Amphi 3 - SESSION 1.3 - 07/02/2018 12:00 > 12:30
Defect detection and classification in sewer pipelines
The aim of the proposed work is to help automate the analysis of the defects in sewer pipelines videos
and reduce the time spent on this task by the operators. Different approaches have been proposed in the
literature to detect defects and classify them in sewer pipes. In some of them, local descriptors are
computed, then, a classifier is applied. More recently, some deep learning-based methods have been
developed. However, a huge amount of data is needed to train these models, hence this operation is
computationally expensive. In this work, we propose a model with a lower training computation time than
a neural networks-based model. As opposed to the datasets used in the literature, the images extracted
from our dataset present high variability among the pipelines and the lightning conditions and the number
of observed defects is high. Our model handles this variability and, on top of that, its architecture attempts
to deal with the high imbalance in the proportion of observed defects.
2.Methodology of the supervised machine learning framework and results
- Preprocessing: ensure the quality of:
the images extracted from the inspection videosand their labellisation.
- Feature engineering: construction of local descriptors
- Train the models to detect defects and classify them while dealing with imbalance.
- Test the models on a validation set.
The detection of defects reaches 87% of accuracy which is similar with state of the art performances. Due
to the fact that many classes of defects are underrepresented, the classification of the defects can be
improved. Nevertheless, this framework is a step forward towards the automation of sewer pipelines
For future work, we expect to collect more samples of underrepresented classes in order to reduce the
dataset imbalance and, if the number of samples is sufficient, a neural network will be trained and its
performances compared with our framework's. If the imbalance in the dataset is not reduced in spite of
the new collected data, we plan to explore neural networks methods based on few shots learning.
Télécharger les slides