PLAMD: Power Line Aerial Multi-Degradation Dataset
收藏资源简介：
## 📖 Overview**PLAMD (Power Line Aerial Multi-Degradation Dataset)** is a large-scale synthetic dataset designed for benchmarking image restoration and enhancement algorithms in the context of **unmanned aerial vehicle (UAV) power line inspection**. It covers **seven realistic degradation scenarios** with paired clean references, supporting full-reference evaluation paradigms.Source images are drawn from two public UAV power line inspection datasets — **InsPLAD** [<sup data-citation='{"id":1,"url":"https://doi.org/10.1080/01431161.2023.2283900","title":"InsPLAD: A Dataset and Benchmark for Power Line Asset Inspection in UAV Images","content":"InsPLAD: A Dataset and Benchmark for Power Line Asset Inspection in UAV Images International Journal of Remote Sensing, 2023 57 citations Power line maintenance and inspection are essential to avoid p"}'>1</sup>](https://doi.org/10.1080/01431161.2023.2283900) and **TTPLA** [<sup data-citation='{"id":2,"url":"https://arxiv.org/abs/2311.01619","title":"InsPLAD: A Dataset and Benchmark for Power Line Asset ...","content":"InsPLAD: A Dataset and Benchmark for Power Line Asset Inspection in UAV Images arXiv Cornell University, 2023 Preprint 1 citation Power line maintenance and inspection are essential to avoid power sup"}'>2</sup>](https://arxiv.org/abs/2311.01619) — ensuring strong domain relevance for the power industry. Each image is then synthetically degraded using physically motivated or learning-based generation pipelines. ## Degradation Types | Type | Synthesis Method | Description || Haze | Atmospheric Scattering Model + Monodepth2 depth estimation | Fog/haze at various concentrations || Rain | Streak-based rain generation | Rainfall with tunable density, length, angle, and opacity || Snow | AutoSnow framework (5 snow conditions) | Streaking snow, fine snow, fluffy snow, and light ice accumulation || Motion Blur | Camera Shake Blur (CSB) generator | Physics-based random camera shake trajectory synthesis || Low-Light | Gamma correction + Poisson-Gaussian noise | Low-illumination evening/dusk scene simulation || Noise | sRGB real noise synthesizing (signal-dependent + signal-independent) | Realistic camera sensor noise || Inpainting | OpenCV-based random mask generation | Partial occlusion simulation (irregular shapes, stripes, mixed) | ## Key Features | Feature | Description ||---------|-------------|| **7 Degradation Types** | Blur (motion), Haze, Inpainting, Low-light, Noise, Rain, Snow || **Realistic UAV Perspective** | All source images from real UAV power line inspection footage (InsPLAD + TTPLA) || **Full-Reference Evaluation** | Each degraded image has a paired clean reference || **3 Severity Levels** | Random parameter ranges during generation, yielding Light / Moderate / Severe degradation per image || **Fine-Grained Component Subsets** | 20 categories organized by power line component type for targeted research || **Intact & Defect Coverage** | Both normal-condition and defect-present component images included | 📂 Dataset StructurePLAMD is distributed as **7 compressed archives**, one per degradation type. Each archive shares an identical internal hierarchy organized by **power line component category**. Images are provided as **degraded–clean pairs** for full-reference training and evaluation.{degradation_type}.zip # e.g., haze.zip, rain.zip, snow.zip ...│├── {degradation_type} # All degraded images (e.g., haze/, rain/, snow/ ...) —— component_mix_1K/damper-preformed/damper-stockbridge/glass-insulator/glass-insulator-big-shackle/glass-insulator-small-shackle/glass-insulator-tower-shackle/lightning-rod-shackle/lightning-rod-suspension/overallL_2K/plate/polymer-insulator/polymer-insulator-lower-shackle/polymer-insulator-tower-shackle/polymer-insulator-upper-shackle/vari-grip/yoke/yoke-suspension└── original # Original clean reference images (identical structure) —— component_mix_1K/damper-preformed/damper-stockbridge/glass-insulator/glass-insulator-big-shackle/glass-insulator-small-shackle/glass-insulator-tower-shackle/lightning-rod-shackle/lightning-rod-suspension/overallL_2K/plate/polymer-insulator/polymer-insulator-lower-shackle/polymer-insulator-tower-shackle/polymer-insulator-upper-shackle/vari-grip/yoke/yoke-suspension Design Notes- **`component_mix_1K`**: a curated subset containing mixed component images at approximately **1K resolution**, useful for quick prototyping and lightweight experiments.- **`overallL_2K`**: wide-view overall-scene images at approximately **2K resolution** (e.g., full transmission towers, multi-component line spans). Ideal for context-aware or scene-level restoration models.- All other categories contain images at **varying resolutions** depending on the original source crops.- Several category includes both **good** and **rust/missing-cap/nest etc.** component images sourced from InsPLAD and TTPLA, enabling research on restoration quality across varying asset conditions. ## 🛠️ Degradation Synthesis MethodsAll degradations are applied synthetically to clean source images, ensuring **pixel-aligned degraded–clean pairs**.