CN103208103B - A kind of enhancement method of low-illumination image based on GPU - Google Patents

Abstract

The present invention is applicable to field of Internet communication, disclose a kind of enhancement method of low-illumination image based on GPU, adopt and calculate Unified Device framework (CUDA) programming model, including: obtain original image RGB channel data and dimension information thereof, GPU initializes, the number of threads performing Kernel function is determined according to original image size, then Kernel function is started, ask for original image luminance component, and original luminance component is carried out image enhancement processing, non-linear modulation and contrast including luminance component strengthen, restart Kernel function, recover color of image, result of calculation is passed back CPU end, save as picture, finally release GPU video memory.Present invention achieves brightness of image Automatic adjusument, utilize higher dimensional space image method of geometry to carry out contrast enhancing, it is to avoid halation phenomenon, especially use GPU to accelerate parallel, be effectively increased processing speed.

Description

A kind of enhancement method of low-illumination image based on GPU

Technical field

The invention belongs to field of Internet communication, particularly relate to a kind of enhancement method of low-illumination image based on GPU.

Background technology

Image enhaucament is an important technology field in Digital Image Processing, owing to being subject to the impact of external environment in image acquisition process, such as night, mist or the condition such as overcast and rainy, will be substantially reduced picture quality.Therefore, the purpose of image enhaucament is in that, by adjusting dynamic range of images, to improve picture contrast to highlight image detail information, improve visual quality of images, thus providing information assurance for the field such as intelligent monitoring, Aero-Space.

Image enhaucament based on Histogram adjustment is widely used as classic algorithm, such as histogram equalization and Histogram Matching.Although the contrast between image light and shade region can be improved based on histogrammic image enhaucament, but owing to this algorithm utilizes the statistical information of image overall intensity, for a kind of Global treatment mode, ignore the local message of image, therefore the local contrast of image does not improve, contrary to carrying out unifying to regulate for the pixel of same grey level in image, weaken the detailed information of image on the contrary.

The image enhaucament mapped based on tone is also a kind of more common method, the method adopts simple nonlinear mapping function to regulate dynamic range of images, such as Sigmoid function, Log function etc., owing to such nonlinear function form is single, it is typically only capable to that luminance level is in a range of image implement effectively to regulate, the method essence is also the mode of Global treatment simultaneously, it is impossible to distinguish the detailed information of image zones of different.

Additionally, Kimmel etc. propose the Retinex algorithm for image enhancement under a kind of variation framework, in this algorithm, object reflecting component and illumination component are all based on slickness a priori assumption, and illumination component is considered as multiplicative noise, thus construct unified variational function.Retinex algorithm for image enhancement form under variation framework is simple, solving speed fast, but it assumes that illumination component is slow even variation, causes that illumination component is estimated inaccurate when image has strong edge, thus being easily generated halation phenomenon.

By above-mentioned analysis it is found that local contrast enhancing, brightness Automatic adjusument and effective halation phenomenon that eliminates are the key issues that image enhaucament to solve.On the other hand, along with technical development, the image resolution ratio collected improves constantly, image enhaucament problem often faced by be mass data, especially when algorithm complex is higher, it is desirable to rely only on CPU realization process high-quality and efficient the completing of task there is great difficulty.

Due to GPU(GraphicProcessingUnit, Graphics Processing Unit) a large amount of transistors are used for ALU, it is made to show huge advantage in parallel computation field, and the development along with GPGPU (GeneralPurposeGPU) technology, memory bandwidth and the floating number computing capability of GPU improve rapidly, therefore compare it with CPU and are suitable for the parallel data process of height calculating density.Especially, after NVIDIA releases CUDA (ComputeUnifiedDeviceArchitecture), provide a set of C language programming platform for developer, greatly reduce programming complexity so that GPU programming technique is widely used in image processing field.

Summary of the invention

Embodiments provide a kind of enhancement method of low-illumination image based on GPU, it is possible to realize brightness Automatic adjusument, be prevented effectively from halation phenomenon, especially use GPU parallel processing, improve the execution efficiency of method.

