Method and circuits for scaling images using neural networks

A method for processing a small block pattern to generate a large block pattern related thereto having a different number of components than the small block pattern, the method comprising:

• a) providing a neural system comprising an artificial neural network including neurons capable of storing said small block pattern or data related thereto as a prototype with a category associated therewith and a memory configured to store coefficients that will subsequently allow determination of the estimated values of the large block pattern in a recognition phase, these coefficients being the components of an intermediate pattern;
• b) providing a selected small block pattern for recognition derived from an input image;
• c) applying said selected small block pattern to a dedicated processor that transmits said input selected small block pattern with or without processing to said artificial neural network to extract the category of the closest prototype stored therein, which in turn acts as a pointer to output a closest intermediate pattern from the memory;
• and
• d) applying said input selected small block pattern and said closest intermediate pattern to said processor for facilitating generating of an up-converted image of the input image using said coefficients of the closet intermediate pattern.

A method for learning an input image consisting of n large blocks of m pixels and a down-converted image consisting of n small blocks of p pixels (p<m), the method comprising:

• a) initializing a block indicia i,;
• b) deriving a large block pattern from a large block i having m components of the input image;
• c) down-converting the large block i of m pixels into a small block i of p pixels, to produce a small block pattern having p components;
• d) applying said small block and large block patterns to a dedicated processor for computing the coefficients of a transform function describing conversion between the small block and large block patterns, the coefficients subsequently facilitating up-converting of a selected small block pattern in a recognition phase, these coefficients being the components of an intermediate pattern;
• e) storing the small block pattern in an artificial neural network composed of neurons capable of storing said small block pattern or data related thereto as a prototype with a category associated with the prototype;
• f) storing said intermediate pattern in a memory at a location that can be addressed by said category as a pointer;
• g) incrementing large block indicia i;
• h) repeating said steps b) to g) until all large blocks of the input image have been processed;
• and
• i) wherein upon receipt of a selected small block pattern of the input image, a closest prototype stored by the artificial neural network is identified and the associated category is extracted to act as a pointer to the stored intermediate pattern, wherein generating of an up-converted image is facilitated by applying the retrieved intermediate pattern to the selected small block pattern.

A method for processing an input image including n blocks of p pixels to generate an up-converted image including n large blocks of m pixels (p<m), the method comprising:

• a) providing a neural system comprising an artificial neural network composed of neurons capable of storing a plurality of small block patterns or data related thereto as a prototype with a category associated with the prototype, the plurality of small block patterns being generated from a down-converted image derived from the input image, and a memory configured to store coefficients subsequently allowing up-converting of a selected small block pattern of the input image in a recognition phase, these coefficients being the components of an intermediate pattern, each intermediate pattern being associated with a corresponding small block pattern;
• b) initializing a block indicia i;
• c) deriving a selected small block pattern from block i having p components of the input image;
• d) applying said selected small block pattern to a dedicated processor that transmits said small block pattern with or without processing to said artificial neural network for extracting the category of the closest prototype stored therein, which in turn acts as a pointer to output said associated intermediate pattern from the memory;
• e) applying said selected small block pattern and the intermediate pattern associated with the closest prototype to said processor for facilitating generating of an up-converted image using said coefficients of the intermediate pattern;
• f) incrementing block indicia i;
• and
• g) repeating steps b) to f) until all blocks of the input image have been processed.

A method for scaling an input image consisting of n blocks of p pixels including both a learning operation and a recognition operation, said learning operation comprising:

• a) initializing a block indicia i;
• b) deriving a large block pattern from block i having p pixels;
• c) down-converting said block i of p pixels into a smaller block of q pixels (q<p), to produce a small block pattern having q components;
• d) applying said small block and large block patterns to a dedicated processor for computing a set of coefficients allowing the determination of estimated up-converted values of a selected small block pattern during the recognition operation, these coefficients being the components of an intermediate pattern;
• e) storing the small block pattern in an artificial neural network composed of neurons capable of storing said small block pattern as a prototype with a category associated thereto;
• f) storing said intermediate pattern in a memory at a location that can be addressed by said category as a pointer;
• g) incrementing block indicia i;
• and
• h) repeating steps b to g) until all large blocks of the input image have been processed;
• and said recognition operation comprising:
• i) initializing a small block indicia j;
• j) deriving a selected small block pattern from small block j having p pixels of the input image;
• k) applying said selected small block pattern of the input image to the dedicated processor that transmits said selected small block pattern with or without processing to said artificial neural network for extracting the category of the closest prototype stored therein, which in turn acts as a pointer for outputing said intermediate pattern from the memory;
• l) applying said selected small block pattern and said intermediate pattern retrieved from memory to said processor for facilitating generating of an up-converted image from the input image;
• m) incrementing small block indicia j;
• and
• n) repeating steps j) to m) until all small blocks of the input image have been processed.