TWI501652B - An image process method having a function of determining the suitable properties of the directional filter - Google Patents

Description

Image processing method with function for judging directional filter applicability

The present invention relates to an image processing method, and more particularly to an image processing method having the function of determining the applicability of a directional filter, by using a mean shift algorithm to determine a wavelet of a frequency domain in a frequency domain. Whether the wavelet transform subband is disassembled by a directional filter can reduce the amount of computation required for encoding.

In today's society where the amount of information on the Internet is booming, there is a growing demand for the processing of data volumes, especially for video and video information. The amount of data is even more alarming. On the server side, in order to efficiently store these multimedia materials and transmit them in the Internet in a more efficient manner, in order to be able to use a variety of applications on the network, it is necessary to compress the video image data more efficiently. , can meet the needs of the diversity of client users.

Wavelet transform is adopted by JPEG2000 specification and used in image compression and circumstance because of its good nonlinear approximation. A typical two-dimensional (2-dimensional, 2-D) wavelet transform divides the image into two one-dimensional (2-dimensional, 1-D), which is to divide the image into vertical and horizontal, so it ignores two. The continuity of the news signal. Since the lines in natural images are not all vertical or horizontal, a typical two-dimensional wavelet transform produces many wavelet coefficients with very small absolute values when processing these non-horizontal or vertical image lines. The coefficient has a negative effect on compression.

In order to solve the problem of poor directionality of wavelet transform, a contour transform image processing method has also been proposed, that is, the first level Laplacian pyramid is used to make the image high and low frequency bands. Dismantle. In the disassembled high and low frequency band signals, the high-pass sub-band uses a directional filter bank for disassembly, and the low-pass sub-band can be sent to the second layer ( The second level of the pull pyramid continues to disassemble the second layer of high and low subbands. In the disassembled high and low frequency band signals, the high pass subband can be disassembled by the directional filter, and the low pass subband can be sent to the third level of the pull pyramid as the third layer. Disassembly of high and low sub-bands. This step is repeated until the desired accuracy. Although the contour conversion method has better representation characteristics in the direction, the pull pyramid used causes an increase in the amount of data after conversion, and thus is disadvantageous to the compression processing.

In order to maintain the data volume architecture before and after the conversion, a method combining the advantages of the wavelet transform method and the contour transform method is proposed, that is, a wavelet-based contourlet transform method. Figure 2 shows the architecture of contour transformation based on wavelet transform, where the directional filter is represented in four directions. The main architecture is to use directional filters on the high-pass sub-band after wavelet transform to achieve more accurate directional disassembly. Figure 2 specifically shows the results of the first layer of wavelet transform. The three high-pass sub-bands (defined as low-high (LH), high-low (HL) and high-height (HH)) are disassembled via directional filters in order to achieve More accurate directional analysis. A low pass sub-band (LL) can further disassemble the second layer of wavelet transform. Similar to the wavelet processing process of the first layer, the three high-pass sub-bands after the wavelet transform of the second layer can also be disassembled for the directional filter, and a low-pass sub-band can also be given to the third. The layer's wavelet transform continues to be disassembled, so it is continuously disassembled until the set resolution is obtained.

However, not all sub-bands after wavelet transform are suitable for directional disassembly. Most of the sub-bands do not help the compression result or become worse after using the directional filter. The context table of the bit-plane coding used by JPEG2000 is also based on wavelet transform, and does not consider other directionality. In addition, the bit plane coding used in the wavelet transform method is EBCOT and 3-D EBCOT specifications. The former is mainly used for video, the latter is mainly used for video, but their adjacent tables are used to process signals in the bit plane. Only considering the horizontal and vertical directions, it is not suitable for the signal after the directional disassembly, so the contour conversion based on wavelet transform still can not achieve better compression effect.

In view of the shortcomings of the prior art, it is necessary to propose an adjustable image processing method, which can reduce the amount of computation required for encoding by determining whether the wavelet transform sub-band in the frequency domain is disassembled by the directional filter.

