JP4700645B2 - High-definition image generation apparatus and high-definition image generation program - Google Patents

Description

  The present invention relates to a high-definition image generation apparatus and a high-definition image generation program that generate a high-definition image with improved resolution of an input image.

  Conventionally, high conversion (Up conversion) for improving the resolution of low-resolution video is performed by, for example, conversion from SDTV (Standard Definition TeleVision) to HDTV (High Definition TeleVision). ing. This higher resolution is achieved when up-converting the resolution per frame (horizontal, vertical) = (720, 480) of SDTV to the resolution per frame (horizontal, vertical) = (1920, 1080) of HDTV. Is performed by converting the sampling rates in the vertical and horizontal directions, and by obtaining intermediate sample values for newly increasing pixels using an FIR digital filter and complementing them.

  This high resolution is performed when up-converting a low-resolution video, that is, a moving image with a low resolution in real time. However, in the case of a still image, it is not always necessary to perform real-time conversion. There is a need for higher quality images and significant resolution improvements. For this reason, instead of obtaining an intermediate sample value by the FIR digital filter, an intermediate sample value is obtained by polynomial approximation by a bi-linear, bi-cubic, Spline function, etc., and interpolation using this intermediate sample value is performed. Yes.

  However, these high resolution techniques that improve the resolution of moving images and still images both keep the amount of information contained in the original signal and keep the Nyquist sampling theorem, that is, the frequency of the original signal. The purpose is to increase the sampling rate for expressing the signal within the band, and it does not increase (addition / improvement) the amount of information contained in the original signal.

  Incidentally, as a technique for adding and generating information such as definition that was not in the original signal at the same time as improving the sampling rate, for example, there is a complementary generation method based on fluency theory. This complement generation method based on fluency theory is characterized in that the intermediate sample value is complemented using a shift-variant complement function based on the B-Spline function, and higher quality sound quality / texture in speech up-conversion. It is known as a technique for generating additional data. The shift-variant complement function is a function having a different complement function for each sample point location.

  In addition to the complementary generation method based on the fluency theory, for example, a “moving image still area determination device” (Patent Document 1) is disclosed as a technique related to video. This moving image still region determination device uses a wavelet transform technique to add a high-frequency component outside the band of the original signal to the signal to realize nonlinear image quality improvement that improves resolution, The purpose was to increase the definition of SDTV content and upconvert to HDTV.

  More specifically, this moving image still region determination device estimates and extracts edges, contours, etc. included in a video by waveform structure analysis by wavelet transform, and receives this result to obtain a higher sampling rate. It is intended to enhance (enhance or enhance) the edge / contour of the converted signal, and not to expand the band itself in the wavelet transform. In other words, the moving image still region determination device mainly aims to sharpen the edge of the video after up-conversion, and only uses the result of wavelet transform of the original signal for this purpose. .

  Furthermore, Non-Patent Document 1 is disclosed as a technique in which a complementary generation method based on fluency theory is extended and applied to video. The method disclosed in Non-Patent Document 1 performs video up-conversion and high-definition, and performs feature extraction at local locations of the video, such as edges and contours of the video components. The sample points are complemented and generated using different B-Spline functions for each location. As described above, even in the technique disclosed in Non-Patent Document 1, it is important to extract features of a local place on a video when high definition is achieved.

As described above, the conventional technique for video up-conversion disclosed in Patent Document 1 and Non-Patent Document 1 uses an external means for extracting features such as edges from the original video, and the up-conversion This is an extrapolated method of enhancing the extracted edge on the signal form.
JP 2001-309206 A Aokage, H. ; Wada, K.; Toraichi, K; High quality conversion of image resolution using two-dimensional sampling function Communications, Computers and signal Processing, 2003.PACRIM.2003 IEEE Pacific Rim Conference on, Volume: 2,28-30 Aug.2003 .Pages: 720-723 vol.2

  However, in the prior art, since features such as edges / contours are extracted at local locations of a video (image) and the corresponding edges / contours are enhanced, an unnatural emphasis portion may occur. There is a problem. In addition, in the prior art, unless the properties of each video (image) are statistically learned, high definition cannot be performed, and depending on the video (image), a high-definition image with stable quality can be obtained. There is a problem that you can not.

  Therefore, the present invention provides a high-definition image generation apparatus and a high-definition image generation program that can solve the above-described problems and can stably obtain a high-definition image that does not cause unnatural emphasis. Objective.