For this, embodiments provide following technical scheme:

A kind of enhancement method of low-illumination image based on GPU, described method adopts and calculates Unified Device framework CUDA programming model, comprises the following steps:

Reading in original low-light (level) image, obtain original image RGB channel data and image dimension information, described image dimension information includes width and height；

GPU initializes, and including arranging texture and CUDA array type, opens up GPU video memory for data, described original image RGB channel data assignment is given CUDA array binded texture internal memory A；

Determine the number of threads performing Kernel function according to described image dimension information, including arranging block size and arranging grid size, described block, refer to GPU thread block unit, described grid, refer to GPU thread net unit；

Start described Kernel function, from described texture memory A, read described original image RGB channel data, ask for original image luminance component, result of calculation is stored in described CUDA array and binds with texture memory B；

Start described Kernel function, described original image luminance component is read from described texture memory B, it is carried out image enhancement processing, enhanced image brightness data being stored in described CUDA array and binds with texture memory C, described image enhancement processing includes non-linear modulation and the contrast enhancing of luminance component；

Start described Kernel function, described original image RGB channel data, described original image luminance component and described enhanced luminance component image is read respectively from described texture memory A, described texture memory B and described texture memory C, color of image is recovered according to the data obtained, result of calculation is passed back CPU end, and output preserves；

Release GPU video memory.

Compared with prior art, embodiments of the invention have the advantage that

First, the enhancement method of low-illumination image based on GPU that the present invention proposes, achieve brightness Automatic adjusument, higher dimensional space image method of geometry is used for Image Warping strengthen, effectively inhibit the appearance of halation phenomenon, and utilize CUDA technology, it is achieved that the GPU of algorithm accelerates parallel, being effectively increased treatment effeciency, speed-up ratio is up to 3-5 times.

And, the Overall Steps of algorithm for image enhancement of the present invention all realizes on GPU, therefore only needs CPU end to transmit to GPU end and GPU end to the data of CPU end, it is to avoid data are frequently transmitted, thus being conducive to the raising of efficiency of algorithm.

Finally, in the present invention, the data reaching GPU end are stored in Texture memory mostly, reduce the reading speed of data, further ensure the treatment effeciency of algorithm.

Accompanying drawing explanation

Fig. 1 is the method flow diagram of the enhancement method of low-illumination image based on GPU that the embodiment of the present invention provides；

Fig. 2 be the embodiment of the present invention provide for non-linear modulation curve with the situation of change figure of parameter θ；

Fig. 3 is that the contrast enhancing that the embodiment of the present invention provides realizes method schematic diagram；

Fig. 4 is the test result comparison diagram that the embodiment of the present invention provides.

Detailed description of the invention

In order to make the purpose of the present invention, technical scheme and advantage clearly understand, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that described herein is only a part of embodiment of the present invention, rather than whole embodiments.Based on the embodiment in the present invention, the every other embodiment that those of ordinary skill in the art obtain under the premise not making creative work, broadly fall into the scope of protection of the invention.

Fig. 1 is the method flow diagram of the enhancement method of low-illumination image based on GPU that the embodiment of the present invention provides, and for the ease of illustrating, illustrate only the part relevant to the embodiment of the present invention.

Image RGB color image used by the present embodiment, image size 600 × 398 pixel, the deep 24bit in position.

Test platform: CPU:Intel (R) Core (TM) 2QuadCPUQ66002.4GHz (4CPUs)；GPU:NVIDIAGeforce8400GS(2SM, 16SP).

As it is shown in figure 1, the method comprises the following steps:

Step 101, reads in original image, obtains RGB data.

Concrete, reading in original low-light (level) image, obtain original image RGB channel data and image dimension information, described image dimension information includes width and height, is designated as imgW and imgH.

Step 102, initializes GPU.

Concrete, GPU initializes, and including arranging texture and CUDA array type, opening up GPU video memory for data, described original image RGB channel data assignment being given CUDA array binded texture internal memory A, thus realizing the transmission to GPU end of the data CPU end.

Meanwhile, determine the number of threads performing Kernel function according to described image dimension information, including arranging block size and arranging grid size, described block, refer to GPU thread block unit, described grid, refer to GPU thread net unit.

Preferably, described block size, may be configured as 128 or 256, described grid size is arranged according to described image dimension information and block size.

Preferably, configurable block is 16 × 16 two dimension thread block, and grid size is determined according to following formula:

GridX=(imgW+blockX-1) blockX

GridY=(imgH+blockY-1) blockY

Wherein, blockX is thread block X-direction size, and blockY is thread block Y-direction size, and gridX is thread net X-direction size, and gridY is thread net Y-direction size.

Step 103, extracts luminance component image.

Concrete, start described Kernel function, from described texture memory A, read described original image RGB channel data, ask for original image luminance component, result of calculation is stored in described CUDA array and binds with texture memory B.