In view of the above disadvantages of the prior art, the main object of the present invention is to provide an image processing method having the function of determining the directional filter applicability, which can determine whether the sub-band of the wavelet transform is suitable for directional disassembly, and then according to all directions. Signal characteristics on the sub-band, propose a new adjacent table to achieve better compression.

To achieve the above objective, the present invention provides an image processing method for determining the applicability of a directional filter for compressing and encoding an image signal. The method includes at least the following steps: Step (1): making the image signal Two-dimensional Fourier transform to obtain an energy spectrum of the image signal in the frequency domain; step (2): correcting the obtained energy spectrum, eliminating the small peak of the energy spectrum; and step (3): correcting the energy Spectrum, find the threshold value for judging the peak of each sub-band in the energy spectrum of the frequency domain; Step (4): Calculating by using the moving average method by comparing the threshold value and the peak of each sub-band peak in the energy spectrum a convergence position of a centroid of each sub-band; and step (5): determining whether each sub-band in the energy spectrum needs to be used by finding a final convergence position of the centroid of each sub-band The directional filter is disassembled.

In summary, the image processing method of the present invention having the function of determining the directional filter applicability will have the following effects:

1. By determining whether the wavelet transform sub-band corresponding to the frequency domain is suitable for disassembly by using a directional filter, and combining the proposed adjacent table for bit-plane coding, the overall compression effect is improved.

2. It can save a lot of calculations by avoiding disassembling the sub-band which does not apply to the directional filter, further saving the cost of the device and the time of compression.

The above and other objects, features, and advantages of the present invention will become more apparent and understood.

While the invention may be embodied in various forms, the embodiments illustrated in the drawings It is not intended to limit the invention to the particular embodiments illustrated and/or described.

Referring now to FIG. 1, there is shown a flow chart of the steps of the image processing method of the present invention having the function of determining the directional filter applicability, which mainly comprises five steps.

First, in step S1, the image signal is subjected to a 2-D fourier transform to the frequency domain, and a plurality of Fourier transform coefficients are obtained, and each coefficient is taken as an absolute value and squared to obtain a representative. The energy spectrum of an image signal, each coefficient of the energy spectrum is called the energy coefficient g(x, y), where (x, y) represents the coordinate value of the energy coefficient, and g is the energy coefficient. Value. Referring to FIG. 3, it is a distribution of sub-bands in each direction in the frequency domain after the image is subjected to contour conversion based on wavelet transform. Assuming that the image signal is a × b two-dimensional Fourier transform, an energy spectrum of a × b size can be obtained, as shown in Fig. 3.

In step S2, the energy spectrum obtained in step S1 is corrected, that is, the energy spectrum is corrected to be an exponential energy spectrum. The modified action includes: S2-1: First, the zero frequency energy coefficient is placed in the middle of the entire energy spectrum, so that the entire spectrum is point-symmetric; S2-2: using a smoothing operator Make the peak of the energy spectrum more obvious and eliminate most of the small peaks; S2-3: take the energy coefficient g(x, y) as an exponential value to get an exponential energy spectrum h(x, y)=log ₁₀ (g(x, y)), the corrected energy spectrum of the exponential type is still a × b.

In step S3, a threshold value for determining the peak of each sub-band is found. Generally, natural images have very powerful low frequency signals, as shown in the gray part of Figure 3. The steps to find the threshold value for judging the peak further include:

S3-1: First, the average value of the low frequency signal in the high-pass sub-bands LH and HL in the exponential energy spectrum, that is, the average value of h(x, y) in the gray region in LH and HL is calculated. The gray area of LH is c×b/4, and the c value is determined according to the characteristics of the image and the number of Fourier transform points used; the gray area of HL is d×a/4, and d is similar to c, that is, It is determined by the characteristics of the image and the number of Fourier transform points used. Among them, to determine the interval of the low-frequency signal, that is, to determine c and d, the energy coefficient g(x, y) is the g(x, y) of other regions before the index value is taken in this interval. W ₁ times or more, you can determine c and d, and then get the average of the exponential energy spectrum h(x, y) of the gray regions in LH and HL, which are set to LH_low and HL_low, respectively. In a preferred embodiment, the parameter W _{1 for} determining the low frequency signal and other signals is preferably between 0 and 10.

