Enhancing Hazy Wildlife Imagery: AnimalHaze3k and IncepDehazeGan
| Authors | Shivarth Rai & Tejeswar Pokuri |
| Year | 2026 |
| Field | Computer Vision |
| arXiv | 2604.16284 |
| Download | |
| Categories | cs.CV |
Abstract
Atmospheric haze significantly degrades wildlife imagery, impeding computer vision applications critical for conservation, such as animal detection, tracking, and behavior analysis. To address this challenge, we introduce AnimalHaze3k a synthetic dataset comprising of 3,477 hazy images generated from 1,159 clear wildlife photographs through a physics-based pipeline. Our novel IncepDehazeGan architecture combines inception blocks with residual skip connections in a GAN framework, achieving state-of-the-art performance (SSIM: 0.8914, PSNR: 20.54, and LPIPS: 0.1104), delivering 6.27% higher SSIM and 10.2% better PSNR than competing approaches. When applied to downstream detection tasks, dehazed images improved YOLOv11 detection mAP by 112% and IoU by 67%. These advances can provide ecologists with reliable tools for population monitoring and surveillance in challenging environmental conditions, demonstrating significant potential for enhancing wildlife conservation efforts through robust visual analytics.
Engineering Breakdown
Plain English
This paper tackles a real problem in wildlife conservation: hazy or foggy images break computer vision systems used for animal detection and tracking. The authors built a synthetic dataset called AnimalHaze3k with 3,477 hazy images generated from 1,159 clear wildlife photos using physics-based rendering, then trained a new GAN architecture called IncepDehazeGan to remove that haze. Their approach outperforms existing methods significantly—6.27% better SSIM and 10.2% better PSNR—and when you feed dehazed images into YOLOv11 detection, you get a 112% improvement in detection mAP and 67% better IoU, which is a massive practical gain for real-world conservation applications.
Core Technical Contribution
The core novelty is the IncepDehazeGan architecture, which combines inception blocks (multi-scale convolutional filters that capture features at different resolutions) with residual skip connections inside a generative adversarial network. This design lets the model learn both local detail recovery and global structure preservation simultaneously, addressing a weakness in prior dehazing GANs that often lost fine details. The authors also contributed AnimalHaze3k, a carefully curated synthetic dataset of haze on wildlife specifically, rather than generic scenes—this domain specificity matters because haze behaves differently depending on the subject matter and you need training data that matches your application. Together, these contributions—the architecture plus the domain-specific dataset—represent the first end-to-end solution optimized for wildlife imagery dehazing.
How It Works
The pipeline starts with AnimalHaze3k: the authors take 1,159 clear wildlife photos and apply a physics-based haze simulation (likely based on the atmospheric scattering model) to create 3,477 hazy variants with different haze densities. These pairs (clear, hazy) become training data for the IncepDehazeGan. The generator takes a hazy image as input and passes it through inception blocks—these blocks run parallel convolutions at different kernel sizes (1x1, 3x3, 5x5, etc.) to capture multi-scale features. Residual skip connections allow gradients to flow directly from input to output layers, preventing degradation during training and helping preserve details. The discriminator (standard GAN setup) learns to distinguish real dehazed images from generator outputs. The loss function combines perceptual loss (via LPIPS), pixel-level loss (PSNR/SSIM), and adversarial loss, forcing the generator to produce output that is both visually sharp and realistic. At inference, you pass a hazy wildlife image and get back a dehazed version ready for downstream detection models.
Production Impact
In a real conservation system using camera traps or drone footage, haze causes detection failures—animals disappear from the model's perspective. Plugging IncepDehazeGan into your preprocessing pipeline before object detection would be transformative: a 112% mAP improvement means you're catching animals you previously missed entirely, and 67% better IoU means bounding boxes are tighter and more usable for downstream analysis (behavior classification, individual identification, etc.). The trade-offs are meaningful: you need GPU memory for the GAN inference (likely 1-4 GB depending on image resolution), which adds latency of 100-500ms per image in a real-time system, and you need the AnimalHaze3k training data or you need to collect and annotate your own domain-specific hazy wildlife pairs. The architecture is relatively lightweight compared to diffusion models, so deployment on edge devices (like drone onboard computers) is feasible but requires optimization. Integration is straightforward—it's a standard PyTorch/TensorFlow model that takes an image tensor and outputs a dehazed tensor—but you must validate performance on your specific animal species and haze conditions since the dataset may not cover your exact use case.
Limitations and When Not to Use This
The paper's main limitation is that AnimalHaze3k uses synthetic haze generated from a physics-based model, which may not perfectly match real-world atmospheric conditions (different haze types—fog, dust, wildfire smoke—scatter light differently, and the model may oversimplify these). The 1,159 source clear images is a modest dataset, meaning the model may struggle with rare animal species or unusual poses not well represented in training. The evaluation metrics (SSIM, PSNR, LPIPS) measure image quality but don't directly measure whether the dehazed output helps downstream tasks—the paper does show YOLOv11 gains, but this doesn't generalize to all detectors or to other applications like animal segmentation or pose estimation. The paper doesn't discuss failure modes: what happens with extreme haze, or with other degradations (rain, night vision noise) mixed in? There's no analysis of domain shift—if you train on images dehazed toward a specific color temperature or contrast, will it work on wildly different camera sensors or lighting conditions?
Research Context
This work builds on a decade of single-image dehazing research (prior GANs like CycleGAN and pix2pix showed GANs could do image-to-image translation) and combines that with recent advances in inception modules for multi-scale processing. The benchmark metrics (SSIM, PSNR, LPIPS) are standard in image restoration, so the 6.27% SSIM improvement is positioned against published results from methods like AODNet and DehazeNet. The novelty is less in the individual components and more in the application to wildlife and the specific architectural choices (inception + residuals in a GAN, not done before for dehazing). This opens a research direction around domain-specific image restoration for conservation: the same pattern (synthetic dataset + custom architecture) could be applied to remove snow, sand, or rain from wildlife images, which is an understudied area compared to general-purpose dehazing.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::