Unsupervised crack identification method based on sliding window variational auto-encoder

Abstract

The invention discloses an unsupervised crack identification method based on a sliding window variational auto-encoder. The method comprises the following steps: acquiring a target image; the target in the target image has a crack; determining a sliding window image sequence of the target image; the sliding window image sequence comprises a plurality of sliding window images; inputting the sliding window image into a variational automatic encoder, and outputting a reconstructed sliding window image corresponding to the sliding window image through the variational automatic encoder; determining a reconstructed image corresponding to the target image according to all the reconstructed sliding window images; the target in the reconstructed image does not have cracks; and determining a crack in the target image according to the reconstructed image and the target image. According to the method, the variational automatic encoder is used for eliminating the abnormal value in the target image, so that the reconstructed image without the crack is obtained, and the crack in the target image is determined. According to the method, positive sample training or crack-specific image feature learning is not needed, so that the problem of relatively low crack recognition accuracy caused by lack of training data or unbalanced data sets is avoided. 本发明公开了一种基于滑窗变分自编码器的非监督裂缝识别方法，包括步骤：获取目标图像；目标图像中的目标存在裂缝；确定目标图像的滑窗图像序列；滑窗图像序列包括多个滑窗图像；将滑窗图像输入变分自动编码器，以通过变分自动编码器输出滑窗图像对应的重建滑窗图像；根据所有重建滑窗图像，确定目标图像对应的重建图像；重建图像中的目标不存在裂缝；根据重建图像和目标图像，确定目标图像中的裂缝。由于本申请是利用变分自动编码器消除目标图像中的异常值，得到不存在裂缝的重建图像，从而确定目标图像中的裂缝。本申请不需要正样本训练或学习特定于裂缝的图像特征，从而避免了缺乏训练数据或不平衡数据集带来的裂缝识别准确性较低的问题。