Deep Feature Extractor with Information-Extreme Decision Rules for Visual Classification of Sewer Pipe Defects and its Training Method

Abstract

CCTV inspections are frequently used to diagnose defects in underground sewer pipes. A model for classification of sewer pipe defects in video frames and a corresponding five-phase training method are proposed. The model is based on deep convolutional neural network with a sigmoid output layer and information-extreme decision rules. The training method includes augmentation, feature extractor training resemblant of a Siamese network with triplet mining and softmax, computation of binary code for each class, training with joint binary cross-entropy loss using binary codes for each class as label and optimization of a hyper-spherical container for each class in Hamming space based on information criterion. Information criterion, expressed in the paper as logarithmic function of accuracy characteristics, provides a robust and reliable model for the most difficult case in the statistical sense. Results of a simulation with the optimal model on a dataset provided by Ace Pipe Cleaning, Inc confirm the efficiency of the proposed model and method and their suitability for practical usage.