Preferably, calculate the maximum of each pixel RGB triple channel gray value, obtain described original image luminance component, and result of calculation is stored in CUDA array and binds with texture memory B.

Step 104, calculates non-linear brightness and regulates parameter.

Concrete, start described Kernel function, from described texture memory B, read described original image luminance component, it is carried out image enhancement processing, enhanced image brightness data is stored in described CUDA array and binds with texture memory C.

Wherein, image enhancement processing includes non-linear modulation (i.e. step 105) and contrast enhancing (i.e. step 106) of luminance component, and concrete, after the non-linear modulation of luminance component, brightness meets functional relationship with original brightness:

$I_1=I^{p_1+p_2\mathrm{\θ}+p_3{\mathrm{\θ}}^2}$

In formula, I is original luminance value, I₁For brightness value after non-linear modulation, p₁, p₂, p₃It is constant, and meets p₁+p₂+p₃=1, θ is Automatic adjusument parameter, and its value is determined according to the overall brightness of input picture, specifically can apply following formula:

$\mathrm{\&theta;}=\left\{\begin{array}{cc}0& I_{0.1N}<l_0\\ (I_{0.1N}-l_0)/l_2& l_0\&le;I_{0.1N}\&le;l_1\\ 1& I_{0.1N}>l_1\end{array}\right.$

Wherein, l₀, l₁, l₂For constant, span meet (0,255], I_0.1NRepresenting that the ascending sequence of brightness value to pixels all in image is in the brightness value of 0.1N position, N is then image slices vegetarian refreshments number.

Step 105, non-linear brightness is modulated.

Concrete, starting Kernel function, from texture memory B, read original luminance component, it is carried out image enhancement processing, described image enhancement processing includes non-linear modulation and the contrast enhancing of luminance component.

Luminance non-linearity modulation adopts function:

$I_1=I^{p_1+p_2\mathrm{\&theta;}+p_3{\mathrm{\&theta;}}^2}$

In formula, I is original luminance value, I₁For brightness value after non-linear modulation, p₁, p₂, p₃It is constant, and meets p₁+p₂+p₃=1, θ is Automatic adjusument parameter, and its value is determined according to the overall brightness of input picture, specifically can apply following formula:

$\mathrm{\&theta;}=\left\{\begin{array}{cc}0& I_{0.1N}<l_0\\ (I_{0.1N}-l_0)/l_2& l_0\&le;I_{0.1N}\&le;l_1\\ 1& I_{0.1N}>l_1\end{array}\right.$

Wherein, l₀, l₁, l₂For constant, span meet (0,255], I_0.1NRepresenting that the ascending sequence of brightness value to pixels all in image is in the brightness value of 0.1N position, N is then image slices vegetarian refreshments number.

Preferably, l is set₀=50, l₁=170, l₂=120, Fig. 2 then illustrates the non-linear modulation curve situation of change (in figure, brightness value is normalized to [0,1]) with parameter θ.

Step 106, contrast strengthens.

Luminance component, after non-linear modulation, need to carry out local contrast enhancing, further to highlight image detail information.Local contrast strengthens employing higher dimensional space image method of geometry, as it is shown on figure 3, kth time iteration result meets:Preferably, α=0.5.

Concrete, contrast strengthens, and adopts higher dimensional space image method of geometry, and image is regarded as the vector in higher dimensional space by described method, comprises the steps:

Step one: original image (i.e. former vector) degree of comparing degeneration is processed, obtains degraded image (vector of namely degenerating) including by low-pass filtering method.

Preferably, contrast degeneration processes selects boxlike filtering BoxFilter, and filter radius is 2.

Step 2: described degeneration vector and described former vector are made difference and determined degeneration direction, its opposite direction is that contrast strengthens direction.

Step 3: described former vector strengthens weighted direction summation with described contrast, it is thus achieved that image after contrast enhancing；

Step 4 (i.e. step 107): repeat step one to step 3, until meeting iteration stopping condition to stop iteration.

Step 107, it may be judged whether meet iteration stopping condition, if so, enters next step, if it is not, repeat step 106.

Concrete, repeat step one to step 3, until meeting iteration stopping condition to stop iteration.

Concrete, described iteration stopping condition includes: before and after meeting, the difference of twice iteration result is not more than a certain constant, or arranges maximum iteration time, after iterations exceedes described maximum iteration time, also stops iteration.