  In order to solve the above-described problem, the high-definition image generation apparatus according to claim 1 is a high-definition image generation apparatus that generates a high-definition image having a resolution higher than the resolution of the input image that has been input. Means, wavelet transform means, filter generation means, high-frequency fine component generation means, and mixing means.

  According to such a configuration, the high-definition image generation apparatus converts the sampling rate of the input image by the sampling rate conversion unit, and obtains an up-conversion image in which the resolution is higher (up-converted) than the resolution of the input image. Subsequently, the high-definition image generation apparatus performs wavelet transform on the up-conversion image obtained by the sampling rate conversion unit by the wavelet transform unit to obtain a plurality of spatial frequency subband region groups. This spatial frequency subband region group includes horizontal low range and vertical low range (LL), horizontal high range and vertical low range (HL), horizontal low range and vertical high range (LH), horizontal high range and vertical high range. It consists of four spatial frequency subband regions in the region (HH).

  Then, the high-definition image generation device uses the filter generation unit to input each of the input images based on the set spatial frequency subband region group set in advance from among the plurality of spatial frequency subband region groups obtained by the wavelet transform unit. A spatial direction vector indicating the direction in which the pixels correlate is extracted, and based on the spatial direction vector, a direction filter used for complementing the up-conversion image is generated. The set spatial frequency subband region group is composed of two spatial frequency subband regions, a horizontal high region and a vertical low region, and a horizontal low region and a vertical high region. Further, the spatial direction vector is an angle (vector) indicating a direction in which a pixel having the highest correlation with the pixel is located with respect to a certain pixel.

  Then, the high-definition image generation device applies the directional filter generated by the filter generation unit to the set spatial frequency subband region group by the high-frequency fine component generation unit, and is a pixel on the sample space of the up-conversion image A high-frequency component is generated, and the high-frequency component generated by the high-frequency component generation unit and the up-conversion image are mixed by the mixing unit to generate a high-definition image.

  The high-definition image generation device according to claim 2 is the high-definition image generation device according to claim 1, wherein the filter generation unit closes a two-dimensional real number in which each coefficient value of the one-dimensional filter is a discrete sample value. A two-dimensional filter is obtained as the directional filter by complementarily generating a potential function defined on a region and sampling an affine-transformed function that rotates the potential function at a predetermined angle at a lattice point. And

  According to such a configuration, the high-definition image generation apparatus first complementarily generates the potential function using the one-dimensional filter by the filter generation unit. Subsequently, the high-definition image generation apparatus generates a two-dimensional filter as a directional filter by performing affine transformation of the potential function and sampling at a lattice point. By applying this two-dimensional filter to the set spatial frequency subband region in the direction along the spatial direction vector, a new pixel is generated.

  The high-definition image generation program according to claim 3, in order to generate a high-definition image whose resolution is higher than the resolution of the input image that has been input, the computer is converted to sampling rate conversion means, wavelet conversion means, filter generation means, The high frequency component generation unit and the mixing unit are configured to function.

  According to such a configuration, the high-definition image generation program converts the sampling rate of the input image by the sampling rate conversion unit, obtains an up-conversion image in which the resolution is higher than the resolution of the input image, and uses the wavelet conversion unit. Then, wavelet transformation is performed on the up-conversion image obtained by the sampling rate conversion means to obtain a plurality of spatial frequency subband region groups. Then, the high-definition image generation program uses the filter generation unit to generate an input image based on a set spatial frequency subband region group set in advance from a plurality of spatial frequency subband region groups obtained by the wavelet transform unit. A spatial direction vector indicating the direction in which each pixel correlates is extracted, and based on this spatial direction vector, a directional filter used for complementing the up-conversion image is generated. The high-definition image generation program applies the directional filter generated by the filter generation means to the set spatial frequency subband region group by the high-frequency fine component generation means, and is a pixel in the sample space of the up-conversion image. A high-frequency component is generated, and the high-frequency component generated by the high-frequency component generation unit and the up-conversion image are mixed by the mixing unit to generate a high-definition image.

  According to the first and third aspects of the present invention, after the input image is up-converted, the spatial direction vector is extracted using the spatial frequency subband region obtained from the result of wavelet transform, and the spatial direction New pixels are complemented and generated from the up-converted image using the vector and mixed with the image, which is unnatural compared to the conventional high-definition image generation by extrapolation such as edge extraction. It is possible to obtain a high-definition image in which no emphasized portion occurs, that is, a more natural high-definition image.

