Method and filtering system for filtering edge directions
A filtering method and filtering system is applied to an edge orientation map obtained from an edge direction detection system in order to keep accurate edge directions and filter out false edges or edges with wrong directions. If an edge direction does not have a certain minimum length, then that direction is filtered out and a default direction is provided. Additional assurances can be obtained by insuring that the edge direction has a certain minimum width. If an edge direction does not have the minimum width, then that edge direction is filtered out and the default direction is provided. A direction smoother can be applied to the directions in the edge orientation map to smooth the changing of neighboring edge directions. This process is found to be effective in improving the visual quality of an image that is interpolated based on edge directions.
Latest Samsung Electronics Patents:
Field of the Invention
The present invention relates to a method and a filtering system for filtering edge directions in an edge orientation map detected from an interlaced or non-interlaced scan. The filtered edge directions can then be used in an image interpolation process to preserve the smoothness of the edges in the interpolated image.
Referring to
Image interpolation is usually performed in both image de-interlacing and image scaling applications. There are various kinds of methods used for image interpolation. With most methods, one common problem is the degradation of edges such as the effects of serrate lines associated with the interpolated image. To overcome this problem, it is desirable to find edge directions in the image and to do image interpolation along those directions so that the smoothness of the edges can be preserved.
There have been a large number of methods published on edge detection, including edge location detection and edge direction detection. However, most edge detection methods are sensitive to image noise. In addition, the neighborhood of a given pixel in an image can have numerous cases for an edge detection method to process. As a result, it is common, if not inevitable, for an edge detection system to provide a certain percentage of false edge locations or wrong edge directions.
Meanwhile, the accuracy of the edge directions serves a crucial role when the directions are used in image applications such as de-interlacing or scaling. In edge direction based image de-interlacing, the edge direction may be detected at the position of each pixel to be interpolated. Then for edge pixels, interpolation can be performed along the edge directions to preserve the smoothness of the edges in the interpolated image. The same idea can be applied to image scaling as well. However, in these applications if a detected edge is false or if the edge direction is not correct, obvious error may be introduced into the interpolated image. Therefore, when edge directions are utilized in these kinds of applications, false edges or edges with wrong directions must be excluded before the edge information is used.
SUMMARY OF THE INVENTIONIt is accordingly an object of the invention to provide a method and a filtering system for filtering edge directions, which overcome the above-mentioned disadvantages of the prior art methods and apparatus of this general type.
In particular, it is an object of the invention to provide a filtering method that can be applied to an edge orientation map obtained from an edge direction detection system. The filtering method should keep accurate edge directions and filter out false edges or edges with wrong directions.
To accomplish the object of the invention, a direction length filter is provided for performing a direction length filtering process. The direction length filter is used to filter the edge direction along the direction indicated by the edge direction. Along the current direction, neighboring edge directions are checked to see whether their orientations are similar to the current direction being checked. If the neighboring directions have very different orientations from the current direction, the current direction is filtered out. The assumption for this filtering is that a valid edge direction should be able to be extended and should therefore have a certain minimum length.
A direction width filter for performing a direction width filtering process may also be provided. The width of a direction in an edge orientation map is defined as the number of consecutive neighboring positions along the horizontal direction where the edge directions are close to the current one. The direction width filter is used to filter out edge directions in the edge orientation map that are too thin. If along the horizontal direction there are a certain minimum number of consecutive neighboring edge directions that have similar orientations to the current one, then the current direction is kept. Otherwise, the current direction is filtered out.
A direction smoother for performing a direction smoothing process may-also be provided. The direction smoother is basically a low pass filter. When the direction smoother is applied to directions in an edge orientation map, the filter can smooth the changing of neighboring edge directions. This process is found to be effective in improving the visual quality of an image that is interpolated based on edge directions.
