A simple and elegant method to design image filters with neural networks is proposed: using small networks that scan the image and perform position invariant filtering. In the theses examples of image filtering with error backpropagation networks for edge detection, image deblurring and noise suppression are given. The main advantages with respect to other filter design methods are: ease of design and the possibility to tune the filters for a specific application and type of image. The main disadvantage is that no insight in the underlaying processes is gained, only a filter design is obtained. For the evaluation of the performance of image filters the average risk (AVR) is proposed as a performance criterion. The AVR has a solid basis in statistical classification theory. It consists of a weighted sum of the probabilities on different types of errors. The errors are weighted with a cost function. Thus both the frequency and the costs of errors are taken into account. An extension of the theory of AVR is proposed to handle multiple and continuous output. It is shown that the AVR can be regarded as a distance measure. As an example an edge detector evaluation method, based on AVR is described. Finally the optimisation of neural networks for minimum AVR is investigated. An example illustrates that this can lead to considerable performance improvements with respect to the optimisation using the 'standard' optimisation criteria as used in the learning rules of neural networks.
|Award date||6 Nov 1992|
|Place of Publication||Enschede|
|Publication status||Published - 6 Nov 1992|