  According to the second aspect of the present invention, a two-dimensional filter is generated each time according to the input image, and the two-dimensional filter is applied to the up-conversion image to generate a new pixel. Compared to a method that requires statistical learning about the properties of each video (image), this need is eliminated, and a high-definition image with stable quality can be obtained.

Next, embodiments of the present invention will be described in detail with reference to the drawings as appropriate.
(Configuration of high-definition image generator)
FIG. 1 is a block diagram of a high-definition image generation apparatus. As shown in FIG. 1, a high-definition image generation apparatus 1 outputs an input image (original image) as a high-definition image, and outputs a sampling rate conversion unit 3, a wavelet conversion unit 5, a filter A generation unit 7, a high-frequency fine component generation unit 9, and a mixing unit 11 are provided.

  The sampling rate conversion means 3 up-converts the image size of the input image, and outputs the up-converted image (up-conversion image Xu) to the wavelet conversion means 5 and the mixing means 11. In this sampling rate conversion means 3, an existing method may be used for up-conversion, but there is a condition that no distortion occurs within the Nyquist frequency limit of the up-conversion image.

  The wavelet transform unit 5 performs wavelet transform on the up-conversion image Xu output from the sampling rate conversion unit 3 to obtain a plurality of spatial frequency subband regions (spatial frequency subband region group). The band region group is output to the filter generation means 7.

  The spatial frequency subband region group is LL (horizontal low frequency, vertical low frequency) where the horizontal component is in the low frequency band, the vertical component is in the low frequency band, and the horizontal component is in the high frequency range. HL (horizontal high frequency, vertical low frequency) where the vertical component is the low frequency band region in the frequency band, the horizontal component is the low frequency band, and the vertical component is the high frequency region LH (horizontal low frequency, vertical high frequency), HH (horizontal high frequency, vertical high frequency) where the horizontal component is the high frequency band and the vertical component is the high frequency band. It consists of a spatial frequency subband region.

Here, wavelet transformation by the wavelet transformation means 5 will be described. This wavelet transform is a “two-dimensional wavelet transform of an image”, where f _L is a low-frequency one-dimensional wavelet transform filter in the horizontal direction, f _H is a high-frequency one-dimensional wavelet transform filter in the horizontal direction, and g _L is vertical. A low-pass one-dimensional wavelet transform filter in the direction, and g _H is a high-pass one-dimensional wavelet transform filter in the vertical direction. Note that these _{f L, f H, g} L and _{g H} is specifically expressed in matrix is used in Equation (2) to (4) described later.

First, “two-dimensional wavelet transform of an image” will be described.
An image having the number of vertical pixels M and the number of horizontal pixels N is set to X [M, N], and a small image area (vertical 2m + 1 pixels, horizontal with the position (i, j) on the image X [M, N] as the center position. 2n + 1 pixels) is x [i, j]. The following example is an example of x [i, j] when m = 2 and n = 3. If the value of the pixel X [i, j] at the center position (i, j) is s _0,0 , x [i, j] is represented by the following formula (1).

Then, f is an odd symmetric one-dimensional FIR filter with 2n + 1 taps f = (f ₋₃ , f ₋₂ , f ₋₁ , f ₀ , f ₁ , f ₂ , f ₃ ), and g is the number of taps 2m + 1. , An oddly symmetric one-dimensional FIR filter g = (g ₋₂ , g ₋₁ , g ₀ , g ₁ , g ₂ ), f is horizontally applied to the small image region x [i, j]. A scalar value y [i, j] as a result of performing the two-dimensional FIR process in which g is applied in the vertical direction is represented by the following formula (2).

In this formula (2), ^{g t} is the transposed matrix of g, matrices × in ○ are, Kronecker product (matrix cross product), "." Indicates a matrix dot product.
An image Y [M, N], which is the result of performing the two-dimensional FIR process over the entire area of the image X [M, N], is expressed by the following formula (3).

  As shown in the equation (3), the image Y [M, N] is a matrix of M rows and N columns where the i row and j column components are y [i, j].

  Then, the spatial frequency subband region group (LL, HL, LH, HH) is expressed by the following formula (4).

