Wavelet Entropy-based Feature Extraction for Crack Detection in Sewer Pipes

This paper describes the use of wavelet entropy as a feature extractor for robust classification of cracks in sewer pipe structures. Video image data was acquired using an infrared camera in an experimental sewer pipe setup. Video frames were partitioned into 64 ∗ 64 pixel sections. 1,885 ’crack’ and 1,675 ’clean’ (non-crack) sections were manually classified. Each section was transformed to a combined spacefrequency representation using a Haar wavelet transform. The concentration of wavelet coefficient distribution energy in 15 orthogonal wavelet subspaces was estimated using Shannon entropy, and the extracted features were used as inputs to a logistic regression model. An average validation set classification rate of 88.31% was obtained over 10 runs.