Pattern classification and filter design for increasing image resolution

1. A method for determining filters for improving image resolution, comprising:

(a) generating a low-resolution sample image from a high-resolution sample image;

(b) generating a vector that characterizes a low-resolution patch about a low-resolution pixel in the low-resolution sample image;

(c) classifying the low-resolution patch into one of a plurality of possible patterns by comparing the vector and the patterns;

(d) repeating steps (a) to (c) for additional low-resolution pixels to classify additional low-resolution patches into the possible patterns;

(e) repeating steps (a) to (d) for additional high-resolution sample images;

(f) ranking the possible patterns by number of low-resolution patches classified into the possible patterns;

(g) determining filters for a plurality of high ranking patterns from the possible patterns using the low-resolution patches classified into the high ranking patterns and corresponding high-resolution patches in corresponding high-resolution sample images.

2. The method of claim 1 , wherein said generating a low-resolution sample image from a high-resolution sample image comprises down-sampling the high-resolution sample image.

3. The method of claim 1 , wherein said generating a vector comprises:

I={I 0 , I 2 , . . . , I 7 }

I k =L k+1 −L 00 , where k= 0, 1, . . . , 7

where I is the vector, L 00 is the low-resolution pixel, and L k+1 is a neighboring pixel around the low-resolution pixel in the low-resolution patch.

4. The method of claim 1 , further comprising, between steps (b) and (c), skipping step (c) for the low-resolution patch if a length of the vector is less than a threshold.

5. The method of claim 1 , wherein said classifying the low-resolution patch into one of a plurality of possible patterns comprises determining distances between the vector and the possible patterns both projected onto a unit sphere, the low-resolution patch belonging to a pattern having the smallest distance.

6. The method of claim 1 , wherein the possible patterns comprises:

C i =[P 0 , P 1 , . . . , P 7 ]

P j ( j =0, 1, . . . , 7)∈(0, 1,−1)

where vector C i is a possible pattern and P j is a component of vector C.

7. The method of claim 1 , wherein the high ranking patterns comprises the top 274 patterns from the possible patterns.

8. The method of claim 1 , wherein said determining filters for a plurality of high ranking patterns comprises applying a least-mean-square estimation on the low-resolution patches classified into the high ranking patterns and the corresponding high-resolution patches.

9. The method of claim 1 , wherein the high ranking patterns, including their variations, comprise:

Class

Pattern

Number of variations

Class 0:

+++00000

(16)

Class 1:

+++0−−−0

 (8)

Class 2:

+++0+++0

 (8)

Class 3:

+0000000

(16)

Class 4:

++000000

(16)

Class 5:

++++++++

 (2)

Class 6:

+++++000

(16)

Class 7:

++++0000

(16)

Class 8:

+++++++0

(16)

Class 9:

+++00−−0

(32)

Class 10:

++++−−−−

 (8)

Class 11:

+++000+0

(32)

Class 12:

+++00++0

(32)

Class 13:

++++0−−0

(16)

Class 14:

+000+000

 (8)

Class 15:

+++000−0

 (32).

10. The method of claim 1 , wherein the high ranking patterns and their patterns are divided into a plurality of quadrants of the unit sphere.

11. The method of claim 1 , further comprising:

(h) selecting another low-resolution pixel in a low-resolution target image;

(i) generating another vector that characterizes another low-resolution patch about the another low-resolution pixel;

(j) classifying the another low-resolution patch into another one of the plurality of possible patterns;

(k) if said another one pattern is one of the high ranking patterns:

retrieving a filter for said another one pattern;

applying the filter to the low-resolution target patch to generate high-resolution pixels that correspond to the another low-resolution pixel;

(l) if said one pattern is not one of the high ranking patterns:

applying an interpolation process to the low-resolution target patch to generate the high-resolution pixels;

(m) repeating steps (h) to (l) for other low-resolution pixels to generate other high-resolution pixels to form a desired high-resolution image.

12. A method for improving image resolution, comprising:

(a) selecting a low-resolution pixel in a low-resolution image;

(b) generating a vector that characterizes a low-resolution patch about the low-resolution pixel;

(c) classifying the low-resolution patch into one of a plurality of possible patterns;

(d) if said one pattern is one of a plurality of high ranking patterns having filters:

retrieving a filter for said one pattern;

applying the filter to the low-resolution patch to generate high-resolution pixels that correspond to low-resolution pixel;

(e) if said one pattern is not one of the high ranking patterns:

applying an interpolation process to the low-resolution patch to generate the high-resolution pixels;

(f) repeating steps (a) to (e) for additional low-resolution pixels to generate additional high-resolution pixels to form a high-resolution image.

13. The method of claim 12 , wherein said generating a vector comprises:

I={I 0 , I 2 , . . . , I 7 }

I k =l k+1 −l 00 , where k= 0, 1, . . . , 7

where I is the vector, l 00 is the low-resolution pixel, and l k+1 is a neighboring pixel around the low-resolution pixel in the low-resolution patch.

14. The method of claim 12 , wherein said classifying the low-resolution patch into one of a plurality of possible patterns comprises:

classifying the low-resolution patch into one of a plurality of quadrants of a unit sphere;

classifying the low-resolution patch into one of a plurality of patterns in said one quadrant.

15. The method of claim 14 , wherein said classifying the low-resolution patch into one of a plurality of quadrant comprises using the sign of components of the vector to classify the low-resolution patch.

16. The method of claim 14 , wherein said classifying the low-resolution patch into one of a plurality of patterns in said quadrant comprises determining distances between the vector and the patterns both projected onto a unit sphere, the low-resolution patch belonging to a pattern having the smallest distance.