S3-2: Find the limit value of the peak of the high-pass sub-band by judging the average value in the high-pass sub-bands LH and HL. The detailed implementation method is as follows: if one of LH_low and HL_low is ₂ times larger than the other W, the small average value is replaced by a large average value, and if the large average value is not ₂ times larger than the small one, Then both averages remain intact. Next, the mean and standard deviation of the energy spectrum h(x, y) of the exponential energy spectrum in each sub-band are calculated. For example, in the LH_4-0 region, the average value of the energy coefficient h(x, y) of the exponential energy spectrum is LH_4-0_μ, and the standard deviation is LH_4-0_σ, and LH_4-0_μ+W ₃ × LH_4-0_σ is compared with LH_low, and the larger one is the limit value LH_4-0_T of the sub-band LH_4-0 in the direction; similarly, in the HL_4-0 region, the energy coefficient h(x, y) of the exponential energy spectrum The mean (mean) HL_4-0_μ and the standard deviation (variance) HL_4-0_σ are compared with HL_4-0_μ+W ₃ ×HL_4-0_σ and HL_low, and the larger one is the boundary of the sub-band HL_4-0 in the direction. The value HL_4-0_T; the threshold values of the other direction sub-bands are the same algorithm. In a preferred embodiment, the parameter W _{2 for} determining which low frequency signal is reserved is preferably set between 0 and 6, and finally the parameter W _{3 of} each sub-band is determined, preferably set to 0. Between 1 and 1.

As can be seen from the above, in step S3, the number of direction splits to which it applies is two or more sub-bands.

In step S4, by comparing the threshold value and the peak of each sub-band peak in the energy spectrum, based on the moving average method, the convergence position of finding a centroid coordinate of each sub-band is calculated. The step further includes: S4-1: in each direction sub-band, such as LH_4-0, we retain the energy coefficient h(x, y) larger than the corresponding limit value LH_4-0_T and take the exponential value, Converting the energy coefficient h(x, y) back to the energy coefficient g(x, y) before the index value is not taken; S4-2: taking g(x, y) as the center of a first search window, The search window size is (2×γ+1)×(2×γ+1), and the first centroid coordinate position (x _mass , y _mass ) of the search window is calculated by the formula (1); wherein, the parameter γ is compared The best is set between 1 and 10; S4-3: repeated (1) until the centroid coordinates no longer change. That is, the first centroid coordinate is regarded as the coordinate of the second search window, and the centroid coordinate of the second search window is further calculated by the formula (1), and the repeated (1) until the centroid coordinates are no longer change.

It should be noted that the range we use to find the sub-band LH_4-0 is the position of the wavelet transform sub-band LH in the frequency domain, and then the LH is extended to the γ energy coefficient g(x, y). Pay attention to the periodicity of the frequency domain signal when extending. If there are b columns in the frequency domain as shown in Figure 3 and each column has a energy coefficient g(x, y), then the b+1th column The value of the value is the value of the first column, and because of the point symmetry in the frequency domain, the direction sub-band LH 4-0 has the upper and lower parts in the third picture, we only need to analyze the upper part and can, The other direction sub-bands are also analyzed to find the convergence position of the final centroid coordinates.

In step S5, a position where the centroid coordinate finally converges is found, and it is determined whether each sub-band needs to be disassembled by using a directional filter, and if disassembly is required, the signal of the sub-band after the disassembly can be used for the fourth time. The adjacent tables of the graph are encoded and compressed. This step further includes:

S5-1: Use a size table of the same size as step S1, which is also a × b. In step S4, as long as the position at which the centroid coordinates finally converge is found, 1 is added to the corresponding position on the score table until the energy coefficient h above the threshold value in all sub-bands in step S4. , y) are all processed.

