Two-dimensional defocusing correction using artificial neural nets

The aim of this work is to show that properly trained Artificial Neural Nets (ANN) can be used as an image restoration tool to correct the effect of defocusing on 2D optical microscopy images. The proposed method can be applied to correct the results of inaccurate range focusing algorithms on fully automated imaging and analysis systems and to compensate the effect of the limited of the depth of focusing on high numerical aperture system. One type of ANN was used : Feedforward Multilayer Perceptron, with supervised backpropagation training. Its performance has been tested with both synthetic images (artificially defocused) and real images (in-focus and out-of-focus) from latex microspheres. The network was trained with sets of pairs of images : each pair consisted of a defocused image and its corresponding in-focus version. Different levels of defocusing were used. The criteria used to select the algorithm parameters to tune the networks and to train them will be presented. The results of the experiments performed to test their ability to 'learn' to correct the defocusing and to generalize the results will also be shown. The results show that, when trained with images with some levels of defocusing, the network was able to learn and accurately correct these defocusing levels, but it can also generalize the results and correct other levels of defocusing.