Progressive Feature Fusion Network for Realistic Image Dehazing

• Single-image dehazing is an ill-posed but practical task, and severe haze can degrade image quality enough to jeopardize applications such as surveillance and autonomous driving.Source: §1 · pp. 1–2
• In the traditional separate pipeline, errors in transmission and atmospheric-light estimation can amplify one another and hinder overall dehazing performance.Source: §1 · p. 2
• Rather than estimate an intermediate transmission function under a simplified physical model, the method learns a direct mapping from hazy input to clear output.Source: §1 · p. 3
• The proposed network improves final restoration performance over several state-of-the-art methods on two public dehazing benchmarks.Source: §1 · p. 3
• A U-Net-like encoder-decoder uses progressive feature fusions between encoder and decoder mappings to learn an input-adaptive restoration model.Source: §1 · p. 3
• It directly processes realistic hazy images up to 4K with reasonable speed and memory, while many deep-learning methods cannot afford that resolution on one TITAN X GPU.Source: §1 · pp. 3–4

2 Related work

• Most dehazing methods follow a three-step scattering-model pipeline: estimate transmission, estimate atmospheric light, then recover the clear image.The section identifies the model as Cantor's [9].Source: §2 · p. 4
• DCP, color attenuation [3], and non-local prior [6] methods use different image or color assumptions to estimate scene information for haze removal.DCP uses a near-zero local dark-channel minimum; color attenuation links haze concentration to brightness-saturation difference; the non-local prior models haze-free colors as clusters.Source: §2 · p. 4
• Learning-based methods include transmission-estimating CNNs, AOD-Net's intermediate transmission variable, and GFN's confidence-map fusion of derived inputs.PFFNet is distinguished as directly learning the input-to-output transformation without estimating intermediate maps.Source: §2 · p. 4

3 Progressive Feature Fusion Network (PFFNet) for Image Dehazing

• The encoder uses five convolution-plus-ReLU layers, beginning with a large-receptive-field layer and then downsampling to encode image information at pyramid scales.The first convolution uses an 11 × 11 receptive field and preserves input spatial size; the following four use stride 2 and double channels at each layer.Source: §3 · p. 5
• For a 4K input, the encoder's D4 feature map has 256 spatial resolution, reducing the memory burden of subsequent processing.The stated D4 dimensions are 1/16 of input width and height, with 256 channels.Source: §3 · p. 6
• The feature-transformation module uses eighteen wide residual blocks, with B(3, 3) as its basic block.The paper presents this choice as balancing computational efficiency and GPU memory usage.Source: §3 · p. 6
• Four decoder deconvolution layers successively recover structural details, while corresponding encoder-decoder skips add features and a global shortcut links the transformation module's input and output.The deconvolution stages upsample to double spatial size and half the channels of the preceding feature map; the additions are channel-wise.Source: §3 · p. 6
• Progressive fusion between spatial-pyramid encoder and decoder mappings preserves input structural details and makes dehazing more input-adaptive.The architecture comprises an encoder, feature transformation, and decoder.Source: §3 · pp. 5, 7, Figure 2 · p. 5

4.1 Dataset

• The NTIRE2018 dataset includes 35 indoor I-HAZE scenes and 45 outdoor O-HAZE scenes, captured with real haze, with all images at 4K resolution.The paired haze-free images were captured under the same illumination parameters.Source: NTIRE2018 Image Dehazing Dataset · p. 7
• RESIDE training uses 13,990 synthetic hazy images, while its SOTS test set contains 500 non-overlapping indoor images.The images are synthesized using depth maps and sampled atmospheric parameters.Source: NTIRE2018 Image Dehazing Dataset · p. 7

4.2 Comparisons and Analysis

• Increasing the feature-transformation module from 6 to 24 residual blocks improved testing performance, and the authors chose 18 blocks to balance performance and computing resources.Figure 4-Left shows the block-size comparisons.Source: Ablation parameter comparisons on… · pp. 8–9, Figure 4 · p. 9
• With 12 residual blocks, the network without encoder-decoder skip connections performed and converged much worse than the network with them.Figure 4-Right compares the two cases on NTIRE2018 outdoor scenes after about 40 epochs.Source: Ablation parameter comparisons on… · p. 9, Figure 4 · p. 9
• PFF-Net ranked in the top 6 of 21 teams on the NTIRE2018 I-HAZE track and received an honorable mention without using O-HAZE data.The authors also report no specific training trick and no O-HAZE retraining at submission time.Source: Experiment results · p. 9
• On challenging 4K I-HAZE and O-HAZE examples, PFF-Net produced relatively satisfactory dehazing with natural color saturation and acceptable perceptual quality.Figures 5 and 6 show the examples.Source: Experiment results · p. 10, Figure 5 · p. 10, Figure 6 · p. 11
• PFF-Net outperformed current state-of-the-art methods on RESIDE's SOTS test set, with Table 1 presenting the comparison.The passage reports superior performance with great improvements but gives no metric values.Source: Experiment results · p. 10, Table 1 · p. 11
• On real-world hazy images around 500 × 600 pixels, the authors describe PFF-Net's results as clear, with moderately enhanced details and natural perceptual qualities.Figure 7 presents these qualitative comparisons.Source: Experiment results · p. 11

5 Conclusion

• PFF-Net uses progressive feature fusion in a U-Net-like encoder-decoder to learn an input-adaptive restoration model.The network treats haze formation as a black box and targets the final clear output.Source: §5 · p. 11
• The authors report superior performance over several popular state-of-the-art methods on two public dehazing benchmarks.The conclusion gives no specific metric values.Source: §5 · p. 11
• The network's efficient memory usage allows it to directly recover hazy images up to 4K resolution.Figure 7 is described as showing comparisons on real-world hazy images.Source: §5 · p. 11, Figure 7 · p. 12