  As described above, in this embodiment, after performing the filtering process shown in Equation (4), the result of performing 2: 1 sub-sampling in the vertical and horizontal axis directions is expressed as “two-dimensional wavelet transform”. It is defined as

  Further, an example of the spatial frequency subband region group will be described with reference to FIG. FIG. 3 is a diagram showing an up-conversion image (a female pupil portion) and a spatial frequency subband region group of this image together.

  By applying wavelet transform to the up-conversion image shown in FIG. 3A, the spatial frequency subband region LL shown in FIG. 3B and the spatial frequency subband region HL shown in FIG. The spatial frequency subband region LH shown in FIG. 3D and the spatial frequency subband region HH shown in FIG.

As shown in FIG. 3B, the spatial frequency subband region LL is a region closest to the state of the up-conversion image.
As shown in FIG. 3C, the spatial frequency subband region HL is a region in which only outstanding vertical lines (shadows) and diagonal lines remain in the up-conversion image.

As shown in FIG. 3D, the spatial frequency subband region LH is a region in which only the outstanding horizontal lines (shadows) and diagonal lines remain in the up-conversion image.
As shown in FIG. 3 (e), the spatial frequency subband region HH is a region in which only a prominent oblique line remains in the up-conversion image.
FIG. 3F shows the arrangement of the spatial frequency subband region groups (LL, HL, LH, HH).

Returning to FIG.
The filter generation means 7 is a two-dimensional filter based on a set spatial frequency subband area group preset from the spatial frequency subband area groups (LL, HL, LH, HH) output from the wavelet transform means 5. Intensity map generation means 7a, directional theta map generation means 7b, and two-dimensional FIR filter generation means 7c.

  In this embodiment, the subband sum (the high frequency component Xhigh of the up-conversion image Xu) is obtained using HL and LH as the set spatial frequency subband region group, and the sum of the subbands is , Xhigh = HL [i, j] + LH [i, j]. Further, Xhigh = HL [i, j] + LH [i, j] + HH [i, j] (horizontal and vertical high-frequency component Xhigh of the up-conversion image Xu) may be used as the sum of the subbands.

  The intensity map generation means 7a uses the values of the HL and LH pixels selected from the spatial frequency subband region group (LL, HL, LH, HH) output from the wavelet transform means 5 to calculate the intensity map. Generate a map. This intensity map shows a significant degree of spatial correlation in a predetermined direction at a certain pixel position (i, j) (spatial correlation θ [i, j] in θ direction described later). And this intensity map is calculated | required by Numerical formula (5) shown next.

  The directional theta map generating means 7b uses the values of the HL and LH pixels selected from the spatial frequency subband region groups (LL, HL, LH, HH) output from the wavelet transform means 5, and θ A direction spatial correlation map (directional theta map) is generated. This directional theta map shows how much a certain pixel is correlated with the θ direction (how much the value of the pixel is correlated), that is, the spatial correlation θ [i, j in the θ direction. ]. And this directional theta map is calculated | required by Numerical formula (6) shown next.

  The two-dimensional FIR filter generation unit 7c performs threshold processing on the intensity map generated by the intensity map generation unit 7a, and the directory generated by the directional theta map generation unit 7b is set at a pixel position equal to or greater than the threshold. A directional LPF to be applied to the pixel position of the high frequency component Xhigh of the up-conversion image Xu is generated using the national theta map. This directional LPF is obtained by extracting the spatial directional vector of each pixel of the up-conversion image Xu.

If this directional LPF is f [θ [i, j]], this f [θ [i, j]] rotates the one-dimensional filter f [0] by θ [i, j] by affine transformation. It has been made. Note that f [θ [i, j]] can be represented by a two-dimensional matrix. For example, when f [0] = {f ₋₃ , f ₋₂ , f ₋₁ , f ₀ , f ₁ , f ₂ , f ₃ }, a two-dimensional matrix is obtained as shown in the following formula (7). Become.

  As shown in Equation (7), f [0] is arranged in a two-dimensional matrix of 11 rows and 11 columns, and the two-dimensional matrix is a matrix in which f [0] is surrounded by “0”. It becomes. This two-dimensional matrix can be regarded as 11 × 11 discrete sample values of a potential function defined on a two-dimensional real closed region.

  Then, the virtual potential function (the definition area is a real continuous area) having 11 × 11 discrete sample values is defined as f (x, y), and this f (x, y) is affine transformed at an angle θ. Let the function be h (x, y). By sampling h (x, y) at the lattice points, a matrix of directional filters can be obtained.

