摘要
数据标注为机器学习提供了大量带标签的数据,在数据集制作时需要借助各种标注工具手动对图像进行画框标注,受主观因素影响较大,且工业现场环境复杂,采集到的图像质量不稳定,也会影响标注效果。因此提出一种改进MSR(Multi-scale retinex)钢板缺陷图像数据集增强算法和基于图像分块和像素差分的自适应目标框标注算法,首先在MSR基础上提出一种自适应权值计算方法对采集到的缺陷图像进行增强,通过计算信息熵占比自动确定权值Wk,克服了传统MSR算法需要人工调整权值的缺点;然后为了解决直接对整幅图像提取目标边界计算量太大的问题,提出一种分块计算像素差分的方法,分别计算每个子块图像的均值矩阵和2阶差分矩阵,通过判别目标在各个子块的分布情况,选取合适的子块分别计算矩形框的4个边界,代替人工画框辅助数据集的标注,并采用Faster R-CNN和YOLOv5进行缺陷检测验证。结果表明:提出算法的平均IoU为0.87,平均检测时间为457 ms,在公开数据集上的平均IoU和检测时间分别为0.84和473 ms,性能均优于其他方法,基于提出算法Faster R-CNN和YOLOv5的检测准确率分别提升了4.8%和5.9%,可以为深度学习模型提供质量稳定的数据集。
Dataset annotation provides a large amount of labeled data for machine learning.In the dataset production,it needs to draw a box manually for annotation by using the various annotation tools.It is greatly affected by subjective factors.Moreover,due to the complex industrial field environment and unstable image quality,it is difficult to achieve the annotation effect.Therefore,an improved MSR(Multi-scale retinex)steel plate defect dataset enhancement algorithm and an adaptive target box annotation method based on the pixel difference are proposed.Firstly,based on MSR,an adaptive weight calculation method was proposed to automatically determine the weight Wk by calculating the image information entropy without manual adjustment.And the collected defect image was enhanced.Then,it was too much to calculate the pixel difference and extract the target boundary directly for the whole image,so a block calculation method was proposed,and the mean matrix and the second-order difference matrix of each sub block were calculated respectively.By considering the distribution of the target in each sub block,the appropriate sub block was selected to calculate the four boundary of the rectangular box.It assists the defect dataset annotation instead of the manual method.The average IoU is 0.87 and the average detection time is 457 ms,and the average IoU and detection time on the open dataset are 0.84 and 473 ms,respectively.The performance is better than that via the other methods.The detection accuracy of Faster R-CNN and YOLOv5 based on the present algorithm are improved by 4.8%and 5.9%respectively,which can provide datasets with stable quality for the deep learning.
作者
杨璐雅
黄新波
任玉成
韩琪
YANG Luya;HUANG Xinbo;REN Yucheng;HAN Qi(School of Mechano-Electronic Engineering,Xidian University,Xi'an 710071,China;School of Electronics and Information,Xi'an Polytechnic University,Xi'an 710048,China;China National Heavy Machinery Research Institute Co.,Ltd.,Xi'an 710032,China)
出处
《机械科学与技术》
北大核心
2025年第3期445-452,共8页
Mechanical Science and Technology for Aerospace Engineering
基金
中国重型机械研究院金属挤压与锻造装备技术国家重点实验室开放课题(N-KY-ZX-1104-201911-5881)。
关键词
数据标注
深度学习
数据集增强
像素2阶差分
自适应目标框标注
data annotation
deep learning
datasets enhancement
pixel second-order difference
adaptive bounding box annotation
