A robot-view image dataset of photovoltaic panel reference lines with paired synthetic soiling for cleaning robot...

This dataset provides robot-view photovoltaic (PV) panel images for reference line perception under visual soiling. The images are captured from a top-down camera viewpoint that approximates the onboard view of a PV cleaning robot. The dataset contains 6200 RGB images at a resolution of $160\times 120$ pixels, with line-instance annotations for three classes: white cell-gap lines, horizontal and vertical frame-cell border lines. This dataset supports real-time navigation of PV cleaning robots under different soiling levels, including heading angle estimation and the trigger signal for path reversal. The training and validation splits include 4800 images derived from 800 manually annotated clean source images. Each source image is stored together with five synthetic soiling variants, forming six soiling conditions: clean, very-light, light, medium, heavy, and extreme. The images in the very-light, light, medium, heavy, and extreme conditions are synthetically produced by a reproducible augmentation pipeline with six foreground effects: bird-dropping, water-stain, puddle, glare, frame-water, and frame-dust. Additional bird-dropping and glare effects are also sampled in background regions away from labeled reference lines. The test split comprises 1400 images: 1200 images stored in six separate folders, one for each soiling condition, and 200 laboratory real-soiling images of two PV panels manually soiled with pigment droplet, sand, wastewater, clean water, and high-intensity light. Each annotation record stores polar line parameters, line segment endpoints, and a structural-line-class identifier. The dataset is designed as a geometry-centered benchmark rather than a general PV inspection collection. Its paired synthetic protocol keeps the structural labels fixed while increasing visual soiling, enabling controlled robustness analysis without changing scene geometry. The augmentation code is released with the dataset to reproduce the synthetic soiling levels from clean annotated images.