BeltCrack: the First Sequential-image Industrial Conveyor Belt Crack Detection Dataset

Dataset IntroductionThe BeltCrack Datasets are designed for industrial conveyor belt crack detection, consisting of two specialized sub-datasets: BeltCrack14ks (strong saliency, dense crack scenes) and BeltCrack9kd (weak saliency, sparse crack scenes), tailored to address diverse industrial inspection challenges.Data Collection & ProcessingData was captured in real-world industrial environments using portable smartphones mounted on adjustable tripods to minimize motion blur. The cameras operated at 30 FPS with 16:9 aspect ratio resolutions (3840×2160 and 1920×1080). Capturing conditions were deliberately diversified to ensure robustness:• Multi-perspective views: top-down and bottom-up angles• Time-varying illumination: from strong morning light to low evening illumination• Multi-weather scenarios: sunny, rainy, and snowy conditions• Dynamic belt speeds: ranging from 2 to 6 m/sOriginal three-channel RGB video streams were split into frame sequences, followed by frame-by-frame pixel-level bounding box annotation, and finally structured into a hierarchical format compatible with the COCO specification. The dataset currently focuses on a single crack class, with inherent support for future multi-class expansion.Sub-datasets DetailsBeltCrack14ks: Contains 14,087 images across 29 sequences, split into training, validation, and test sets at a 6:2:2 ratio (8,773 training, 2,596 validation, 2,718 test images). It has an average of 485.76 images per sequence and 2.65 cracks per image, featuring strong crack saliency and dense crack distributions. Cracks are categorized by pixel area into tiny (<100), small (100–1000), medium (1000–10000), large (10000–100000), and huge (≥100000), with a long-tailed normal area distribution.BeltCrack9kd: Comprises 9,645 images from 42 sequences, split into training and test sets at an 8:2 ratio (7,697 training, 1,948 test images). It has an average of 229.64 images per sequence and 1.34 cracks per image, characterized by weak crack saliency and sparse, discrete crack distributions. It also follows the same crack area categorization and exhibits higher crack discreteness, posing greater challenges for sparse crack detection.

数据集简介
BeltCrack数据集专为工业传送带裂纹检测任务打造，包含两个专用子数据集：BeltCrack14ks（高显著性、密集裂纹场景）与BeltCrack9kd（低显著性、稀疏裂纹场景），可应对多样化的工业巡检挑战。

数据采集与处理
本数据集采集自真实工业场景，采用安装于可调三脚架的便携智能手机进行拍摄，以尽可能降低运动模糊。拍摄设备帧率设为30 FPS，采用16:9宽高比分辨率（3840×2160与1920×1080）。为保证数据集的鲁棒性，采集场景被刻意设置为多样化：
• 多视角拍摄：包含俯视与仰视两种角度
• 光照条件随时间变化：从清晨强光到傍晚弱光的全区间覆盖
• 多天气场景：涵盖晴天、雨天与雪天
• 传送带动态速度范围：2~6 m/s

原始三通道RGB视频流先被拆解为帧序列，随后对每一帧进行像素级边界框标注，最终整理为适配COCO（Common Objects in Context）规范的层级化数据格式。目前本数据集仅针对单裂纹类别设计，但预留了后续扩展至多类别检测的能力。

子数据集详情
BeltCrack14ks：包含29个序列共计14087张图像，按照6:2:2的比例划分为训练集、验证集与测试集（训练集8773张、验证集2596张、测试集2718张）。该子数据集单序列平均图像数为485.76张，单张图像平均裂纹数为2.65个，具有裂纹显著性高、分布密集的特点。裂纹按像素面积分为极小（<100）、小（100~1000）、中（1000~10000）、大（10000~100000）与超大（≥100000）五个类别，面积分布呈长尾正态分布。

BeltCrack9kd：包含42个序列共计9645张图像，按照8:2的比例划分为训练集与测试集（训练集7697张、测试集1948张）。该子数据集单序列平均图像数为229.64张，单张图像平均裂纹数为1.34个，以裂纹显著性低、分布稀疏离散为主要特征。其裂纹面积分类标准与BeltCrack14ks保持一致，且裂纹离散程度更高，对稀疏裂纹检测任务提出了更大挑战。