Exploring Image Enhancement for Salient Object Detection in Low Light Images

1 INTRODUCTION

• Figure 1 contrasts bright-image R3Net examples with low-light results, where the authors report considerable results from their approach.The figure shows inputs, R3Net outputs, the approach's outputs, ground truths, and enhanced outputs for the low-light examples.Source: Figure 1 · p. 2
• Low contrast and changing artificial light in low-light images degrade salient-object information, making objects difficult to detect with methods developed for sufficiently bright images.The introduction notes that existing methods can lose object details and include non-salient backgrounds on degraded low-light images.Source: §1 · p. 2
• The approach enhances low-light images by modeling environment lighting as a point-wise, locally changing variable, then uses a Non-Local-Block Layer to extract salient-object features.The stated aim is to retain salient-object detail and improve SOD performance.Source: §1 · pp. 2–3
• The paper builds a low-light image dataset for SOD and uses it to verify that low illumination can reduce SOD performance.The dataset is presented as a contribution for the SOD community.Source: §1 · p. 3

2 RELATED WORKS

• The related-work section frames low illumination as one of several sources of interference in separating salient objects from their surroundings.It situates the paper's focus on low-light SOD among broader challenges such as viewpoint, background, and resolution variation.Source: §2 · p. 3

2.1 Salient Object Detection

• Deep SOD methods use global, local, or combined features; the paper notes that low-light SOD work is limited and direct detection on degraded images may be non-optimal.Earlier approaches include global-feature models, local-feature models, and methods combining global and local information.Source: §2.1 · pp. 3–4

2.2 Low Light Image Enhancement

• Global histogram equalization can reduce boundary sharpness, while local histogram equalization may over-enhance bright regions.These trade-offs motivate the review of alternatives for low-light image enhancement.Source: §2.2 · p. 4
• Retinex-based approaches rely on a Lambertian scene assumption and piece-wise smooth illumination, whereas artificial-light interference can produce rapidly changing illumination and halo effects.The review also notes that some physical-lighting approaches may lose information in dark regions.Source: §2.2 · p. 4
• Deep enhancement methods learn image mappings or combine illumination estimation with detail reconstruction, but the reviewed methods mainly target visual quality rather than subsequent SOD.The cited examples are RetinexNet and GLADNet.Source: §2.2 · p. 4

3 PROPOSED METHOD

• PIE enhances the low-light input using atmospheric light A(z) and transmission t(z), then NLD predicts saliency from the enhanced image J(z).The two subnetworks form the method’s overall pipeline.Source: §3 · p. 5, §3.1 · p. 5

3.1 PIE

• PIE adapts the atmospheric scattering model to non-uniform nighttime illumination by treating atmospheric light as a point-wise random variate, rather than a constant.The model includes a U-Net, an encoder-decoder network, atmospheric scattering, and a joint discriminator.Source: §3.1 · p. 5, §3.1.1 · p. 5

3.1.1 U-Net.

• An 8-block U-Net predicts atmospheric light A(z), with four Conv-BN-Relu encoder blocks and symmetric Dconv-BN-Relu decoder blocks.For training, A(z) is randomly valued to create corresponding atmospheric-light maps; α is set to 0.5.Source: §3.1.1 · p. 5

3.1.2 Encoder-decoder Network.

• A DenseNet-121-based encoder-decoder estimates transmission t(z), using pretrained encoder weights and a decoder to reconstruct the map at the original resolution.The architecture is intended to preserve object structure and produce high-resolution feature maps.Source: §3.1.2 · p. 6

3.1.3 Atmospheric Scattering Model.

• Atmospheric scattering models the observed image as direct attenuation plus scattered light, with A(z) varying by pixel according to local scene light.The model estimates enhanced image J(z) from observed low-light image I(z), atmospheric light A(z), and transmission t(z).Source: §3.1.3 · p. 6

3.1.4 Joint Discriminator Learning.

• Joint discriminator learning uses the structural correlation between transmission t(z) and enhanced image J(z) to distinguish estimated pairs from their ground-truth counterparts.The estimated transmission map and enhanced image are concatenated for discrimination.Source: §3.1.4 · p. 6

3.2 NLD

• NLD applies a Non-Local-Block Layer to features from enhanced images, capturing differences from neighboring regions to retain object detail and combine local and global context.Its architecture derives multi-scale local feature maps, applies non-local processing, then fuses local and global features to predict saliency.Source: §3.2 · pp. 6–8

