Method, device and storage medium for non-destructive semantic segmentation of high-resolution images of concrete cracks

Abstract

The invention discloses a method, device and storage medium for non-destructive semantic segmentation of high-resolution images of concrete cracks, wherein the method includes: obtaining high-resolution images of concrete surfaces; using sliding windows to intercept high-resolution images of concrete surfaces into several partial images; Input the pre-trained concrete crack identification model one by one, and screen out the local images with crack probability greater than the preset threshold value; input the screened local images into the pre-trained concrete crack semantic segmentation model one by one, corresponding to Output the semantic segmentation map of concrete cracks classified pixel by pixel, and realize the lossless semantic segmentation of high-resolution images of concrete surfaces. This scheme can non-destructively identify the crack pixels in the original image, thereby ensuring the pixel recognition accuracy of the crack; it can effectively alleviate the category imbalance, excessive calculation, too many false cracks and the presence of false cracks in the original high-scoring image caused by other semantic segmentation models. Problems such as messy distribution.

Description

technical field [0001] The invention relates to the field of concrete safety monitoring, in particular to a method, device and storage medium for non-destructive semantic segmentation of concrete crack high-score images. Background technique [0002] In many cases, it is necessary to monitor and detect the safety status of concrete. Among them, patrol inspection is an indispensable part of concrete safety monitoring. Through patrol inspection, cracks and other defects that affect structural safety can be found in time. However, patrol inspection currently mainly relies on manual work, which inevitably has problems such as limited inspection space, time-consuming and labor-intensive inspection, and difficulty in controlling crack details and development process. [0003] In recent years, deep learning technology represented by Convolutional Neural Networks (CNN) has continuously surpassed previous image processing technologies in image classification tasks, and is gradually b...

AI Technical Summary

Problems solved by technology

However, due to the large amount of background information during model training, if all local images are used for semantic segmentation, it will lead to extreme category imbalance and excessive calculation

If only partial images with cracks are selected for semantic segmentation, a large amount of background information will be ignored, thereby reducing the ability of the model to recognize background information. For images with a large amount of background information, more false cracks will inevitably appear Phenomenon

Images