Ober Pipe Condition Classifier

Introduction

We are building a deep learning CNN algorithm for drainage pipe condition classification. We are training this system on manually labelled imagery from drainage CCTV - hours and hours of it.

Aims

We are trying to build a model, which will correctly classify drainage condition from CCTV footage without human input.

Methodology

We collect and label the CCTV survey imagery, group the images into categories of respective condtion, and load it into Keras/TensorFlow, CNN sequential model and train the system from scratch. We then compile our model and run the algo on a set of previously unseen imagery. We then manually correct the output and train a new model on a larger dataset. It's an iterative process.

Conclusions

We completed the first phase of the project and the resutls are satisfactory for some pipe condition scenarios - the model performs well on pipe with water and is rubbish at distinguishing if the pipe is damaged, filled with rubbish/rubble or completely collapsed. We are going to need...