IncepDehazeGan: Wildlife Image Dehazing GAN

• The paper presents IncepDehazeGan, a GAN-based system that integrates inception blocks and residual skip connections to enhance dehazing quality in wildlife images.
• It leverages a physics-based synthetic dataset (AnimalHaze3k) to benchmark performance, achieving notable improvements in SSIM, PSNR, FSIM, and LPIPS over competing methods.
• Dehazing with IncepDehazeGan significantly boosts downstream YOLOv11 animal detection, with over 112% mAP and 67% mIoU improvement, emphasizing its conservation impact.

Searching arXiv for the cited papers to ground the article in current preprints. IncepDehazeGan is a GAN-based single-image dehazing architecture introduced for wildlife imagery in “Enhancing Hazy Wildlife Imagery: AnimalHaze3k and IncepDehazeGan” (Rai et al., 17 Apr 2026). It is presented as a dehazing model tailored specifically for conservation-oriented visual data, especially camera-trap and field imagery in which atmospheric haze reduces visibility, shifts colors unnaturally, lowers contrast between animal and background, and weakens depth cues. Within that framing, the model is coupled to a synthetic paired benchmark, AnimalHaze3k, and evaluated not only by image-restoration metrics but also by downstream animal detection performance. The work positions IncepDehazeGan as an end-to-end dehazing system that combines inception blocks with residual skip connections in a generator–discriminator framework, while its broader significance lies in connecting dehazing quality to animal detection, tracking, species identification, behavior analysis, and poaching surveillance (Rai et al., 17 Apr 2026).

1. Problem domain and conservation motivation

The paper frames haze in wildlife imagery as more than a cosmetic degradation. In the stated problem setting, atmospheric haze makes wildlife images less informative for both humans and automated systems, with direct consequences for animal detection, tracking, species identification, behavior analysis, and poaching surveillance (Rai et al., 17 Apr 2026). This domain emphasis distinguishes IncepDehazeGan from generic dehazing systems whose primary evaluation target is image fidelity alone.

The motivating application context is the Northeast Tiger and Leopard National Park (NTLNP) dataset, which contains 25,567 camera-trap images of 17 species captured between 2014 and 2020 under daylight and nighttime conditions. For the dehazing study, only daylight images were used. The resulting focus is therefore explicitly on wildlife scenes where haze obscures fine animal structure, background separation, and scene interpretability (Rai et al., 17 Apr 2026).

A plausible implication is that the model is designed under conditions where local texture recovery has operational value, not merely perceptual value. The paper’s discussion of fur texture, object boundaries, and background separation supports that interpretation, particularly because the restored imagery is later tested with YOLOv11 detection rather than treated as a purely aesthetic restoration task (Rai et al., 17 Apr 2026).

2. AnimalHaze3k and the physics-based synthesis pipeline

To address the lack of paired real-world hazy/clear wildlife data, the paper introduces AnimalHaze3k, described as a synthetic dataset comprising 3,477 hazy images generated from 1,159 clear wildlife photographs through a physics-based pipeline (Rai et al., 17 Apr 2026). The clear-image source is NTLNP. From the full dataset, 1,159 clear wildlife images representing 11 species were selected, standardized to $640 \times 480$ resolution, and stripped of date/time stamps. Each clear image was then converted into three hazy variants, yielding 3,477 hazy images total (Rai et al., 17 Apr 2026).

The dataset is split into training, validation, and test subsets. The training set contains 1,041 clear and 3,123 hazy images, while the validation and test sets each contain 59 clear and 177 hazy images (Rai et al., 17 Apr 2026). The pairing structure is therefore explicit, supervised, and suitable for direct image-to-image dehazing.

Haze synthesis follows the standard atmospheric scattering model:

$I(x) = J(x)\cdot t(x) + A\cdot(1- t(x))$

where $I(x)$ is the observed hazy image, $J(x)$ is the clean scene radiance, $A$ is global atmospheric light, and $t(x)$ is the transmission map. The transmission is computed as

$t(x) = e^{ - \beta \cdot d(x)}$

with $\beta$ the scattering coefficient and $d(x)$ the depth map (Rai et al., 17 Apr 2026).

The synthesis pipeline has three stages. First, a depth map is estimated using HybridDepth, which the authors state was chosen because it provides state-of-the-art metric depth accuracy, better generalization than ZoeDepth, DFV, and Depth Anything, and more consistent depth across zoom levels. Second, the transmission map is computed from the depth map, with $\beta$ sampled uniformly from $I(x) = J(x)\cdot t(x) + A\cdot(1- t(x))$0. Third, the hazy image is generated using atmospheric light $I(x) = J(x)\cdot t(x) + A\cdot(1- t(x))$1 randomly chosen from $I(x) = J(x)\cdot t(x) + A\cdot(1- t(x))$2, following prior benchmarking practice (Rai et al., 17 Apr 2026).