Preferably, iteration stopping condition is then by calculating the Euclidean distance of kth time iteration result and kth-1 result, it may be judged whether meet formula:And set maximum iteration time: Iter_max=10.

Afterwards, enhanced image brightness data is stored in CUDA array and binds with texture memory C, for subsequent step (i.e. step 108).

Step 108, color is recovered.

Concrete, start described Kernel function, described original image RGB channel data, described original image luminance component and described enhanced luminance component image is read respectively from described texture memory A, described texture memory B and described texture memory C, above three part data according to obtaining recover color of image, result of calculation is passed back CPU end, and output preserves.

Preferably, the described data according to acquisition recover color of image, can adopt linear restoring method, and concrete formula is as follows:

$R_i^{\&prime;}=R_i{I}_i^{\&prime;}/I_i,i=1,...,N$

$G_i^{\&prime;}=G_i{I}_i^{\&prime;}/I_i,i=1,...,N$

$B_i^{\&prime;}=B_i{I}_i^{\&prime;}/I_i,i=1,...,N$

In formula, R_i,G_i,B_iFor the RGB triple channel gray value of pixel a certain in original image, I_iFor the brightness value of a certain pixel,WithA certain pixel RGB triple channel gray value and this brightness after expression process, N represents pixel number, applies the whole pixels in described formula traversing graph picture, it is possible to obtain enhanced image.

Step 109, image after output processing.

After adopting step 108, result of calculation is passed back CPU end, saves as picture, as shown in Figure 4.

Step 110, discharges GPU.

Release GPU video memory, namely discharges global memory and CUDA array internal memory that GPU end is opened up.

The embodiment of the present invention calculates Unified Device framework programming model by adopting, obtain original image RGB channel data and dimension information thereof, GPU initializes and determines the number of threads performing Kernel function according to original image size, then Kernel function is started, ask for original image luminance component, original luminance component is carried out image enhancement processing, non-linear modulation and contrast including luminance component strengthen, finally start Kernel function, recover color of image, result of calculation is passed back CPU end, save as picture and discharge GPU video memory, achieve brightness of image Automatic adjusument, higher dimensional space image method of geometry is utilized to carry out contrast enhancing, avoid halation phenomenon, especially GPU is used to accelerate parallel, it is effectively increased processing speed.

Through the above description of the embodiments, those skilled in the art is it can be understood that can add the mode of required general hardware platform by software to the present invention and realize, naturally it is also possible to by hardware, but in a lot of situation, the former is embodiment more preferably.Based on such understanding, the part that prior art is contributed by technical scheme substantially in other words can embody with the form of software product, this computer software product is stored in a storage medium, including some instructions with so that a station terminal equipment (can be mobile phone, personal computer, server, or the network equipment etc.) perform the method described in each embodiment of the present invention.

The above is only the preferred embodiment of the present invention; it should be pointed out that, for those skilled in the art, under the premise without departing from the principles of the invention; can also making some improvements and modifications, these improvements and modifications also should look protection scope of the present invention.

Claims (9)

1., based on an enhancement method of low-illumination image of GPU, described method adopts and calculates Unified Device framework CUDA programming model, it is characterised in that comprise the following steps:

A: read in original low-light (level) image, obtains original image RGB channel data and image dimension information, and described image dimension information includes width and height；

B:GPU initializes, and including arranging texture and CUDA array type, opens up GPU video memory for data, described original image RGB channel data assignment is given CUDA array binded texture internal memory A；

C: determine the number of threads performing Kernel function according to described image dimension information, including arranging block size and arranging grid size, described block, refers to GPU thread block unit, described grid, refers to GPU thread net unit；

D: start described Kernel function, reads described original image RGB channel data from described texture memory A, asks for original image luminance component, result of calculation is stored in described CUDA array and binds with texture memory B；

E: start described Kernel function, described original image luminance component is read from described texture memory B, it is carried out image enhancement processing, enhanced image brightness data being stored in described CUDA array and binds with texture memory C, described image enhancement processing includes non-linear modulation and the contrast enhancing of luminance component；

F: start described Kernel function, described original image RGB channel data, described original image luminance component and described enhanced luminance component image is read respectively from described texture memory A, described texture memory B and described texture memory C, color of image is recovered according to the data obtained, result of calculation is passed back CPU end, and output preserves；

G: release GPU video memory；

Wherein, described step E is further comprising the steps of:

E2: described contrast strengthens, adopts higher dimensional space image method of geometry, and image is regarded as the vector in higher dimensional space by described method, comprises the steps:

E201: original image degree of comparing degeneration processed, obtain degraded image, described original image and former vector including by low-pass filtering method, namely described degraded image degenerates vector；

E202: described degeneration vector and described former vector are made difference and determined degeneration direction, its opposite direction is that contrast strengthens direction；

E203: described former vector strengthens weighted direction summation with described contrast, it is thus achieved that image after contrast enhancing；

E204: repeat step e201 to e203, until meeting iteration stopping condition to stop iteration.

