Truncated edges estimation using MLP neural nets applied to regularized image restoration

This paper describes the implementation of a method for image restoration based on neural nets. The idea is to apply a well-known deterministic regularisation function in an iterative energy minimisation procedure to a multilayer perceptron (MLP). At every iteration, the neural net simulates an approach to both the degradation process, using the aperiodic model (backward), and the restoration process (forward), including the truncation (loss) of the borders and their regeneration (approach) with no a priori knowledge, assumption, or estimation concerning those lost borders. As no net training is required, no images are involved in this non-existent process. Within the strategy of global energy minimisation, the borders are regenerated while adapting the centre of the image to the optimum linear solution. The proposed tests compare the restoration results with those given by the pure aperiodic model (no border truncation) and the circulant model. The restoration error is computed for different concentric areas of the image, showing the adaptation provided by the neural net up to the borders. Moreover, pure aperiodic and circulant models can be obtained as particular implementations of the method proposed in this work.