3.3 Overall Loss Function

• The overall loss combines cross-entropy with NLDF’s IoU boundary loss, which is used to make the boundary robust to background noise.PIE also jointly minimizes atmospheric-light estimation error and transmission-map estimation error.Source: §3.3 · p. 9

4 NIGHTTIME IMAGE DATASET FOR SOD

• NTI-V1 contains 577 nighttime images, each with a pixel-level human-labeled ground-truth annotation.Source: §4 · pp. 9–10
• Collection comprised 224 surveillance-camera images captured from 7–9 PM and 353 smartphone images captured from 9–11 PM.Source: §4 · p. 9
• The dataset includes single people, multiple people, and vehicles, with varied viewpoints, illumination, and scenes.Source: §4 · pp. 9–10
• The evaluation protocol assigns 457 images to training and 120 to testing.Source: §4 · p. 10
• Table 1 compares NTI-V1 with low-light enhancement and SOD datasets; the paper identifies NTI-V1 as the first available low-light SOD benchmark.Source: §4 · p. 10, Table 1 · p. 10

5.1 Datasets and Experimental Settings

• Experiments use four public daytime datasets and the nighttime NTI-V1 dataset; NTI-V1 has 577 images, with 457 used for training and 120 for testing.The evaluation reports PR curves, max Fβ, and MAE; the NTI-V1 set includes hand-labeled person and vehicle categories.Source: §5.1 · pp. 11–12

5.2 Comparison with State-of-the-art Methods

• Results on NTI-V1 are better for images captured from 7–9 PM than from 9–11 PM, while object type shows no obvious effect.Table 2 reports the proposed method’s results across three object types and two capture stages.Source: §5.2 · p. 12, Table 2 · p. 12
• The comparison evaluates SOD methods on original low-light images, SOD models on PIE-enhanced images, and NLD on images enhanced by alternative methods.The tested SOD methods include MR, BSCA, DSSC, LPS, R3Net, and BASNet; enhancement methods include Gamma, LIME, LECARM, Dong, Ying, RetinexNet, and GLADNet.Source: §5.2 · p. 12

5.2.1 Comparison with state-of-the-art SOD methods on Low light images.

• The proposed SOD method beats the compared methods on NTI-V1 but not on the four public daytime datasets, and its PR curve is best on NTI-V1.Qualitative results describe sharper, more uniformly highlighted objects and saliency maps closer to ground truth.Source: §5.2.1 · pp. 12–13, Figure 7 · p. 13

5.2.2 Comparisons with several SOD methods on low light images enhanced via PIE.

• PIE-enhanced NTI-V1 improves SOD performance over the unenhanced setting, and PIE benefits both NLD and other tested SOD models.The comparison reports a better PR curve for the proposed method on enhanced NTI-V1; qualitative results also show PIE+LPS outperforming LPS on the original images.Source: §5.2.1 · p. 13, Figure 7 · p. 13, §5.2.2 · p. 13

5.2.3 Comparisons among PIE and several image enhancement methods.

• On NTI-V1, NLD with PIE achieves the best results among the tested image-enhancement methods, though PIE’s improvement differs between LPS and NLD.PIE improves LPS less than LECARM does, but improves NLD more; the qualitative comparisons also show the proposed method closer to ground truth than PIE+LPS and LECARM+NLD.Source: Table 5 · p. 14, §5.2.3 · pp. 14–15

5.3 Ablation Study

• Using point-wise random A(z) instead of constant A(z) improves MAE by 1.4% and max Fβ by 0.4% on NTI-V1.The authors attribute this design to treating atmospheric light as varying rather than constant.Source: §5.3 · p. 15

5.3.2 The Ablation Study of Non-Local-Block Layer.

• On NTI-V1, adding the Non-Local-Block (NLB) Layer improves MAE by 2.1% and max Fβ by 4.1% compared with the method without the NLB Layer.Source: §5.3.2 · p. 15

• The proposed low-light SOD method embeds a physical lighting model in a deep network and uses a Non-Local-Block Layer to extract non-local salient-object features.Source: §6 · p. 15
• NTI-V1 contains 577 low-light images with pixel-wise object-level annotations for the SOD community.Source: §6 · p. 15
• The experiments verify that low illumination can reduce SOD performance, while the proposed method enhances local content to facilitate SOD.Source: §6 · pp. 15–16