With the foregoing and other objects in view there is provided, in accordance with the invention, a method for filtering an edge direction. The method includes steps of: in a given row in an edge orientation map, selecting a given pixel having an edge direction; finding neighboring upper pixels having neighboring upper edge directions for the given pixel by traversing along the edge direction of the given pixel from the given pixel to a neighboring upper row of the edge orientation map; finding neighboring lower pixels having neighboring lower edge directions for the given pixel by traversing along the edge direction of the given pixel from the given pixel to a neighboring lower row of the edge orientation map; determining whether the edge direction of the given pixel can be extended to the neighboring upper row by checking an orientation similarity between the edge direction of the given pixel and the neighboring upper edge directions; determining whether the edge direction of the given pixel can be extended to the neighboring lower row by checking an orientation similarity between the edge direction of the given pixel and the neighboring lower edge directions; if the edge direction of the given pixel can be extended to both the neighboring upper row and the neighboring lower row, then assuming that the edge direction of the given pixel is valid; if the edge direction of the given pixel cannot be extended to the neighboring lower row, but the edge direction of the given pixel can be extended to one of the neighboring upper pixels, and a neighboring upper edge direction at the one of the neighboring upper pixels can be extended to a further neighboring upper row of the edge orientation map, then assuming that the edge direction of the given pixel is valid; if the edge direction of the given pixel cannot be extended to the neighboring upper row, but the edge direction of the given pixel can be extended to one of the neighboring lower pixels, and a neighboring lower edge direction at the one of the neighboring lower pixels can be extended to a further neighboring lower row of the edge orientation map, then assuming that the edge direction of the given pixel is valid; and otherwise, assuming that the edge direction of the given pixel is not valid.
In accordance with an added feature of the invention, the edge direction of the given pixel (x,y) is represented by d(x,y), and the step of finding the neighboring upper pixels includes using equations:
f(x,y)=d(x−2,cu(x,y))
fl(x,y)=d(x−2,cu(x,y)−1), and
fr(x,y)=d(x−2,cu(x,y)+1), where cu(x,y)=y+d(x,y)*2.
In addition, the step of finding the neighboring lower pixels includes using equations:
g(x,y)=d(x+2,cd(x,y))
gl(x,y)=d(x+2,cd(x,y)−1), and
gr(x,y)=d(x+2,cu(x,y)+1), where cu(x,y)=y−d(x,y)*2.
In accordance with an additional feature of the invention, the step of checking the orientation similarity between the edge direction of the given pixel and the neighboring upper edge directions includes:
-
- a) defining a checking function chkup(x,y) as follows:
- b) concluding that the edge direction of the given pixel can be extended to the neighboring upper row if chkup(x,y) equals one; and
the step of checking the orientation similarity between the edge direction of the given pixel and the neighboring lower edge directions includes: - a) defining a checking function chkdn(x,y) as follows:
- b) concluding that the edge direction of the given pixel can be extended to the neighboring lower row if chkdn(x,y) equals one.
Here, T1 is a threshold value chosen to be greater than zero and small enough to verify correlations of neighboring edge directions, but not small enough such that slightly different directions will remain.
- a) defining a checking function chkup(x,y) as follows:
In accordance with another feature of the invention, the method includes:
- defining a checking function chkup2(x,y) as:
- defining a checking function chkdn2(x,y) as:
if the checking function chkup2(x,y) takes a value of one, concluding that the neighboring upper edge direction at the one of the neighboring upper pixels can be extended to the further neighboring upper row of the edge orientation map; and if the checking function chkdn2(x,y) takes a value of one, concluding that the neighboring lower edge direction at the one of the neighboring lower pixels can be extended to the further neighboring lower row of the edge orientation map.
In accordance with a further feature of the invention, if the edge direction of the given pixel is within a threshold value from two horizontally neighboring directions in the given row, then a final conclusion is made that the edge direction of the given pixel is valid.
In accordance with a further added feature of the invention, the method includes: checking a width of the edge direction of the given pixel by determining an orientation similarity between the edge direction of the given pixel and two horizontally neighboring directions in the given row.
In accordance with a further additional feature of the invention, the edge direction of the given pixel (x,y) is represented by d(x,y); and the step of checking the width of the edge direction of the given pixel includes using equations:
max(|d(x,y)−d(x,y−1)|,|d(x,y+1)−d(x,y)|)≦T2,
max(|d(x,y)−d(x,y−1)|,|d(x,y−1)−d(x,y−2)|)≦T2, and
max(|d(x,y+1)−d(x,y)|,|d(x,y+2)−d(x,y+1)|≦T2.
Here, T2 is a threshold value chosen to be greater than zero and small enough to verify correlations of neighboring edge directions, but not small enough such that slightly different directions will remain.
In accordance with yet an added feature of the invention, the method includes: performing a filtering process to smooth a transition between neighboring edge directions.
