Sewer Pipe Defects Classification Based on Deep Convolutional Network with Information-Extreme Error-Correction Decision Rules

Abstract

CCTV inspection is a modern approach for diagnosing defects in underground sewer pipes. A new deep model and multi-phase learning method for recognizing sewer pipe defects in fixed parts of video frame are proposed. The deep model is based on a convolutional feature extractor with a sigmoid output layer and information-extreme error-correction decision rules. The first phase of the proposed machine learning method includes feature extractor training on augmented data using triplet mining and softmax triplet loss with regularization term to approximate the discrete representation of data. Next phases aim to reduce intra-class dispersion of discrete data representation by computing averaged discretized representations of each class which are used as labels during training with joint binary cross-entropy loss. Last phase of the earning method is an optimization of a hyper-spherical container for each class in Hamming space based on information criterion. Shannon’s measure of information expressed by logarithmic is used as an information criterion; it provides a reliable solution for the difficult case in the statistical sense. Training results on test data provided by Ace Pipe Cleaning (Kansas City, USA) confirm the efficiency of the proposed model and method and their suitability for practical application.