.Source: §3.5 · p. 4, Table 1 · p. 4

3.6 Te Proposed Architecture

• Each block predicts a negative residual, with deeper blocks receiving the input reconstruction and residual features from earlier blocks; all block outputs are fused and supervised.The residual estimates are trained to approximate Y − X, making earlier residual guidance available to deeper blocks.Source: §3.6 · pp. 4–5

4.1 Implementation details

• Across experiments, filters were 3 × 3 except for the merge convolution, convolution layers had 16 feature maps, and ResGuideNet training required 50,000 iterations using RMSProp.Training used batch size 16 and an initial learning rate of 0.001 on a NVIDIA GeForce GTX 1080 with 8GB GPU memory.Source: §4.1 · p. 6

4.2 Dataset

• For fair comparison, models are trained and tested on synthetic Rain100H and Rain100L datasets, the Rain100L-trained model tests on Rain12, and training randomly generates 0.8 million 128×128 rainy/clean patch pairs.The four comparison datasets are Rain100H, Rain100L, Rain12, and a dataset of 10K rainy/clean pairs.Source: §4.2 · p. 6

4.3 Evaluation on Synthetic dataset

• ResGuideNet has SSIM values comparable to JORDER while outperforming other methods on the synthetic evaluation.The third-block output also has a decent result compared with other methods, despite the network using far fewer parameters and being sliceable for light-rain conditions.Source: §4.3 · p. 6

4.4 Evaluation on Real-world dataset

• ResGuideNet trained on synthetic data produces less-blurred qualitative results on real-world rainy images with varied streaks.No ground truth is available for this real-world evaluation.Source: §4.4 · p. 6

4.5 Running Time

• ResGuideNet runs quickly on GPU relative to other methods, and its third block runs faster still for light rain.The comparison averages runtime across 100 test images at 500×500 resolution; GMM is the slowest method.Source: §4.5 · p. 6

5.1 Generalization to other image processing tasks

• 0.927 SSIM is reported on BSD500 test images with Gaussian noise of standard deviation 0.1, supporting generalization to similar image-restoration problems.Training used 300 BSD500 images and testing used 100.Source: §5.1 · p. 6

5.2 Pre-processing for high-level vision tasks

• Figure 9 presents three synthetic rainy-image results, while Figure 10 shows five real-world examples spanning rain magnitudes and shapes.Source: Figure 9 · p. 7, Figure 10 · p. 7
• Using ResGuideNet before Faster R-CNN improves detection performance over feeding the degraded rainy image directly.The example concerns a pretrained detector that misses some objects under heavy rain and produces low recognition confidence.Source: §5.2 · pp. 6–7

• Shallow-block negative residuals guide deeper blocks toward coarse-to-fine rain-streak residual estimates, and the final reconstruction combines intermediate outputs.The network is presented as an ensemble across blocks.Source: §6 · pp. 7–8
• Figure 11 gives a denoising example with SSIM = 0.964, and Figure 12 shows an example of detection.Source: Figure 11 · p. 8, Figure 12 · p. 8
• ResGuideNet has 37K parameters, while ResGuideNet3 has fewer than 20K, and the model can be detached into smaller versions for different rain conditions and computational resources.Source: §6 · p. 8