A sewer pipe defect dataset optimization method based on ResNet50 middle layer feature embedding

Abstract

The application relates to the technical field of pipeline detection, and specifically discloses a drainage pipeline defect data set optimization method based on ResNet50 middle layer feature embedding, which comprises the following steps: acquiring detection video data of a drainage pipeline to obtain an image sequence to be processed; performing cross-frame labeling on the image sequence to be processed to generate an initial drainage pipeline defect image sequence containing defect positions and category labels; screening and discriminating the initial drainage pipeline defect image sequence by adopting a multi-stage deduplication strategy to output a deduplicated drainage pipeline defect image data set; and the multi-stage deduplication strategy comprises pre-screening based on a perceptual hashing algorithm and feature discrimination based on middle layer features of a convolutional neural network. The drainage pipeline defect data set optimization method based on ResNet50 middle layer feature embedding solves the problems that video data redundancy in drainage pipeline detection, similar frame interference and analysis efficiency lead to difficulty in realizing accurate deduplication and retaining defect features.

Images