| Degradation | Method | Description ||-------------|--------|-------------|| **Haze** | Atmospheric Scattering Model [3] + Monodepth2 [4] | Depth-aware fog/haze at various concentrations; Monodepth2 estimates depth maps from single images to modulate transmission maps || **Rain** | Rain-Generation-Python [5] | Streak-based rain rendering with tunable density, length, angle, and opacity || **Snow** | AutoSnow [6] | Synthetic winter image generation framework supporting 5 snow conditions: streaking snow, fine snow, fluffy snow, and light ice accumulation || **Motion Blur** | Camera Shake Blur (CSB) generator | Physics-based random camera shake trajectory synthesis producing realistic motion blur kernels || **Low-Light** | Gamma correction + Poisson-Gaussian noise | Low-illumination evening/dusk scene simulation via non-linear intensity attenuation and realistic noise injection || **Noise** | NoiseDiff (TPAMI 2025) [7] | Dark noise diffusion model for realistic sRGB noise synthesis (signal-dependent + signal-independent) || **Inpainting** | OpenCV-based random mask generation | Irregular shapes, stripes, and mixed mask patterns simulating partial occlusion | ## How to Build an All-in-One Training Set Since PLAMD is organized by degradation type, researchers canconstruct an All-in-One training set by sampling from each category.A recommended mixing ratio is provided below as a starting point: | Degradation | Suggested Ratio ||-------------|:---:|| Haze | 25% || Rain | 20% || Snow | 15% || Blur | 15% || Low-Light | 15% || Noise | 5% || Inpainting | 5% | The mixing ratio can be adjusted based on specific deploymentrequirements (e.g., higher snow ratio for winter operations). ## Dataset Statistics | Property | Value ||----------|-------|| Total degradation types | 7 || blur | 38,865 || haze | 38,865 || inpainting | 39,699 || lowlight | 38,865 || noise | 27,802 || rain | 39,699 || snow | 39,699 || File format | JPG || Compression | ZIP | ## Intended Use 1. **Single-task restoration**: training and evaluating models on individual degradation types (dehazing, deraining, etc.)2. **All-in-One restoration**: mixing multiple categories to train a unified model handling diverse degradations3. **Domain adaptation validation**: comparing models trained on PLAMD vs. generic datasets to demonstrate the necessity of power-domain-specific data4. **Downstream task evaluation**: assessing whether restored images improve power asset detection/segmentation performance Note on Class Balance: Due to the stochastic nature of the data augmentation pipeline, a small subset of generated images exhibited suboptimal quality and were manually filtered out during post-processing. As a result, the seven degradation tasks exhibit minor sample count discrepancies. Specifically, the noisy subset contains the fewest images due to file-size-limitaion, while the other six tasks differ by no more than 5% from one another. These variations are negligible for standard training workflows but should be noted in fairness-aware or class-balanced evaluation settings. ## Citation If you use PLAMD in your research, please cite: @dataset{plamd2025, title = {PLAMD: A Comprehensive Multi-Degradation Dataset for UAV Power Line Image Restoration}, author = {[Jingke Yang]}, year = {2025}, publisher = {Zenodo}, doi = {[10.5281/zenodo.21321669]}, } ## 🙏 AcknowledgementPLAMD builds upon two publicly available UAV power line inspection datasets. Please also cite:InsPLAD [1]:@article{InsPLAD2023, author = {André Luiz Buarque Vieira e Silva and Heitor de Castro Felix and Franscisco Paulo Magalhães Simões and Veronica Teichrieb and Michel dos Santos and Hemir Santiago and Virginia Sgotti and Henrique Lott Neto}, title = {{InsPLAD}: A Dataset and Benchmark for Power Line Asset Inspection in {UAV} Images}, journal = {International Journal of Remote Sensing}, volume = {44}, number = {23}, pages = {1--27}, year = {2023}, publisher = {Taylor \& Francis}, doi = {10.1080/01431161.2023.2283900}} TTPLA [2]:@inproceedings{TTPLA2020, author = {Abdelfattah, Rabab and Wang, Xiaofeng and Wang, Song}, title = {{TTPLA}: An Aerial-Image Dataset for Detection and Segmentation of Transmission Towers and Power Lines}, booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)}, year = {2020}}[3] Narasimhan, S. G., & Nayar, S. K. (2002). Vision and the atmosphere. IJCV, 48(3), 233–254. [4] Godard, C., Mac Aodha, O., Firman, M., & Brostow, G. J. (2019). Digging into self-supervised monocular depth estimation. ICCV 2019. [5] Shen Zheng. Rain-Generation-Python. https://github.com/ShenZheng2000/Rain-Generation-Python [6] Amardeep Sarang. AutoSnow: A Synthetic Winter Image Generator Framework. https://github.com/AmardeepSarang/AutoSnow-A-synthetic-winter-image-generator-framework [7] NoiseDiff: Dark Noise Diffusion — Noise Synthesis for Low-Light Image Denoising (TPAMI 2025). IVRL. https://github.com/IVRL/NoiseDiff
