Color band elimination method for dark light noise reduction image
By combining image linear enhancement and Atkinson dithering jitter algorithm in dark light environments, the problem of decreasing image ribbon phenomenon in dark light environments is solved, and a good balance between noise patterns and noise reduction effects is achieved.
Patent Information
- Application Number
- CN202311815858.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-27
- Publication Date
- 2025-06-27
AI Technical Summary
In dark light environments, the image after noise reduction is prone to ribbon phenomenon, and the prior art is difficult to maintain the balance between the noise form and the noise reduction effect when removing the ribbon.
Combining image linear enhancement and Atkinson dithering jitter algorithm, the image after noise reduction is first linearly enhanced to reduce the density of ribbons, and then the noise pattern is improved by quantizing error diffusion.
Effectively eliminate the ribbon phenomenon of noise-decreasing images in dark light environments, improve the noise form, and reduce the impact on the noise reduction effect in non-ribbon areas.
Smart Images
Figure CN120219210A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of image noise reduction, and particularly relates to a method for eliminating color bands in low-light noise-reduced images. Background Art
[0002] In the prior art, imaging electronic devices usually process and convert the original image and then output it for display. The color bit depth that a conventional display device can represent is limited, and the image needs to be quantized before being output to the display device, thereby reducing the bit depth. For imaging in low-light environments, temporal noise reduction and spatial noise reduction are essential steps. Among them, spatial denoising usually processes the image in the form of filtering. Since the noise in the image is uneven, some spatial denoising methods adopt local filtering. The spatial denoising model based on neural network adaptively performs different degrees of noise reduction on different intensities of noise in the image. In this case, between adjacent regions with and without filtering processing, or between adjacent regions with different degrees of denoising, it is usually a smooth gradient region. Due to the small color change and limited bit depth, the number of coding values used to quantify these small color differences is insufficient, resulting in the color gradient being reflected as color bands on the display device.
[0003] In low-light environments, especially in scenes containing light sources, the dynamic range of the image is higher. After noise reduction and low-light enhancement processing, color bands are more likely to appear in regions with a gentle brightness gradient, seriously affecting the image quality.
[0004] To solve the color band problem in the image, it is usually necessary to perform color band elimination processing on the image, such as using image dithering or adding noise and other means. Common image dithering algorithms include the Floyd-Steinberg dithering algorithm and its variant, the Atkinson dithering algorithm. Among them, the Floyd-Steinberg dithering algorithm uses error diffusion to achieve dithering. Its principle is to distribute the residual quantization error of the pixel to its adjacent pixels. The algorithm scans the image from left to right and from top to bottom, quantifying the pixel values one by one, transferring the quantization error of the current pixel to the adjacent pixels, and not affecting the already quantized pixels at the same time. The Atkinson dithering algorithm has the same principle as the Floyd-Steinberg dithering algorithm. The difference is that only 3 / 4 of the error diffuses outward, the dithering is more local, and the obtained image contrast is richer.
[0005] For low-light noise-reduced images, a part of the noise in the original image can be retained by reducing the spatial denoising coefficient, which is equivalent to adding a layer of noise to the noise-reduced image to avoid the generation of color bands in the smooth gradient region of the image. However, it will cause residual noise in the non-color band area, affecting the overall noise reduction effect of the image.
[0006] However, the deficiencies of the prior art are as follows:
[0007] 1. For low-light scenes, directly using the dithering algorithm on the denoised image will result in poor residual noise morphology in the color band area;
[0008] 2. Reducing the spatial domain denoising coefficient can effectively avoid the color band problem, but if the denoising coefficient is too small, the overall denoising effect of the image will deteriorate.
[0009] In addition, the commonly used terms in the prior art include:
[0010] Bit depth: When a computer records the color of a digital image, the number of bits occupied by each pixel is the bit depth. For example, for an image with a bit depth of 8 bits, each pixel occupies 8 bits and can represent 256 colors.
[0011] Quantization: Adjusting data with a large numerical range to data with a small numerical range. For example, adjusting an image with a 16-bit depth to an 8-bit depth.
[0012] Spatial domain denoising: A 2D denoising method that only processes the noise within a single frame of the image. According to different implementation principles, denoising algorithms can be divided into many types, such as linear / non-linear, spatial domain / frequency domain, and the frequency domain includes wavelet transform, Fourier transform, or other transforms.
