Image Processing Using Masked Restricted Boltzmann Machines

An object recognition system comprising:

• an input arranged to receive an image of a scene depicting at least one object;
• a memory storing a data structure holding a first field of masked restricted Boltzmann machines each comprising a plurality K of restricted Boltzmann machines each arranged to receive input from a patch of the image and having shared parameters and also comprising an image mask having a variable for each image element specifying which of the plurality of masked restricted Boltzmann machines is to be used for that image element;
• a processor arranged to input the image to the first field of masked restricted Boltzmann machine and obtain first superpixel elements comprising hidden node activations of the plurality of restricted Boltzmann machines;
• wherein the memory further comprises a second field of masked restricted Boltzmann machines and the processor is arranged to input the first superpixel elements to the second field of masked restricted Boltzmann machines and obtain second superpixel elements comprising hidden node activations of the restricted Boltzmann machines of the second field;
• a classifier arranged select an object class from a plurality of specified object classes on the basis of at least some of the first and second superpixel elements such that the at least one object depicted in the image is a member of the selected object class.

A method of processing an image comprising:

• at an input receiving an image of a scene;
• at a memory storing a data structure holding a first field of masked restricted Boltzmann machines each comprising a plurality K of restricted Boltzmann machines each arranged to receive input from a patch of the image and having shared parameters and also comprising an image mask having a variable for each image element specifying which of the plurality of masked restricted Boltzmann machines is to be used for that image element;
• at a processor inputting the image to the first field of masked restricted Boltzmann machines and obtaining first superpixel elements comprising hidden node activations of the plurality of restricted Boltzmann machines;
• and storing the superpixel elements at the memory.