印刷品表面缺陷检测图像数据

数据包括不同类型的印刷产品在不同光照条件下扫描的图像，可用于印刷品质量控制和自动化检测相关的目标检测模型训练。训练好的模型可以集成到印刷产品的自动化生产线中，实现实时、高效的缺陷检测。这不仅可以减少人工检测的成本和错误率，还可以提高生产线的运行效率和灵活性。模型可以显著提高缺陷检测的准确性和效率，为企业带来显著的经济效益和社会效益。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，数据采集与标注阶段利用高分辨率成像设备对印刷品表面进行扫描。图像经过预处理，采用随机旋转-随机缩放-随机亮度调整-随机对比度调整-高斯噪声对图像增强。为提高模型对细微缺陷的识别能力，引入局部对比度增强算法：首先对原始图像应用高斯模糊，然后将原图像与模糊后的图像做差，得到局部细节图。最后，将这个局部细节图与原图像进行融合，融合程度由增强系数（本数据集中为0.5）控制，有效突出局部细节，使用labelme等标注工具基于视觉特征判断对缺陷的位置进行精确标注，使用COCO包围框(x,y,width,height)进行定位和目标检测标注，将COCO包围框转换为CSV格式。转换公式如下：x1=x,y1=y,x2=x+width,y2=y+height,cls∈{"scratch","spot"," blurred image","Ink bleeding"," bubble"}，cls为类别分别代表划痕、斑点、图像模糊、窜墨、气泡5种情况，filenam为图像，x1,y1,x2,y2是坐标。

This dataset contains scanned images of various types of printed products under different lighting conditions, and can be used for training object detection models related to printed product quality control and automated inspection. The trained models can be integrated into automated production lines for printed products to enable real-time and efficient defect detection. This not only reduces labor detection costs and error rates, but also improves the operating efficiency and flexibility of production lines. The models can significantly improve the accuracy and efficiency of defect detection, bringing substantial economic and social benefits to enterprises.

This dataset is an object detection dataset desensitized from internal test data generated during patent operation evaluation of several patents. During the data collection and annotation stages, high-resolution imaging equipment was used to scan the surface of printed products. The images underwent preprocessing, with image augmentation performed via random rotation, random scaling, random brightness adjustment, random contrast adjustment, and Gaussian noise addition. To improve the model's ability to recognize subtle defects, a local contrast enhancement algorithm was introduced. First, Gaussian blur was applied to the original image, then the difference between the original image and the blurred image was calculated to obtain a local detail map. Finally, this local detail map was fused with the original image, with the fusion degree controlled by the enhancement coefficient (set to 0.5 in this dataset), which effectively highlights local details. Defect locations were accurately annotated using annotation tools such as LabelMe based on visual feature judgment. Object detection and localization annotations were performed using COCO bounding boxes in the format (x, y, width, height), which were then converted to CSV format using the following conversion formula: x1 = x, y1 = y, x2 = x + width, y2 = y + height. The class labels cls belong to {"scratch","spot","blurred image","Ink bleeding","bubble"}, where each class respectively represents scratch, spot, blurred image, ink bleeding, and bubble. Each entry corresponds to an image filename, with x1, y1, x2, y2 representing the bounding box coordinates.