In accordance with yet an additional feature of the invention, the method includes: using a three-tap filter for the filtering process; and implementing the three-tap filter using equation: dS(x,y)=(d(x,y−1)+2*d(x,y)+d(x,y+1))/4. Here, dS(x,y) is a filtered output direction.
With the foregoing and other objects in view there is also provided, in accordance with the invention, an edge direction filtering system including: a direction length filter for extending a current edge direction to neighboring upper rows and to neighboring lower rows in an edge orientation map. The direction length filter is for checking an orientation similarity between the current edge direction and neighboring edge directions located along the current edge direction. The direction length filter indicates that the current edge direction is valid if the current edge direction can be extended to a certain number of neighboring rows in the edge orientation map.
In accordance with an added feature of the invention, a direction width filter is provided for filtering the current edge direction along a horizontal direction. The direction width filter is for checking an orientation similarity between the current edge direction and neighboring edge directions on a row with the current edge direction in an edge orientation map. The direction width filter indicates that the current edge direction is valid if a certain number of consecutive ones of the neighboring edge directions on the row with the current edge direction have a similar orientation to the current edge direction.
In accordance with a concomitant feature of the invention, a low pass filter is provided for smoothing transitions between horizontally neighboring edge directions of the edge orientation map.
BRIEF DESCRIPTION OF THE DRAWINGS
In this specification, {d(.,.)} denotes an edge direction detected in an image produced by an interlaced or non-interlaced scan operation. In an image de-interlacing application, the edge directions are detected at the position of each pixel that will be interpolated. This case is shown in
A similar method can be applied to image scaling applications as well. For instance, edge directions can be detected between neighboring rows as shown at each location marked with an “⋄” in
For the kind of applications mentioned above, the accuracy of the edge directions is crucial to the overall system performance. Since image interpolation is done along the edge direction at each location, obvious error may be introduced if the edge direction is wrong or inaccurate. Therefore, edge direction filtering is important so that wrong or inaccurate directions can be removed from the edge orientation map before the map is used for interpolation.
For simplicity, the following description of the invention is based on an edge orientation map generated from an image de-interlacing application. But the described procedures can be applied to an edge orientation map generated from an image scaling application as well.
The edge orientation map is represented by assigning a different value to each direction with a different orientation. Neighboring directions have neighboring values. The direction numbering method is shown in
Once an incorrect direction is filtered out, no edge direction information should be used for interpolation at that location. In such a case, interpolation may be performed along the vertical direction. Therefore, the vertical direction can be considered as a default interpolation direction.
As shown in
In this invention, the length requirement for a valid edge is defined as: being able to extend the edge along at least three rows in the edge orientation map. To further clarify the length requirement, some functions are defined as follows:
f(x,y)=d(x−2,cu(x,y)) (1)
g(x,y)=d(x+2,cd(x,y)) (2).
Here, d(x,y) is the edge direction value at the position of pixel (x,y). Functions cu(x,y) and cd(x,y) are defined in equations (3) and (4) as:
cu(x,y)=y+d(x,y)*2 (3)
cd(x,y)=y−d(x,y)*2 (4).
As shown in
Based on equations (1) and (2), the following functions can also be defined:
fl(x,y)=d(x−2,cu(x,y)−1) and fr(x,y)=d(x−2,cu(x,y)+1) (5)
gl(x,y)=d(x+2,cd(x,y)−1) and gr(x,y)=d(x+2,cd(x,y)+1) (6).
Obviously, fl(n,m) is the direction located to the left of f(n,m) and fr(n,m) is the direction located to the right of f(n,m). In
Based on the above function definitions, two direction checking functions can be defined. One of them is used for checking direction consistency on the immediate upper row relative to the current position in the edge orientation map. The other is used for checking direction consistency on the immediate lower row relative to the current position in the edge orientation map. The two checking functions are named as chkup(x,y) and chkdn(x,y) respectively.
In the chkup(x,y) function, when any of the following three conditions shown in (7)˜(9) is true, it's said that there is direction consistency for the current direction in its immediate upper row.
|d(x,y)−f(x,y)|≦T1 (7)
|d(x,y)−fl(x,y)|≦T1 (8)
|d(x,y)−fr(x,y)|≦T1 (9).
The flowchart of this checking process is shown in
T1 is a threshold value that is greater than zero. When T1 takes a small value, it requires that the neighboring edge directions shown in (7), (8) or (9) have very similar orientations to each other. If T1 equals zero, it requires that the neighboring edge directions have exactly the same orientation. The value of T1 should be chosen in a way such that it's small enough to verify correlations of neighboring edge directions, but not too small to allow slightly different directions to remain.
