Research On Defect Detection Method Of Sewer Pipeline Based On Deep Active Learning

Sewer pipelines play an irreplaceable role in urban construction.With the increase of service life,various types of defects inevitably appear inside the pipeline.These defects seriously affect the service life of sewer pipelines and even affect the health and safety of urban people.Therefore,it is very important to detect whether there are defects inside the sewer pipeline in time and formulate a professional repair plan.Due to the large number of pipelines and the complex internal environment,the traditional manual observation efficiency is low.The target detection technology based on deep learning aims to judge the category and location of pipeline defects by training the learning model,which can greatly reduce the cost of manual observation and improve the detection efficiency.At present,many scholars have applied target detection technology to intelligent detection of pipeline defects.However,there are still two main problems: many scholars have not optimized the model structure according to the characteristics of pipeline internal defect images,resulting in low model detection accuracy;lack of large-scale labeled data sets,model training labor costs are high.In view of the above problems,a sewer pipeline defect detection method based on deep learning and active learning is proposed.The main work is as follows:1.Aiming at the problems of low detection accuracy and poor robustness of current deep learning models,an object detection algorithm IYOLOv5 suitable for sewer pipeline defect detection is improved.In this improved algorithm,the original CSPNet module is replaced by the translation variable convolution network module TVCSPNet.This module improves the ability of the model to capture the dependencies between different pixels in the feature map by enhancing the perception level of the model to specific image structures.At the same time,compared with the original module,it greatly reduces the number of model parameters.Secondly,before the model prediction head,the ECA-Net attention module is added to enhance the cross-channel interaction of the model.By improving the boundary box loss function of YOLOv5,α-EIOU is used instead of CIOU loss.The purpose is to accelerate the convergence speed of the model and facilitate the tuning of hyperparameters under different thresholds.Finally,the prior bounding box of the model is re-clustered to make the size of the prior bounding box closer to the actual value.2.The sewer pipeline defect data set required for model training needs to be manually labeled,which greatly increases the training cost.With the help of active learning ideas,the value loss prediction module VLPNet is constructed.The most representative samples are actively selected by the model for manual labeling,so that the model can obtain the maximum performance improvement.The value loss prediction function is designed.By comparing the value priority training loss prediction modules of different samples,the loss prediction module and the target detection algorithm model are end-to-end trained to converge together.3.Combining the IYOLOv5 object detection algorithm with the VLPNet value loss prediction module,a deep active learning framework IYOLOv5-VLPNet is proposed.IYOLOv5-VLPNet can start from zero number of labeled samples,and drive IYOLOv5 to select samples that need to be annotated for next training through the VLPNet module.Through active selection of samples and iterative training between detection models,this study can use the lowest manual labeling cost to improve the representation ability of deep learning models.The algorithm in this paper is compared with the mainstream target detection algorithm.The experimental results show that the improved model IYOLOv5 proposed in this paper can effectively improve the detection accuracy of sewer pipeline defects in different scenarios.Through the efficient sample selection ability of VLPNet,the cost of manual sample labeling is minimized,and the task of pipeline internal defect detection under the condition of small labeled samples is effectively completed.