In this paper, a new methodology consisting of space-varying templates in cellular nonlinear networks (CNNs) for real-time visual pattern recognition in nuclear fusion devices is presented. The development of space-varying templates is a new upgrade, driven by the need to process different parts of the images in different ways. The new approach has been applied to the identification in real time of various objects present in the Joint European Torus videos of both infrared (IR) and visible cameras. IR videos are here used to detect hot spots and the regions of the walls in which dangerously high temperatures are reached, whereas visible cameras provide information about multifaceted asymmetric radiations from the edge, which are dangerous instabilities that can lead to disruptions. Their identification is particularly difficult because of their movement and their shape which is similar to other objects present in the frames. Therefore, in addition to space-varying template CNNs, quite sophisticated morphological operators have to be deployed and their outputs processed by machine learning tools, such as support vector machines. The implementation of the whole methodology was performed in a field-programmable gate array board, obtaining, in both applications, a final success rate close to 100% and a frame rate higher than 200 frames/s.