If chkup(x,y) takes a value of one, it's said that the direction d(x,y) can be extended to its neighboring upper row in the edge orientation map. Similarly, in the chkdn(x,y) function, when any of the following three conditions shown in (11)˜(13) is true, it's said that there's direction consistency for the current direction in its immediate lower row.
|d(x,y)−g(x,y)|≦T1 (11)
|d(x,y)−gl(x,y)|≦T1 (12)
|d(x,y)−gr(x,y)|≦T1 (13).
The flowchart of this checking process is shown in
When chkdn(x,y) takes a value of one, it's said that the direction d(x,y) can be extended to its neighboring lower row in the edge orientation map.
According to the length requirement for a valid edge, if both chkup(x,y) andchkdn(x,y) take a value of 1, i.e. chkup(x,y)+chkdn(x,y)=2, then the direction d(x,y) can be extended to both its neighboring upper row and neighboring lower row. Therefore, d(x,y) meets the length requirement. If both chkup(x,y) andchkdn(x,y) take a value of 0, then d(x,y) does not meet the length requirement and it is invalid. In the case of chkup(x,y)+chkdn(x,y)=1, further checking is needed.
Assume that chkup(x,y)=1 and chkdn(x,y)=0. In this case, d(x,y) can be extended to its neighboring upper row, but not to its neighboring lower row in the edge orientation map. As shown in
The checking function used in this case is named as chkup2(x,y). This function takes a value of one if any of the following three cases shown in (15)˜(17) is true:
|d(x,y)−f(x,y)≦T1 and chkup(x−2,cu(x,y))=1 (15)
|d(x,y)−fl(x,y)|≦T1 and chkup(x−2,cu(x,y)−1)=1 (16)
|d(x,y)−fr(x,y)|≦T1 and chkup(x−2,cu(x,y)+1)=1 (17).
The flowchart of this checking process is shown in
When chkup2(x,y) takes a value of one, it means that the direction d(x,y) can be extended to both the row x-2 and x-4. Therefore, d(x,y) also meets the edge length requirement in this case. In
In a similar way, when chkup(x,y)=0 and chkdn(x,y)=1, a checking function chkdn2(x,y) is defined for this case and it takes a value of one if any of the following three cases shown in (19)˜(21) is true:
|d(x,y)−g(x,y)≦T1 and chkdn(x+2,cd(x,y))=1 (19)
|d(x,y)−gl(x,y)≦T1 and chkdn(x+2,cd(x,y)−1)=1 (20)
|d(x,y)−gr(x,y)≦T1 and chkdn(x+2,cd(x,y)+1)=1 (21).
The flowchart of this checking process is shown in
Based on the descriptions above, the direction length filtering conditions can be summarized as follows:
chkup(x,y)+chkdn(x,y)=2 (23)
chkup2(x,y)=1 (24)
chkdn2(x,y)=1 (25).
When any condition in (23)˜(25) is true, the direction d(x,y) is said to meet the direction length requirement. Assume that the direction length filtering result on {d(.,.)} is {dL(.,.)}, the flow chart of the direction length filtering process is, shown in
If a direction d(n,m) is valid according to the checking process shown in
It's found that the Direction Length Filter 1 is highly efficient in removing false edge directions by exploiting the correlation among neighboring directions along edge.
The width requirement for the position of pixel (n,m) can be expressed as the following three cases:
max(|d(n,m)−d(n,m−1)|,|d(n,m+1)−d(n,m)|)≦T2 (26)
max(|d(n,m)−d(n,m−1)|,|d(n,m−1)−d(n,m−2)|)≦T2 (27)
max(|d(n,m+1)−d(n,m)|,|d(n,m+2)−d(n,m+1)|≦T2 (28).
T2 is a threshold value greater than zero. When T2 takes a small value, it requires that neighboring edge directions shown in (26), (27) or (28) should have very similar orientations to each other. If T2 equals to zero, it requires that neighboring edge directions have exactly the same orientation. The value of T2 should be chosen in a way that it's small enough to verify correlations of neighboring edge directions, but not too small to allow slightly different directions to remain.
