伞齿常见缺陷高效识别数据

数据集包含了不同加工阶段、不同磨损程度的伞齿图像，可用于机械加工质量控制和自动化检测相关的目标检测模型训练。评估模型对具有改进几何特征的伞齿的泛化性能，特别是在微小缺陷的识别和特征描述性能。应用于实际生产检验能快速识别缺陷，从而及时采取措施，减少次品率，提高生产效率和产品质量。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，涵盖伞齿不同加工阶段和磨损程度。表面反射增强：亮度调整：对图像进行亮度和对比度的微调模拟不同光照条件下金属表面的反光效果，增强或减弱图像中的高光溢出，使模型能够适应不同光照强度下的伞齿表面特征。对比度线性变换：对调整后的图像进行对比度线性变换，将图像像素值先减去128，乘以一个随机系数（范围在0.8到1.2之间），最后再加回128。使得金属表面的细微纹理和潜在缺陷更加明显。之后以50%的概率对处理后图像进行随机水平翻转，增强图片多样性和模型鲁棒性。采用COCO格式进行单类"Defect"目标检测标注。使用CVAT工具，通过包围框(x,y,width,height)标注缺陷。标注规则：缺陷最小边长不小于5像素。磨损表现为局部光泽度导致的反光不一致，视为缺陷标注；边缘破损指轮廓不连续或缺口。标注时遇到部分边缘反射不一致视为异常，按缺陷标注。将COCO格式转换为CSV格式：x1=x,y1=y,x2=x+width,y2=y+height,cls="Defect"。cls为类别（单类，缺陷），x1,y1,x2,y2是坐标。

This dataset contains bevel gear images at different processing stages and wear levels, which can be used for training object detection models related to machining quality control and automated inspection. It is designed to evaluate the generalization performance of models on bevel gears with improved geometric features, particularly the ability to identify and characterize tiny defects. Applying this dataset in actual production inspection enables rapid defect identification, allowing timely implementation of corrective measures to reduce the defective product rate, and improve production efficiency and product quality.

This dataset is a de-identified object detection dataset derived from internal test data during patent operation evaluation of several patents, covering bevel gears at various processing stages and wear levels.

Two data augmentation methods are adopted:
1. Surface reflection enhancement and brightness adjustment: Fine-tune the brightness and contrast of the images to simulate the reflective effects of metal surfaces under different lighting conditions, and enhance or suppress overexposed highlights in the images, enabling the model to adapt to the surface features of bevel gears under varying light intensities.
2. Contrast linear transformation: Perform linear contrast transformation on the adjusted images: subtract 128 from the pixel values, multiply by a random coefficient ranging from 0.8 to 1.2, and then add 128 back. This approach makes the subtle textures and potential defects on the metal surface more distinguishable.
Additionally, a 50% probability random horizontal flip is applied to the processed images to enhance image diversity and model robustness.

The dataset uses COCO format for single-class "Defect" object detection annotations. Annotations are created using the CVAT tool, with bounding boxes in the format (x, y, width, height) for defects. The annotation rules are as follows:
- The minimum side length of a defect shall not be less than 5 pixels.
- Wear manifested as inconsistent reflectance caused by local gloss is regarded as a defect and annotated.
- Edge damage refers to discontinuous contours or notches, which are also annotated as defects.
- Inconsistent reflectance on partial edges encountered during annotation is considered abnormal and annotated as a defect.

The COCO format annotations are converted to CSV format with the structure: x1=x, y1=y, x2=x+width, y2=y+height, cls="Defect", where cls represents the category (single-class, "Defect"), and x1, y1, x2, y2 are the coordinates of the bounding box.