  Here, an example of generating a directional LPF will be described with reference to FIG. FIG. 4 shows an example of generating a directional LPF. (A) and (d) are diagrams showing the filter coefficients of the LPF in pixel shading, and (b) and (e) are the LPF filter coefficients. It is the figure which showed the filter coefficient by the height of the impulse, (c) and (f) are the figures which showed the frequency response of the said LPF.

  In FIG. 4, the one-dimensional filter f [0] serving as a base is f = {− 1/8, 1/4, 3/4, 1/4, −1/8}, and θ is π / 4. The impulse response of the two-dimensional FIR filter obtained by rotating is shown in FIGS. 4 (a) and 4 (b), and the two-dimensional spatial frequency response is shown in (c). Also, the impulse response of the two-dimensional FIR filter obtained by rotating θ by −π / 4 is shown in FIGS. 4D and 4E, and the two-dimensional spatial frequency response is shown in FIG.

  The pixel shading shown in FIG. 4A shows the difference in values applied to the high frequency components of the up-conversion image due to the difference in hatching (difference in shading) for nine pixels. The values applied to the pixels and the lower left pixel are configured to be high.

  The impulse height shown in FIG. 4 (b) indicates the difference in value applied to the high frequency component of the up-conversion image depending on the spectrum, and the near side position (x, y) = (6, 4 ) And the spectrum at the back position (x, y) = (4, 6).

  The frequency response shown in FIG. 4C shows the difference in values applied to the high frequency components of the up-conversion image by the contour lines, and the position (H, V) = (− 0.5, 0.5) It is configured to be high in the vicinity.

  The pixel shading shown in FIG. 4 (d) shows the difference in values applied to the high frequency components of the up-conversion image due to the difference in hatching (difference in shading) for nine pixels. A value applied to the lower pixel and the upper left pixel is configured to be high.

  The impulse height shown in FIG. 4 (e) indicates the difference in value applied to the high frequency component of the up-conversion image depending on the spectrum, and the left position (x, y) = (4, 4). And the spectrum at the right position (x, y) = (6, 6) are configured to be higher.

  The frequency response shown in FIG. 4 (f) shows the difference in values applied to the high frequency components of the up-conversion image by the contour lines, and the position (H, V) = (0.5, − 0.5) It is configured to be high in the vicinity.

Returning to FIG.
The high-frequency fine component generating means 9 is the sum of the high-frequency components of the set spatial frequency subband area group selected from the spatial frequency subband area groups (LL, HL, LH, HH) output from the wavelet transform means 5. (HL and LH), that is, the directional LPF generated by the two-dimensional FIR filter generating means 7c is applied to the high frequency component Xhigh of the up-conversion image Xu to generate the high frequency fine component Xext.

  The directional LPF is set to f [θ [i, j]], and the high frequency component Xext of the up-conversion image Xu is subjected to two-dimensional filtering using this directional LPF. It is calculated | required by (8). Note that f [θ [i, j]] is a two-dimensional matrix.

  Here, xu [i, j] is an image small area centered on the position (i, j) of the image Xu, and this size is the same size as f [θ [i, j]].

  In the high-frequency fine component generating means 9, the mathematical formula (6) used in the directional theta map generating means 7b is defined as a “spatial direction vector” in the pixel X [i, j], and this pixel X [i , J] are treated as having the strongest pixel correlation in the θ direction in space. The high-frequency fine component generation means 9 assumes that the spatial frequency component in the θ direction is as close as possible to zero with regard to the strongest pixel correlation, and predicts the pixel value in the θ direction (the high-frequency fine component to be generated). It can be done by simple LPF processing.

  Here, an example of generating an image (component) obtained by the filter generation unit 7 and the high-frequency fine component generation unit 9 will be described with reference to FIG. FIG. 5 shows, as an example of image generation, superior pixels detected by threshold processing of the high-frequency component Xhigh and intensity map of the up-conversion image Xu for the same input image as shown in FIG. FIG. 5 is a diagram showing a distribution, a directional theta map, and a high-frequency fine component Xext.

  The high frequency component Xhigh of the up-conversion image Xu shown in FIG. 5A is the sum (HL + LH) of the high frequency components of the set spatial frequency subband region group, and the low frequency component that is the vertical low frequency of HL and the LH It is possible to observe a prominent vertical line (shadow) and a prominent horizontal line (shadow) as compared to FIGS. 3C and 3D because the horizontal low-frequency components of each other cancel each other out. Can not.