As shown in
For the position of pixel (n,m), if the condition in any one of (26), (27) or (28) is true, then d(n,m) is said to meet the direction width requirement. Assume that the direction width filtering result on {d(.,.)} is {dW(.,.)}. The flow chart for the direction width filtering process that insures that the edge direction has the required width is shown in
If a direction d(n,m) is valid according to the width filtering process shown in
Similar to Direction Length Filter 1, the Direction Width Filter 2 also utilizes the correlation among neighboring directions, but it does this along the horizontal direction instead of the edge direction. When the Direction Width Filter 2 is combined with the Direction Length Filter 1 in a system, it further improves the system's capability to filter out false edge directions in an edge orientation map.
The Direction Smoother 3 is basically a 3-tap low pass filter. Assume the direction smoothing result on {d(.,.)} is {dS(.,.)}. The Direction Smoother 3 can be expressed as the following equation (11):
dS(n,m)=(d(n,m−1)+2*d(n,m)+d(n,m+1))/4 (29).
This process is found to be effective in improving the visual quality of the interpolated image.
Claims
1-12. (Canceled).
13. An edge direction filtering system, comprising:
- a direction filter that extends a current edge direction to neighboring rows in an edge orientation map;
- said direction filter further determines an orientation similarity between the current edge direction and neighboring edge directions located along the current edge direction;
- wherein the direction filter indicates if the current edge direction is valid.
14. The system of claim 13 wherein the direction filter determines if the current edge direction is valid by determining if the current edge direction can be extended to a certain number of neighboring rows in the edge orientation map.
15. The system of claim 13 wherein the direction filer extends the current edge direction to neighboring upper and lower rows in the edge direction map.
16. The system of claim 15 wherein the direction filter determines if the current edge direction is valid by determining if the current edge direction can be extended to a certain number of neighboring rows in the edge orientation map.
17. The system of claim 13 wherein the direction filter further extends the current edge direction in the same row as the current edge direction in the edge direction map.
18. The system of claim 17 wherein the direction filter determines if the current edge direction is valid by determining if the current edge direction can be extended to a certain number of neighboring rows in the edge orientation map.
19. The system of claim 13 wherein the direction filter extends the current edge direction in the same row as the current edge direction in the edge direction map, wherein the direction filter checks an orientation similarity between the current edge direction and neighboring edge directions on a row with the current edge direction in the edge orientation map, whereby the direction filter indicates that the current edge direction is valid if a certain number of consecutive ones of the neighboring edge directions on the row with the current edge direction have a similar orientation to the current edge direction.
20. The system of claim 13 further comprising a low pass filter for smoothing transitions between neighboring edge directions of the edge orientation map.
21. A method for filtering edge directions, comprising the steps of:
- extending a current edge direction to neighboring rows in an edge orientation map;
- determining an orientation similarity between the current edge direction and neighboring edge directions located along the current edge direction; and
- indicating if the current edge direction is valid.
22. The method of claim 21 wherein the step of extending the current edge direction further includes the steps of determining if the current edge direction is valid by determining if the current edge direction can be extended to a certain number of neighboring rows in the edge orientation map.
23. The method of claim 21 further including the steps of extending the current edge direction to neighboring upper and lower rows in the edge direction map.
24. The method of claim 23 further including the steps of determining if the current edge direction is valid by determining if the current edge direction can be extended to a certain number of neighboring rows in the edge orientation map.
25. The method of claim 21 further including the steps of extending the current edge direction in the same row as the current edge direction in the edge direction map.
26. The method of claim 25 further including the steps of determining if the current edge direction is valid by determining if the current edge direction can be extended to a certain number of neighboring rows in the edge orientation map.
27. The method of claim 21 further including the steps of extending the current edge direction in the same row as the current edge direction in the edge direction map, and checking an orientation similarity between the current edge direction and neighboring edge directions on a row with the current edge direction in the edge orientation map, to indicate that the current edge direction is valid if a certain number of consecutive ones of the neighboring edge directions on the row with the current edge direction have a similar orientation to the current edge direction.
28. The method of claim 21 further including the steps of applying a low pass filter for smoothing transitions between neighboring edge directions of the edge orientation map.
Type: Application
Filed: Aug 5, 2004
Publication Date: Jan 13, 2005
Applicant: Samsung Electronics Co., Ltd. (Suwon City)
Inventor: Xianglin Wang (Santa Ana, CA)
Application Number: 10/913,663