This synthesis strategy is physically grounded, but the paper also identifies its principal limitation: AnimalHaze3k is synthetic rather than captured from truly hazy wildlife scenes, so a domain gap to real atmospheric conditions remains (Rai et al., 17 Apr 2026). That limitation places the work in a broader discussion about synthetic-data bias in dehazing. A related line of work argues that conventional synthetic datasets can induce a depth prior that harms real-world generalization, and proposes depth-agnostic haze synthesis to decouple haze density from scene depth (Xu et al., 2024). This suggests that AnimalHaze3k’s physically grounded construction improves controllability and supervision, while still inheriting the general problem that synthetic haze distributions may not fully match real conditions.

3. Generator and discriminator design

IncepDehazeGan is described as a GAN-based single-image dehazing system built around a generator–discriminator pair (Rai et al., 17 Apr 2026). The generator is a dense encoder–decoder architecture that integrates Inception blocks with residual skip connections. Its encoder has two parallel processing streams.

One stream is a conventional convolutional encoder composed of repeated blocks containing two convolution layers, ReLU activation, and max pooling. The second stream uses Inception blocks, each running four parallel convolutions with receptive fields $I(x) = J(x)\cdot t(x) + A\cdot(1- t(x))$3, $I(x) = J(x)\cdot t(x) + A\cdot(1- t(x))$4, $I(x) = J(x)\cdot t(x) + A\cdot(1- t(x))$5, and $I(x) = J(x)\cdot t(x) + A\cdot(1- t(x))$6 (Rai et al., 17 Apr 2026). The stated purpose of this design is multi-scale capture of haze and structural cues, since wildlife scenes contain both small details such as fur edges and leaves and larger scene structures.

The decoder is a single series of blocks, each consisting of two standard convolution layers followed by a transpose convolution layer. Residual skip connections link encoder blocks to their corresponding decoder blocks. The paper explicitly states that these skip connections carry low-level feature maps forward and that this multilayer feature fusion reduces information loss caused by encoder downsampling (Rai et al., 17 Apr 2026). In architectural terms, the generator is therefore intended to combine global haze removal with preservation of fine-grained animal features.

The discriminator is a conventional convolutional classifier with six convolutional blocks followed by a final sigmoid activation. Each block contains a convolution layer, leaky ReLU, and batch normalization. Its output is a single-channel map, described as being in the spirit of patch-based GAN discrimination, encouraging realism at a local texture level and mitigating overfitting by focusing on small image regions (Rai et al., 17 Apr 2026).

This design places IncepDehazeGan within the broader family of paired conditional GAN dehazing systems that directly map hazy images to clear images without separately estimating the transmission map or atmospheric light. A closely related precedent is the input-size-flexible cGAN of “A GAN-Based Input-Size Flexibility Model for Single Image Dehazing” (Kan et al., 2021), which likewise formulates dehazing as direct image-to-image translation. The difference in emphasis is that IncepDehazeGan centers its generator on inception-based multiscale feature extraction and residual skip fusion for wildlife scenes, whereas the earlier cGAN emphasizes UR-Net, spatial pyramid pooling in the discriminator, consistency loss, and input-size flexibility (Kan et al., 2021).

4. Training objective and implementation regime

Training uses a hybrid objective combining adversarial loss and $I(x) = J(x)\cdot t(x) + A\cdot(1- t(x))$7 reconstruction loss (Rai et al., 17 Apr 2026). The adversarial term is defined against target labels of ones for the generator, so the generator is rewarded when the discriminator classifies its output as real. The reconstruction term is mean absolute error:

$I(x) = J(x)\cdot t(x) + A\cdot(1- t(x))$8

The final generator objective is stated as a weighted sum of adversarial loss and $I(x) = J(x)\cdot t(x) + A\cdot(1- t(x))$9 loss, with $I(x)$0 (Rai et al., 17 Apr 2026). The discriminator is trained with binary cross entropy to classify real clear images versus generator outputs.

The reliably stated implementation details are limited. Experiments were run on an NVIDIA Tesla P100 GPU with 16 GB memory and 3584 CUDA cores. All dehazing models, including baselines and IncepDehazeGan, were trained for 50 epochs with batch size 4 (Rai et al., 17 Apr 2026). The paper does not report optimizer hyperparameters such as learning rate or momentum settings. That omission is explicitly noted in the source material and should be treated as part of the method’s reporting constraints rather than filled by inference.