2. the enhancement method of low-illumination image based on GPU as claimed in claim 1, it is characterised in that described step C is further comprising the steps of:

C1: described block size, may be configured as 128 or 256, and described grid size is arranged according to described image dimension information and block size.

3. the enhancement method of low-illumination image based on GPU as claimed in claim 2, it is characterised in that described step c1 is further comprising the steps of:

C2: configurable block is 16 × 16 two dimension thread block, and grid size is determined according to following formula:

GridX=(imgW+blockX-1)/blockX

GridY=(imgH+blockY-1)/blockY

Wherein, blockX is thread block X-direction size, and blockY is thread block Y-direction size, and gridX is thread net X-direction size, and gridY is thread net Y-direction size, and imgW is picture traverse, and imgH is picture altitude.

4. the enhancement method of low-illumination image based on GPU as claimed in claim 1 or 2, it is characterised in that described step D is further comprising the steps of:

D: calculate the maximum of each pixel RGB triple channel gray value, obtain described original image luminance component.

5. the enhancement method of low-illumination image based on GPU as claimed in claim 1 or 2, it is characterised in that described step E is further comprising the steps of:

E1: after the non-linear modulation of described luminance component, brightness meets functional relationship with original brightness:

$I_1=I^{p_1+p_2\mathrm{\&theta;}+p_3{\mathrm{\&theta;}}^2}$

In formula, I is original luminance value, I₁For brightness value after non-linear modulation, p₁, p₂, p₃It is constant, and meets p₁+p₂+p₃=1, θ is Automatic adjusument parameter, and its value is determined according to the overall brightness of input picture, specifically can apply following formula:

$\mathrm{\&theta;}=\left\{\begin{array}{cc}0& I_{0.1N}<l_0\\ (I_{0.1N}-l_0)/l_2& l_0\&le;I_{0.1N}\&le;l_1\\ 1& I_{0.1N}>l_1\end{array}\right.$

Wherein, l₀, l₁, l₂For constant, span meet (0,255], I_0.1NRepresenting that the ascending sequence of brightness value to pixels all in image is in the brightness value of 0.1N position, N is then image slices vegetarian refreshments number.

6. the enhancement method of low-illumination image based on GPU as claimed in claim 1, it is characterised in that described step e201 is further comprising the steps of:

E2011: described contrast degeneration processes selects boxlike filtering BoxFilter, and filter radius is 2.

7. the enhancement method of low-illumination image based on GPU as claimed in claim 1, it is characterised in that described step e204 is further comprising the steps of:

E2041: described iteration stopping condition includes: before and after meeting, the difference of twice iteration result is not more than a certain constant, or arranges maximum iteration time, after iterations exceedes described maximum iteration time, also stops iteration.

8. the enhancement method of low-illumination image based on GPU as claimed in claim 7, it is characterised in that the described iteration stopping condition in described step e2041, comprises the steps:

E2042: by calculating the Euclidean distance of kth time iteration result and kth-1 result, it may be judged whether meet formula:And set maximum iteration time: Iter_max=10, whereinFor kth time iteration result,For-1 iteration result of kth.

9. the enhancement method of low-illumination image based on GPU as claimed in claim 1 or 2, it is characterised in that described step F, further comprising the steps of:

F1: the described data according to acquisition recover color of image, can adopt linear restoring method, and concrete formula is as follows:

R_i'=R_iI′_i/I_iI=1 ..., N

G_i'=G_iI′_i/I_iI=1 ..., N

B_i'=B_iI′_i/I_iI=1 ..., N

In formula, R_i,G_i,B_iFor the RGB triple channel gray value of pixel a certain in original image, I_iFor the brightness value of a certain pixel, R_i',G_i',B_i' and I_i' expression process after a certain pixel RGB triple channel gray value and this brightness, N represents pixel number；

F2: apply the whole pixels in described formula traversing graph picture.