S5-2: If within a range of τ × τ on the score table, the center is the largest value in the range and its value λ, then put all in the range The values of λ are added together, and the sum is placed at the center of τ × τ and the other values are set to 0. We call this a convergence point. In a preferred embodiment, the size τ to be processed on the score table is preferably between 3 and 5, and the threshold λ is preferably set to 5.

S5-3: If in a sub-band of a wavelet transform, there is a convergence point and its value λ, and the convergence point is not within the convergence interval of the low frequency signal, then the subband can be disassembled using a directional filter.

S5-4: The sub-band after disassembly can be encoded and compressed by the adjacent table of FIG.

In an embodiment, for example, the convergence region of the low frequency signal of the LH sub-band is a gray-scale oblique line region in the LH sub-band of FIG. 3, the size of which is an interval of e×b/4, and the low-frequency signal of the HL sub-band The convergence interval is the gray-scale oblique line region in the HL sub-band of Fig. 3, and the size is in the interval of f × a / 4. If the convergence point obtained in step 4 is not within the interval, the frequency band can be disassembled using the directional filter. After the disassembly, the adjacent table in Fig. 4 can be used for encoding and compression, which can improve the overall compression. effect. The symbols in the adjacent table of Fig. 4 indicate the position of each sub-band. In a preferred embodiment, the gray-scale oblique line region of the LH sub-band shown in FIG. 3 is a convergence region of the low-frequency signal of the LH sub-band, and the size thereof is e×b/4, wherein the size of e is recommended to be at least c, and the gray-scale oblique line region of the HL sub-band shown in FIG. 3 is a convergence region of the low-frequency signal of the HL sub-band, and its size is f×a/4, and the size of f is recommended to be at least d.

In summary, the image processing method of the present invention having the function of determining the directional filter applicability will have the following effects:

1. By determining whether the wavelet transform sub-band corresponding to the frequency domain is suitable for disassembly by using a directional filter, and combining the proposed adjacent table for bit-plane coding, the overall compression effect is improved.

2. It can save a lot of calculations by avoiding disassembling the sub-band that does not apply to the directional filter, further saving the cost of the device and the time of compression.

While the present invention has been described in its preferred embodiments, it is not intended to limit the scope of the invention, and various modifications and changes can be made without departing from the spirit and scope of the invention. As explained above, all types of corrections and changes can be made without destroying the spirit of this creation. Therefore, the scope of the invention is defined by the scope of the appended claims.

Figure 1 is a flow chart showing the steps of the image processing method of the present invention;

Figure 2 is an architecture of contour transformation based on wavelet transform, wherein the directional filter represented is disassembled in four directions;

Figure 3 is a distribution of the sub-bands in each direction in the frequency domain after the image is transformed by the architecture of Figure 2;

Figure 4 is an adjacent table of signals for which the direction of the invention has been disassembled for encoding.

Claims (14)

