Powerline Image Dataset (Infrared-IR and Visible Light-VL)

Despite a relatively crowded literature about detection of power lines for aircraft safety, the works are mostly optimized with very few training images; some even use artificially generated images. The IR imaging case is even less utilized. The reason is clearly the tremendous workload to obtain real images. With this demand in mind, the authors cooperated with the Turkish Electricity Transmission Company (TEIAS) to obtain video captures from actual aircrafts. Later, the authors made a thorough inspection over the video frames to isolate, capture and clean thousands of valuable images.

In this content, 4000 IR and 4000 VL images are acquired and scaled to a size of 128x128. The IR folder contains 2000 IR images with power line and 2000 IR images with no power line. The VL folder contains 2000 VL images with power line and 2000 VL images with no power line. The IR and VL groups were deliberately constructed to contain both regular and especially confusing scenes. "TY" shows that there are no power lines in the images. "TV" shows that there are power lines in the images.

The videos were captured from 21 different regions all over Turkey at different seasonal days. Due to varying background behavior, varying temperatures and weather conditions, and varying lighting conditions, the achieved positive set contains several difficult scenes where low contrast causes close-to invisibility for power lines. The original video resolutions were 576x325 for IR and full HD for VL, however, the captured frames were scaled down to smaller sizes and the effect of resizing was tested for various image sizes. An image size of 128x128 is sufficient for a consistently accurate power line recognition.

尽管航空器安全输电线路检测领域的相关研究文献已较为丰富，但现有工作大多仅基于极少量训练图像进行优化，部分研究甚至仅使用人工生成的图像开展实验。红外（IR）成像场景的相关研究则更少被采用，究其原因，显然在于获取真实红外图像的工作量极为庞大。
基于这一需求，研究团队与土耳其输电公司（Turkish Electricity Transmission Company, TEIAS）合作，从实际运行的航空器上获取了视频录制素材。随后，团队对视频帧进行了全面筛查，从中提取、筛选并清洗出数千张高质量的有效图像。
本数据集共包含4000张红外（IR）图像与4000张可见光（VL）图像，所有图像均被统一缩放至128×128的尺寸。其中，红外图像文件夹内包含2000张含输电线路的红外图像与2000张不含输电线路的红外图像；可见光图像文件夹内则包含2000张含输电线路的可见光图像与2000张不含输电线路的可见光图像。红外与可见光图像组均刻意构建了常规场景与高干扰混淆场景。标签"TY"代表图像中无输电线路，"TV"代表图像中存在输电线路。
所有视频素材均采集自土耳其境内21个不同区域，拍摄时间覆盖不同季节。由于背景环境、温度、天气状况与光照条件均存在差异，最终得到的正样本集包含诸多高难度场景：在部分场景中，输电线路因对比度极低而几乎难以辨识。原始视频的分辨率分别为：红外视频为576×325，可见光视频为全高清（Full HD）。随后团队将采集到的视频帧统一缩放至更小尺寸，并针对多种图像尺寸的缩放效果进行了测试，最终确定128×128的图像尺寸足以实现稳定且准确的输电线路识别。