Convolutional Neural Networks for Image Processing: An Application in Robot Vision

Convolutional neural networks (CNNs) represent an interesting method for adaptive image processing, and form a link between general feedforward neu-ral networks and adaptive filters. Two dimensional CNNs are formed by one or more layers of two dimensional filters, with possible non-linear activation functions and/or down-sampling. Conventional neural network error minimization methods may be used to optimize convolutional networks in order to implement quite powerful image transformations. CNNs possess key properties of translation invariance and spatially local connections (receptive fields). CNNs are an interesting alternative when the the input is spatially or temporally distributed, and the desired output of a system may be specified. The present paper presents a description of the convolutional network architecture, and an application to a practical image processing application on a mobile robot. As a formal CNN framework has not yet been specified in the literature, we describe CNNs in some detail, conceptually and formally. A CNN is used to detect and characterize cracks on an autonomous sewer inspection robot. Although cracks are relatively easy to detect by a human operator, autonomous sewer inspection necessitates the detection of pipe damage using computer vision methods. This is an appropriate application for trainable data-based computer vision methods, since prior specification of appropriate of the filtering / detection method is quite difficult. The The CNN architecture used involved a total of five layers: a single input and output map, and three hidden layers. The filter sizes used in all cases were 5x5, and the common activation function used was a log-sigmoid. The number