Deep learning-based damage detection for sewer pipe inspection using faster R-CNN

Sewer pipe systems are important components of civil infrastructures to provide essential urban services. Severe defects of sewer pipes can lead to disturbance to residential life, serious incidents and public property loss.... [ view full abstract ]

Sewer pipe systems are important components of civil infrastructures to provide essential urban services. Severe defects of sewer pipes can lead to disturbance to residential life, serious incidents and public property loss. Sewer pipe inspection aims to discover pipe defects such as cracks, tree roots and deposits such as to ensure normal sewer operations. Currently, visual inspection techniques especially CCTV robots are commonly utilized and large number of inspection videos or images have accumulated. However, inspectors are required to check and analyse the videos manually to identify the defect type and their locations, which is time consuming, error prone and subjective. Conventional computer vision approaches used in previous studies for automatic sewer pipe defect classification require complex handcrafted feature extraction and are susceptible to external conditions. With the promising performance of deep learning models in computer vision, region-based convolutional neural networks (R-CNNs) have shown greater potentials in both image classification and object detection.

This study proposed a deep learning-based defect detection approach for sewer pipe inspection using faster region-based convolutional neural network (faster R-CNN). 1260 images containing four types of sewer pipe defects were collected from CCTV inspection videos. Data augmentation was conducted to improve the dataset size to 3000 images. The augmented images were annotated with ground-truth classes and bounding boxes, among which 85% were used for training and 15% for testing. The faster R-CNN model was designed and trained with three datasets of different size. Through experiments, the proposed approach was demonstrated to be applicable for identifying the defect type and localizing the defect accurately with a high mean average precision (mAP) and a fast detection speed. In the end, the mAP was further improved to 83% by adjusting the hyper-parameters of the network.