Learning matte extraction in green screen images with MLP classifiers and the back-propagation algorithm

This paper considers the problem of automatically extracting a foreground element with its alpha matte in green screen images by training a multi-layer perceptron (MLP) with the back-propagation algorithm. The classifier learns to identify green backgrounds, foreground object contours, and the corresponding alpha values for subsequent digital compositing. We developed our own dataset to train and test the MLP for alpha matte extraction. To speed up the generation of the training set, a second method for semi-automatic alpha matte extraction is proposed. Different experiments show that automatic matte extraction, based on the MLP, generates high-quality matting visually and it is also shown that results depend on the training and the architecture of the classifier. To the best of our knowledge, this is the first effort that applies neural networks to the problem of alpha matte extraction.