基于计算机视觉的钢结构表面锈蚀程度检测方法

A Method for Detecting Surface Corrosion Degree of Steel Structures Based on Computer Vision

针对钢结构表面锈蚀损伤,传统的人工检测方法耗时、费力,且受检测人员技术水平限制。计算机视觉技术为钢结构表面锈蚀的检测和分类提供了一种快速准确的替代方法。目前常用的锈蚀程度检测技术多基于卷积神经网络结构,由于网络结构自身的缺陷,在进行锈蚀程度分类时存在忽略图像中部分锈蚀特征的问题,导致错误的检测结果。为此,提出了一种基于Vision Transformer网络结构的钢结构表面锈蚀程度识别方法,通过引入自注意力机制(SA)在进行特征提取的过程中保证数据的完整性,并在自建的锈蚀程度图像数据集上进行验证,该方法对锈蚀程度的分类准确率可达到90%。此外,还提出一种基于滑动窗口法的钢结构表面锈蚀程度检测方法,对待检测的钢结构图像进行切割,利用训练好的网络结构进行锈蚀程度检测,将检测后的图像重新拼接,实现钢结构表面锈蚀程度的可视化。

Traditional manual inspection methods for surface corrosion damage in steel structures are timeconsuming,labor-intensive,and limited by the technical expertise of the inspection personnel.Computer vision technology provides a fast and accurate alternative method for detecting and classifying the surface corrosion on steel structures.Currently,commonly used methods for corrosion degree detection are based on convolutional neural network(CNN)structures.However,due to inherent flaws in the network structure,there is a problem of neglecting certain rust features in the image during corrosion degree classification,leading to incorrect detection results.A steel structure surface corrosion degree recognition method based on the Vision Transformer network structure was proposed.By introducing self-attention mechanisms(SA)during the feature extraction process,data integrity could be ensured.The proposed method was validated on a self-built dataset of rust severity images,achieving an accuracy ratio of 90%in corrosion degree classification.Furthermore,a steel structure surface corrosion degree detection method based on the sliding-window method was also proposed.This method involves segmenting the steel structure images to be inspected,utilizing a trained network structure for corrosion degree detection,and reassembling the detected images to achieve intelligent detection of corrosion degree on the surface of steel structures.