Feature Extraction of Sewer Pipe Defects Using Wavelet Transform and Co-Occurrence Matrix

In general, the sewer inspection usually employs a great number of CCTV images to discover sewer failures by human interpretation. A computer-aided program remains to be developed due to human's fatigue and subjectivity. To enhance the efficiency of sewer inspection, this paper attends to apply artificial intelligence to extract the failure features of the sewer systems that is demonstrated on the sewer system in the eastern Taichung City, Taiwan. Wavelet transform and gray-level co-occurrence matrix, which have been widely applied in many texture analyses, are adopted in this research to generate extracted features, which are the most valuable information in pattern recognition of failures on CCTV images. Wavelet transform is capable of dividing an image into four sub-images including approximation sub-image, horizontal detail sub-image, vertical detail sub-image, and diagonal detail sub-image. The co-occurrence matrices of horizontal orientation, vertical orientation, and 45° and 135° orientations, respectively, were calculated for the horizontal, vertical, and diagonal detail sub-images. Subsequently, the features including angular second moment, entropy, contrast, homogeneity, dissimilarity, correlation, and cluster tendency, can be obtained from the co-occurrence matrices. However, redundant features either decrease the accuracy of texture description or increase the difficulty of pattern recognition. Thus, the correlations of the features are estimated to search the appropriate feature sets according to the correlation coefficients between the features. In addition, a discriminant analysis was used to evaluate the discriminability of the features for the pipe failure defection, and entropy, correlation, and cluster tendency were found to be the best features based on the discriminant accuracy through an error matrix analysis.