An image processing method for determining the applicability function of a directional filter for compressing and encoding an image signal, the method comprising at least the following steps: Step (1): performing a two-dimensional Fourier transform on the image signal to obtain The image signal is in an energy spectrum of the frequency domain; step (2): correcting the obtained energy spectrum, eliminating the small peak of the energy spectrum; and step (3): finding the used energy spectrum in the correction to determine the a threshold value of each sub-band peak in the frequency domain energy spectrum; step (4): calculating a centroid of each sub-band by a moving average method by comparing a threshold value and a peak of each sub-band peak in the energy spectrum Convergence position of coordinates; and step (5): determining whether each sub-band in the energy spectrum needs to be disassembled using a directional filter by finding the final convergence position of the centroid coordinate of each sub-band. The image processing method of claim 1, wherein the step (2) further comprises: step (2-1): placing a zero frequency energy coefficient in the middle of the energy spectrum to make the energy spectrum point symmetric Step (2-2): using a smoothing operator to cancel the wavelet of the energy spectrum; and step (2-3): taking the energy coefficient g(x, y) of the energy spectrum as an exponential value to obtain an exponential type The energy spectrum h(x, y) = log ₁₀ (g(x, y)), and the corrected energy spectrum of the exponential type is still a × b. The image processing method of claim 2, wherein the step (3) further comprises: step (3-1): calculating an average value of the low-frequency signal of the exponential energy spectrum in a high-pass sub-band, wherein the high-pass sub-band includes A low high frequency band (LH) and a high and low frequency band (HL), the area of the low high frequency band is c×b/4, and the size of the high and low frequency bands is d×a/4, c value and d value Determining according to the characteristics of the image and the number of Fourier transform points used; Step (3-2): finding the high-pass sub-band by judging the average of the low-high frequency band and the high-low frequency band in the high-pass sub-band The threshold value of the peak. The image processing method of claim 3, wherein in the step (3-1), the method of determining the c value and the d value: before the index value is not taken in the low high frequency band and the high and low frequency band intervals, The energy coefficient g(x, y) is W ₁ times or more of g(x, y) of other regions, and c and d are determined, and the average values of the low-high frequency band and the high-low frequency band obtained are respectively determined as LH_low and HL_low. The image processing method of claim 4, wherein the parameter W ₁ is set to be between 0 and 10. The image processing method of claim 4, wherein in step (3-2), if one of LH_low and HL_low is ₂ times larger than the other, the small average value is replaced by a large average value. If the large average is not ₂ times larger than the small one, the two averages remain. The image processing method of claim 6, wherein the parameter W ₂ is set to be between 0 and 6. For example, in the image processing method of claim 4, wherein the small values of LH_low and HL_low are left untouched, and the average value of the energy coefficient h(x, y) of the exponential energy spectrum in each sub-band is calculated. The standard deviation, and the sum of the mean and the standard deviation multiplied by W ₃ is compared with the average of the exponential energy spectrum, with the larger value being the threshold of the sub-band in that direction. The image processing method of claim 8, wherein the parameter W ₃ is set to be between 0 and 1. For example, in the image processing method of claim 1, wherein the step (4) further comprises: step (4-1): retaining the index larger than the corresponding threshold value and taking the index value in each direction sub-band The energy coefficient h(x, y) of the energy spectrum of the type, the energy coefficient h(x, y) is converted back to the energy coefficient g(x, y) before the index value is not taken; step (4-2): g ( x, y) is the center of a first search window, the size of the search window is (2 × γ + 1) × (2 × γ + 1), wherein the parameter γ is preferably set between 1 and 10; Step (4-3): Repeat step (4-2) until the centroid coordinates no longer change. For example, in the image processing method of claim 1, wherein the step (5) further comprises: step (5-1): using a score table of the same size as a×b, as long as each centroid is finally converged. , adding 1 to the corresponding position on the score table until the energy coefficient h(x, y) above the threshold value is processed in all the sub-bands in step (4); step (5-2): In the range of τ × τ on the score table, the center is the largest value in the range and its value is greater than or equal to λ, then all the values in the range greater than or equal to λ are added up to obtain the sum, and the sum is Place it at the center of τ × τ as a convergence point, and set all other values to 0, where the λ value is set to 5 or more; Step (5-3): If in a sub-band of a wavelet transform, there is Contains convergence points and their values λ, and the convergence point is not located in the convergence interval of the low frequency signal, then the subband can be disassembled using a directional filter; and step (5-4): the subband after disassembly can be encoded by an adjacent table and compression. For example, the image processing method of claim 10, wherein the size τ to be processed on the score table is set between 3 and 5. The image processing method of claim 3, wherein in step (3), a threshold value for determining a peak of each sub-band in the energy spectrum is found, and the number of direction splits applicable is two or more sub-bands. . The image processing method of claim 1, wherein a convergence region of the low-frequency band of the low-high frequency band has a size of e×b/4, wherein the size of e is set to be at least c, and the low-frequency signal of the high-low frequency band is The size of the convergence region is f × a / 4, where the size of f is set to at least d.