  The distribution of dominant pixels (pixels above the threshold) detected by the threshold processing of the intensity map shown in FIG. 5B represents pixels above the threshold on a white background, and other pixels on a black background. It is a thing. As can be seen from FIG. 5B, the spatial correlation in a predetermined direction at a certain pixel position is high, here, the boundary between the hair and the face, the edge of the eyebrows, the boundary between the pupil and the face, the pupil portion, and the like are observed. be able to.

  The directional theta map shown in FIG. 5C indicates how much a certain pixel is correlated with the θ direction. As can be seen from FIG. 5C, the same part as in FIG. 5B can be observed more remarkably. Although not confirmed in FIG. 5C, the color of the pixel differs based on the value of θ indicating the direction in which the pixels are correlated, based on FIG.

In the high-frequency fine component Xext shown in FIG. 5 (d), as in FIG. 5 (a), it is impossible to observe an outstanding vertical line (shadow) and an outstanding horizontal line (shadow). On the contrary, since this high-frequency fine component Xext does not include a markedly emphasized portion, it is mixed with the up-conversion image Xu by the mixing means 11 described later, thereby generating a high-definition image that does not produce an unnatural emphasized portion. be able to.
FIG. 5 (e) shows the color display of θ (display by RGB). Although it cannot be confirmed in FIG. 5 (e), blue (B), clockwise on the circle depending on the value of θ, It shows that the hue changes to green (G) and red (R).

Returning to FIG.
The mixing means 11 generates and outputs a high-definition image obtained by mixing the up-conversion image Xu output from the sampling rate conversion means 3 and the high-frequency fine component Xext generated by the high-frequency fine component generation means 9. It is.

  Assuming that the high-definition image generated by the mixing unit 11 is Yu [i, j], the high-definition image Yu [i, j] is obtained by the following formula (9) or (10).

In Expression (9), α is a positive real number, and the high-definition image Yu is controlled to add the high-frequency fine component Xext to the up-conversion image Xu. Further, in Equation (10), the high-definition image Yu is obtained by adding the high-frequency fine component Xext to the spatial frequency subband region group (LL). In this case, since the spatial frequency subband region group (LL) is a component that is additively superimposed on the up-conversion image Xu, even if the high-definition image Yu is obtained in this way, a change occurs in the effect of the present invention. Absent.

  According to the high-definition image generating apparatus 1, the input image is up-converted by the sampling rate converting unit 3, and then the filter generating unit is used by using the spatial frequency subband region obtained from the result of wavelet transform by the wavelet converting unit 5. 7 generates a directional LPF (extracted spatial directional vector), and the high-frequency fine component generation means 9 complements and generates a new pixel from the up-converted image using the directional LPF, and the mixing means 11 is mixed with the image, so that a high-definition image in which an unnatural emphasis portion is not generated, that is, a more natural image, is generated as compared with the generation of a high-definition image by a conventional extrapolation technique such as edge extraction. High definition images can be obtained.

  Further, according to the high-definition image generating device 1, the filter generating means 7 generates a directional LPF (two-dimensional filter) each time according to the input image, and applies the directional filter to the up-conversion image. Since new pixels are generated, this is not necessary as compared with a conventional method that requires statistical learning about the properties of each video (image), and a high-definition image with stable quality can be obtained.

(Operation of high-definition image generator)
Next, the operation of the high-definition image generating apparatus 1 will be described with reference to the flowchart shown in FIG. 2 (see FIG. 1 as appropriate).
First, the high-definition image generating apparatus 1 uses the sampling rate conversion unit 3 as the up-conversion image Xu in which the sampling rate of the input image is changed, and outputs the up-conversion image to the wavelet conversion unit 5 and the mixing unit 11 (step S1). ).

  Subsequently, the high-definition image generating apparatus 1 obtains a spatial frequency subband region group (LL, HL, LH, HH) from the up-conversion image Xu by the wavelet transform unit 5, and this spatial frequency subband region group (LL). , HL, LH, HH) are output to the filter generation means 7 and the high-frequency fine component generation means 9 (step S2).