This restricted reporting contrasts with other cGAN-based dehazing studies that expose a more elaborate optimization stack. For example, the input-size-flexible cGAN paper specifies TensorFlow implementation, Adam with learning rate $I(x)$1, $I(x)$2, $I(x)$3, batch size 1, discriminator updates once every 4 iterations, and a multi-term loss including consistency, $I(x)$4, SSIM, PSNR, and $I(x)$5 regularization in multi-scale training (Kan et al., 2021). IncepDehazeGan does not provide comparable optimizer granularity, so the method’s most firmly supported technical description is architectural and evaluative rather than optimization-centric (Rai et al., 17 Apr 2026).

5. Quantitative performance on AnimalHaze3k

The paper reports IncepDehazeGan as the best-performing method on the AnimalHaze3k test set across SSIM, PSNR, FSIM, and LPIPS (Rai et al., 17 Apr 2026). The comparison is made against FFA-Net, FD-GAN, DEA-Net, and DehazeFormer.

Method	Metrics
FFA-Net	SSIM 0.5468, PSNR 12.1842, FSIM 0.6665, LPIPS 0.3610
FD-GAN	SSIM 0.5580, PSNR 17.5573, FSIM 0.8314, LPIPS 0.1637
DEA-Net	SSIM 0.8303, PSNR 18.6481, FSIM 0.8936, LPIPS 0.2102
DehazeFormer	SSIM 0.8388, PSNR 17.4550, FSIM 0.8917, LPIPS 0.2375
IncepDehazeGan	SSIM 0.8914, PSNR 20.5404, FSIM 0.9363, LPIPS 0.1104

The abstract summarizes this as a 6.27% improvement in SSIM and a 10.2% improvement in PSNR over competing approaches (Rai et al., 17 Apr 2026). The table values also show that the model attains the lowest LPIPS among the listed methods, which the paper interprets as being perceptually closer to the clean image.

The work therefore claims simultaneous gains in both reference-based fidelity and perceptual similarity on its synthetic wildlife benchmark. However, because all quantitative comparisons are reported on AnimalHaze3k, the strongest supported conclusion is benchmark-specific: IncepDehazeGan outperforms the listed baselines on that paired synthetic test set (Rai et al., 17 Apr 2026). A broader claim of cross-domain superiority would require evidence not provided in the source data.

The paper also reports qualitative examples in which IncepDehazeGan restores clearer fur texture, object boundaries, and background separation compared with the baselines (Rai et al., 17 Apr 2026). This suggests that the multi-branch encoder and skip-connected decoder are functioning in the way the architecture intends, namely by jointly modeling haze removal and detail preservation.

6. Downstream detection impact, interpretation, and limitations

A central empirical result is the practical effect of dehazing on downstream animal detection using YOLOv11 (Rai et al., 17 Apr 2026). In this experiment, YOLOv11 is trained on the NTLNP dataset and tested on hazy images from the AnimalHaze3k test set as well as on the corresponding dehazed outputs produced by IncepDehazeGan.

On hazy images, YOLOv11 achieves mAP 0.3216 and mIoU 0.4313. On dehazed images, these values rise to mAP 0.6842 and mIoU 0.7201 (Rai et al., 17 Apr 2026). The paper states this as more than a 112% increase in mAP and more than a 67% increase in IoU/mIoU. Within the paper’s framing, this is the key argument that dehazing meaningfully improves detection quality in wildlife scenes rather than merely producing visually cleaner outputs.

This downstream emphasis is important because dehazing literature has often been evaluated primarily through PSNR- and SSIM-oriented protocols. A related cGAN-based study also reports that dehazed outputs improve downstream detection performance (Kan et al., 2021), but IncepDehazeGan places that observation in a wildlife-conservation setting and ties it directly to operational tasks such as population monitoring, movement tracking, behavioral analysis, and surveillance (Rai et al., 17 Apr 2026).

The main strengths identified in the paper are the AnimalHaze3k dataset, the multiscale inception-based generator, and the evidence that dehazing improves downstream detection (Rai et al., 17 Apr 2026). The main limitation is that AnimalHaze3k is synthetic rather than derived from truly hazy wildlife scenes, leaving a domain gap to real atmospheric conditions. The paper also does not report some potentially useful implementation details such as optimizer hyperparameters, learning rates, or broader cross-dataset validation (Rai et al., 17 Apr 2026).

A common misconception in synthetic dehazing research is that strong benchmark performance automatically implies robust real-world dehazing. The evidence provided here does not support that generalization unconditionally. Instead, the defensible conclusion is narrower: the combination of a physically grounded synthetic dataset and a multiscale GAN architecture is effective on AnimalHaze3k and improves YOLOv11 detection on the paper’s hazy-to-dehazed wildlife evaluation protocol (Rai et al., 17 Apr 2026). A plausible implication is that future work could combine wildlife-specific data curation with synthetic regimes designed to weaken depth-dependent shortcuts, as advocated in depth-agnostic dehazing research (Xu et al., 2024), in order to reduce the remaining synthetic-to-real gap.