[0013] Temporal domain denoising: A 3D denoising method whose main idea is to use the temporal correlation of multiple frames of images to achieve denoising.
[0014] Color band: A smooth gradient area in the image. Due to small color changes and limited bit depth, the number of coding values used to quantify these small color differences is insufficient, resulting in the color gradient being reflected as bands with color differences on the display device.
[0015] Atkinson dithering algorithm: A variant of the Floyd-Steinberg dithering algorithm. Its principle is to use error diffusion to achieve dithering. The characteristic is that only 3 / 4 of the error spreads outward, the dithering is more local, and the obtained image has richer contrast. Summary of the Invention
[0016] To solve the above problems, the purpose of this application is to: By combining image linear enhancement and the dithering algorithm, eliminate the color band phenomenon caused by quantization of the denoised image in low-light environments, make the residual noise have a good morphology, and reduce the impact on the denoising effect of non-color band areas.
[0017] Specifically, the present invention provides a method for eliminating color bands in low-light denoised images, and the method includes the following steps:
[0018] S1. Perform low-light image denoising processing and output a denoised image with the same bit depth as the RAW image;
[0019] S2. Perform linear enhancement on the denoised image to reduce the color band density; in low-light environments, the pixel values after image normalization are mostly concentrated in the range less than 0.1. Performing linear enhancement on it can amplify the color changes in the smooth gradient region and improve the color band density; it includes:
[0020] S2.1, Calculate the image mean and select the linear enhancement coefficient;
[0021] S2.2, Solve the linear enhancement coefficient of the current frame by moving average;
[0022] S2.3, Image linear enhancement;
[0023] S3. Quantization error diffusion; first calculate the quantization error of the denoised image, and then use the Atkinson dithering image dithering algorithm to diffuse the quantization error;
[0024] S4. Output the quantized denoised image: Quantize the image after error diffusion from the high bit depth d1 to the low bit depth d0, as shown in formula (7),
[0025]
[0026] Obtain the denoised image after processing.
[0027] The denoising in step S1 can be separate spatial domain denoising, or it can be first denoised in the time domain and then in the spatial domain. Regardless of the form of denoising, the output is a denoised image with the same bit depth as the RAW image.
[0028] Step S2 further includes:
[0029] S2.1, Calibrate the linear enhancement coefficient in different illuminance intervals and establish the correlation between the image brightness mean and the linear enhancement coefficient:
[0030] First, determine the lower limit of the illuminance of the imaging device's working environment, calibrate the required linear enhancement coefficient at this illuminance, and record the image brightness mean;
[0031] Then, increase the illuminance in sequence, calibrate the corresponding linear enhancement coefficient at each illuminance, and record the corresponding image brightness mean until there are no color bands in the image when the linear enhancement coefficient is 1;
[0032] Among them, the calibration method of the linear enhancement coefficient is as follows:
[0033] 3) Normalize the denoised image and calculate the 99th percentile, denoted as percentile_99;
[0034] 4) To ensure that no more than 99% of the pixels in the image are overexposed after linear enhancement, the method for solving the linear enhancement coefficient amplifier is shown in Equation (1):
[0035]
[0036] In the above formula, floor represents the floor operation;
[0037] After calibration, multiple groups of image average brightness intervals and corresponding linear enhancement coefficients can be obtained. Input a denoised image. This denoised image has the same bit depth as the RAW image and has not been quantized to a lower bit depth. Select the corresponding linear enhancement coefficient according to its brightness mean value;
[0038] S2.2, Perform a moving average on the linear enhancement coefficient to avoid image brightness jumps caused by frequent changes of the linear enhancement coefficient at the interval boundaries; In the storage device, cache the linear enhancement coefficient amplifier(t0) at the current time t0 and the amplifier(t0-1), amplifier(t0-2),..., amplifier(t0-N) at the previous N time moments. The calculation method of the amplifier_curr used in the current frame is as follows:
[0039]
[0040] S2.3, Perform linear enhancement on the denoised image; The denoised image is denoted as I dn , with a bit depth of d0 bits. After solving the linear enhancement coefficient amplifier through a moving average, enhance the image. The calculation method is as follows:
[0041]
[0042] In the above formula, clip represents the value range limiting operation. Let clip(x, a, b), then it represents limiting the value range of x to the interval (a, b).
