特种设备成果转化库内热成像检测数据

数据来源于不同环境条件（如室内/室外、昼/夜）和不同温度范围的各种特种设备、建筑物和工业设施的热成像扫描，可用于热泄漏检测、能源效率评估和预防性维护相关的目标检测模型训练。通过对建筑物进行热成像扫描，利用数据集训练的目标检测模型可以快速识别出墙体、屋顶、窗户等部位的热量泄漏点，为建筑物节能改造提供精准的数据支持，帮助业主和管理者制定有效的节能措施。在工业上，通过热成像检测和数据集验证的模型，可以及时发现并定位这些设备的热泄漏问题，促进工业设施的能效提升和安全生产。本数据是评估若干专利进行专利运营时的内部测试数据脱敏后的目标检测数据集，包含各种特种设备、建筑物和工业设施的热成像扫描，对图像预处理，保留原始热成像分辨率，采用随机裁剪-随机旋转-随机翻转-温度偏移（±5摄氏度）对图像进行增强。为提高模型对温度边界的感知能力，引入温度梯度感知预处理算法。该算法首先计算温度梯度图，使用Sobel算子计算X和Y方向的梯度，并取最大值作为温度梯度。然后，将这个温度梯度图与原始特征图进行融合，融合程度由系数λ（选用0.2）控制，有效增强温度边界的特征，同时保留原有的温度分布信息，有助于模型更好地识别热泄漏区域。标注工具使用CVAT，采用COCO格式(x,y,width,height)包围框进行单一类别目标检测标注，类别定义为"THERMAL-LEAK"。标注时，只关注图中温度最明显的部分，其余温差不大的忽略不标。
将COCO格式的各个包围框转换为CSV格式：
x1=x,y1=y,x2=x+width,y2=y+height,cls="THERMAL-LEAK"，代表热泄漏。cls为类别，x1,y1,x2,y2是坐标。

This dataset is sourced from thermal imaging scans of various special equipment, buildings and industrial facilities under different environmental conditions (e.g., indoor/outdoor, day/night) and varying temperature ranges, and can be used for training object detection models related to thermal leak detection, energy efficiency assessment and predictive maintenance. Through thermal imaging scans of buildings, the object detection models trained with this dataset can quickly identify heat leak points on walls, roofs, windows and other parts, providing accurate data support for building energy retrofitting and helping property owners and managers formulate effective energy-saving measures. In industrial scenarios, models validated and trained using this dataset can timely detect and locate thermal leaks of such equipment, promoting energy efficiency improvement and safe production of industrial facilities. This is a desensitized object detection dataset derived from internal test data used during patent operation evaluation for several patents. It contains thermal imaging scans of various special equipment, buildings and industrial facilities. Image preprocessing was performed while retaining the original thermal imaging resolution, and image augmentation was conducted using the pipeline: random cropping → random rotation → random flipping → temperature shift (±5°C). To improve the model's perception of temperature boundaries, a temperature gradient-aware preprocessing algorithm was introduced. This algorithm first calculates the temperature gradient map: the Sobel operator is used to compute gradients in the X and Y directions, and the maximum value is taken as the temperature gradient. Then, this temperature gradient map is fused with the original feature maps, with the fusion strength controlled by the coefficient λ (set to 0.2), which effectively enhances the features of temperature boundaries while preserving the original temperature distribution information, helping the model better identify thermal leak regions. The annotation tool used was CVAT, and single-class object detection annotations were performed using COCO format (x, y, width, height) bounding boxes, with the category defined as "THERMAL-LEAK". During annotation, only the parts with the most significant temperature differences in the image are focused on, while areas with minor temperature differences are ignored and not annotated. Convert each bounding box in COCO format to CSV format: x1=x, y1=y, x2=x+width, y2=y+height, cls="THERMAL-LEAK", where "THERMAL-LEAK" represents thermal leaks. cls denotes the category, and x1, y1, x2, y2 are the bounding box coordinates.

搜集汇总

数据集介绍

特点

特种设备成果转化库内热成像检测数据集包含1265条热成像扫描数据，主要用于热泄漏检测和能源效率评估。数据经过多种预处理和增强方法，标注为'THERMAL-LEAK'类别，适用于目标检测模型训练。