  Then, the high-definition image generating apparatus 1 uses the intensity map of the filter generating unit 7 from the set spatial frequency subband region group (HL, LH) selected from the spatial frequency subband region group (LL, HL, LH, HH). The intensity map is generated by the generating means 7a, the directional theta map is generated by the directional theta map generating means 7b, and based on these, the directional filter is generated by the two-dimensional FIR filter means 7c (step S3).

  Further, the high-definition image generating apparatus 1 uses the high-frequency fine component generating means 9 to sum the high-frequency components of the set spatial frequency sub-band region group selected from the spatial frequency sub-band region groups (LL, HL, LH, HH). (HL and LH), that is, the directional LPF generated by the two-dimensional FIR filter generating means 7c is applied to the high frequency component Xhigh of the up-conversion image Xu to generate the high frequency fine component Xext (step S4).

  Thereafter, the high-definition image generating apparatus 1 mixes the up-conversion image Xu and the high-frequency fine component Xext generated by the high-frequency fine component generating means 9 by the mixing unit 11 and outputs a high-definition image ( Step S5).

  As mentioned above, although embodiment of this invention was described, this invention is not limited to the said embodiment. For example, in the present embodiment, the high-definition image generation apparatus 1 has been described. However, each configuration of the apparatus can be regarded as a high-definition image generation program by grasping each component as a unit that causes a computer to function. In this case, the same effect as that of the device 1 can be obtained.

1 is a block diagram of a high-definition image generation device according to an embodiment of the present invention. 2 is a flowchart illustrating an operation of the high-definition image generation device illustrated in FIG. 1. It is the figure which showed the example of the spatial frequency subband area | region group. It is the figure shown about the example of production | generation of a directional two-dimensional LPF. It is the figure shown about the example of a production | generation of the image (component) obtained by a filter production | generation means and a high region fine component production | generation means.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 High-definition image generation apparatus 3 Sampling rate conversion means 5 Wavelet conversion means 7 Filter generation means 9 High frequency-definition component generation means 11 Mixing means

Claims (3)

A high-definition image generation device that generates a high-definition image having a higher resolution than the resolution of an input image that has been input,
Sampling rate conversion means for converting a sampling rate of the input image and obtaining an up-conversion image having a higher resolution than the resolution of the input image;
Wavelet transform means for obtaining a spatial frequency subband region group composed of a plurality of spatial frequency subband regions by performing wavelet transform on the up-conversion image obtained by the sampling rate conversion means;
Based on a set spatial frequency subband region group set in advance from the spatial frequency subband region group obtained by this wavelet transform means, a spatial direction vector indicating the direction in which each pixel of the input image correlates is extracted. And, based on the spatial direction vector, a filter generation means for generating a direction filter used for complementing the up-conversion image,
A high-frequency fine component generating means for applying a directional filter generated by the filter generating means to the set spatial frequency subband region group to generate a high-frequency fine component that is a pixel on a sample space of the up-conversion image; ,
Mixing means for mixing the high-frequency fine component generated by the high-frequency fine component generating means and the up-conversion image to generate the high-definition image;
A high-definition image generating apparatus comprising: The filter generation means includes
Complementary generation of a potential function defined on a two-dimensional real closed region with each coefficient value of the one-dimensional filter as a discrete sample value, and a function subjected to affine transformation that rotates the potential function at a predetermined angle at a lattice point The high-definition image generation apparatus according to claim 1, wherein a two-dimensional filter is obtained as the directional filter by sampling. In order to generate a high-definition image whose resolution is higher than the resolution of the input image input,
Sampling rate conversion means for converting a sampling rate of the input image and obtaining an up-conversion image having a higher resolution than that of the input image,
Wavelet transform means for applying a wavelet transform to the up-conversion image obtained by the sampling rate conversion means to obtain a spatial frequency subband region group composed of a plurality of spatial frequency subband regions,
Based on a set spatial frequency subband region group set in advance from the spatial frequency subband region group obtained by this wavelet transform means, a spatial direction vector indicating the direction in which each pixel of the input image correlates is extracted. Filter generating means for generating a directional filter used for complementing the up-conversion image based on the spatial direction vector,
Applying the directional filter generated by the filter generation means to the set spatial frequency subband region group, a high-frequency fine component generation means for generating a high-frequency fine component that is a pixel on the sample space of the up-conversion image,
Mixing means for mixing the high-frequency fine component generated by the high-frequency fine component generating means and the up-conversion image to generate the high-definition image;
A high-definition image generation program characterized by functioning as