DefectAtlas-20: A Real-World Image Dataset for Joint Material and Surface-Condition Classification

ScratchAtlas-20 is a real-world image dataset for joint material and surface-condition classification. It contains 13,233 RGB images organized into 20 mutually exclusive classes, formed by combining ten commonly encountered material categories with scratched and non-scratched surface conditions. The dataset comprises 7,060 defect images and 6,173 no-defect images. The included material categories are uncoated metal, coated metal, uncoated wood, coated wood, rubber, textile, cardboard, glass, cable, and hard plastic. The dataset addresses the limited availability of image collections that capture material-dependent scratch appearance under real-world conditions. Existing surface-defect datasets often focus on a single material, component, defect type, or controlled inspection environment. ScratchAtlas-20 instead covers heterogeneous everyday objects and surfaces photographed under varying illumination, viewing angles, backgrounds, surface textures, reflectance conditions, and signs of use. Images were captured by multiple contributors using consumer smartphone cameras at indoor and outdoor locations in Hamburg and the surrounding metropolitan area, Germany, between October and December 2025. Contributors followed a predefined acquisition protocol specifying camera distance, viewing angle, focus, lighting, background, and image quality. Flash, digital zoom, and live-photo capture modes were prohibited. Scratched surfaces contained naturally occurring damage, while non-scratched images depicted surfaces from the same or visually comparable material categories without visible scratches. Blurred, poorly exposed, duplicated, mislabeled, or visually ambiguous images were excluded through contributor review and centralized quality control. The Zenodo repository contains two files: images.zip: the complete image dataset containing all 13,233 images resized to 512 × 512 pixels overview.csv: the accompanying image-level labels and metadata All released images are stored as lossless PNG files in the sRGB color space. Each image has a spatial resolution of 512 × 512 pixels. File names follow the pattern [class]_[identifier].png, for example, 00_glass_defect_00001.png. The class name consists of a two-digit class index, the material category, and either defect or no_defect. Within images.zip, the images are organized into one directory for each of the 20 material–surface-condition classes. The overview.csv file contains one row for each image. The sample_name column provides the unique image file name, while the class, material, and defect_label columns provide the corresponding 20-class label, material category, and binary surface-condition label. Scratched images are represented by the label defect, and non-scratched images by no_defect. The width and height columns document the released image dimensions. The CSV file additionally contains a non-identifying subset of the available acquisition metadata. These fields comprise camera_make, camera_model, lens_model, focal_length_mm, focal_length_35mm_equiv, f_number, exposure_time_s, iso, and source_format. The source_format column records the format of the original image before conversion to PNG. Acquisition metadata may be unavailable for individual images where the corresponding information was not present in the original source file. Potentially identifying metadata, including GPS coordinates, capture timestamps, embedded thumbnails, device identifiers, and owner-related information, are not included in the released files. All images were re-encoded at the pixel level before publication so that metadata from the original source files were not retained in the released PNG images. The primary intended task is 20-class material–surface-condition classification. The structured labels can additionally be decomposed to support: ten-class material classification binary defect recognition material-specific scratch classification robustness and domain-generalization studies transfer learning and foundation-model evaluation analysis of material-specific error patterns and cross-material transfer ScratchAtlas-20 is intended for research on automated visual inspection, material-aware computer vision, surface-condition assessment, maintenance documentation, refurbishment, resale assessment, and warranty-related screening. The current release focuses on image-level classification and may also serve as a basis for future extensions with scratch-localization or pixel-level segmentation annotations.