融合自适应光照预处理和深度学习的钢桥腐蚀检测方法

Steel Bridge Corrosion Detection Method Combining Adaptive Illumination Preprocessing and Deep Learning

针对钢结构桥梁实桥腐蚀图像往往由于光照条件较差或光照不均匀导致腐蚀区域检测困难的问题,提出了一种融合自适应光照预处理方法和深度学习的钢桥腐蚀检测方法。首先,采用Global and Local Fusion(GLF)对比度增强算法结合KinD++低光增强模型的方法,对图像进行预处理;其次,采用粗标注结合K-means算法标注腐蚀区域得到分割标签;最后,采用原始图像和预处理后图像分别对UNet++网络进行了训练和测试,验证了所提出的预处理方法的有效性和优越性。结果表明:所提出的自适应光照预处理方法有效改善了实桥腐蚀图像的光照不均和低光照问题,修复和增强了细节和纹理特征信息,颜色保真度较高;所提出的数据标注方法能够精准标注腐蚀区域,减少边缘描绘工作;与原始图像相比,该方法预处理后的图像训练的模型在准确率、精确率、召回率、F1-score、交并比IoU和AUC上分别提高了5.2%、2.7%、22.5%、19.4%、25.4%和10.5%;对于光照良好的均匀腐蚀图像,预处理对分割精度提高有限,对于点蚀图像,分割精度有较大的提高,对于低光照或光照均匀性较差的图像,分割精度得到了大幅提高。

Detecting corrosion areas from images of steel bridge structures is challenging because of poor or uneven lighting conditions of the bridges.In this study,a novel method integrating adaptive lighting preprocessing and deep learning for steel bridge corrosion detection is proposed.Initially,the images are preprocessed using the global and local fusion contrast enhancement algorithm combined with the KinD++low-light enhancement model.Subsequently,coarse annotation combined with the K-means algorithm is employed to label the corrosion areas for segmentation.Finally,the effectiveness and superiority of the proposed preprocessing method are validated by training and testing the UNet++network with the original and preprocessed images.The results indicated that the proposed adaptive lighting preprocessing method effectively resolves issues of uneven and low lighting in images of bridge corrosion,while repairing and enhancing details and texture feature information with high color fidelity.The data annotation method introduced accurately labels corrosion areas,reducing the need for extensive edge delineation.Compared to models trained with original images,models trained with images processed by the proposed method show improvements of 5.2%in the accuracy,2.7%in the precision,22.5%in the recall rate,19.4%in the F1-score,25.4%in the intersection over union,and 10.5%in the area under the curve.For images with uniform corrosion and good lighting,the preprocessing leads to limited improvement in terms of segmentation precision.However,for pitting images and images with low or uneven lighting,a significant enhancement in segmentation precision is observed.