Optimizing motion and colour segmented images with neural networks

摘要

Segmentation of independently moving foreground elements from background, is a very common procedure in digital postproduction. The conventional technique, known as rotoscoping, is carried out manually and is therefore too reliant on human effort. The industry is interested in an automated method that can correctly locate the boundary and be robust given rapid motion and non-static backgrounds. A cellular neural network is presented that labels pixels by estimated motion, colour, neighbouring and previous labels. The method is accurate, labour-saving and many times faster than manual rotoscoping. Moreover, due to the inherent parallelism and the local nature of the network, the whole process can be implemented on hardware boosting up performance.