[0043] In step S2.1, the lower illuminance limit is selected according to the requirements for the usage environment of the imaging device, usually 0.1 lux or less, including 0.01 lux, 0.0001 lux; The illuminance is increased sequentially. Based on the lower illuminance limit, the illuminance is increased sequentially, usually twice the lower illuminance limit. For example, if the lower illuminance limit is 0.001 lux, it is increased by 0.002 lux sequentially;
[0044] Step S3 further includes:
[0045] S3.1, Calculate the quantization error; Assume that the bit depth of the quantized image is d1. The calculation method of the quantization error quant_error is shown in Equation (4):
[0046]
[0047] In formula (4), round represents the rounding operation;
[0048] S3.2, Error diffusion; The Atkinson dithering algorithm is used to diffuse the quantization error, and all pixels in the image are traversed in the order from left to right and from top to bottom;
[0049] After completing the traversal and error diffusion of all pixels in the image in order, rounding and value range limitation are performed on the image.
[0050] For traversing all pixels in the image in the order from left to right and from top to bottom, assuming that the current pixel traversed in the current image is I(x, y), the quantization error is diffused to the 6 neighboring pixels of I(x, y+1), I(x, y+2), I(x+1, y-2), I(x+1, y-1), I(x+1, y) and I(x+2, y-1) with the same weight, and the calculation method is as follows:
[0051]
[0052]
[0053]
[0054]
[0055]
[0056]
[0057] For rounding and value range limitation of the image, as shown in formula (6),
[0058]
[0059] Therefore, the advantages of this application are:
[0060] 1. First performing linear enhancement on the denoised image can reduce the density of color bands, and then performing the image dithering algorithm can effectively improve the noise form remaining after color band removal;
[0061] 2. It has less impact on the denoising effect in the non-color band area of the image. BRIEF DESCRIPTION OF THE DRAWINGS
[0062] The drawings described herein are used to provide a further understanding of the present invention, form a part of this application, and do not constitute a limitation to the present invention.
[0063] Figure 1 It is a schematic diagram of the method flow of this application.
[0064] Figure 2 It is a schematic diagram of the Atkinson dithering algorithm in method step S3 of this application.
[0065] Figures 3(1) and 3(2) are schematic diagrams of the effect comparison between the original noise-reduced image and the noise-reduced image after algorithm processing. Among them, Figure 3(1) is the original noise-reduced image, and Figure 3(2) is the noise-reduced image after color band elimination. Specific implementation manner
[0066] In order to more clearly understand the technical content and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.
[0067] A method for eliminating color bands in a low-light noise-reduced image in this embodiment has an overall process as Figure 1 shown, and the method includes the following steps:
[0068] S1. Perform noise reduction processing on the low-light image, and output a noise-reduced image with the same bit depth as the RAW image;
[0069] S2. Perform linear enhancement on the noise-reduced image to reduce the color band density; in a low-light environment, the pixel values after image normalization are mostly concentrated in the interval less than 0.1. Performing a linear enhancement operation on it can amplify the color change in the smooth gradient area and improve the color band density; including:
[0070] S2.1, calculate the image mean value and select the linear enhancement coefficient;
[0071] S2.2, solve the linear enhancement coefficient of the current frame by moving average;
[0072] S2.3, perform image linear enhancement;
[0073] S3. Quantization error diffusion; first calculate the quantization error of the noise-reduced image, and then use the Atkinson dithering image dithering algorithm to diffuse the quantization error;
[0074] S4. Output the quantized noise-reduced image: Quantize the image after error diffusion from the high bit depth d1 to the low bit depth d0, as shown in formula (7),
[0075]
[0076] Obtain the noise-reduced image after processing.
[0077] Specifically, it includes the following steps:
[0078] Step S1. Denoising processing of low-light images; the denoising here can be separate spatial domain denoising, or it can be first denoised in the time domain and then in the spatial domain. Regardless of the form of denoising, the output is a denoised image with the same bit depth as the RAW image;
[0079] Step S2. Linearly enhance the denoised image to reduce the density of color bands; the main reasons for the generation of color bands are that the color change in the smooth gradient area is small and the bit depth is insufficient after image quantization. In a low-light environment, most pixels in the image are concentrated in a very small value range. Performing a linear enhancement operation on it can amplify the color change in the smooth gradient area and improve the density of color bands;
[0080] Step S2.1, Calibrate the linear enhancement coefficient in different illumination intervals and establish the correlation between the image brightness mean and the linear enhancement coefficient; first determine the lower limit of the illumination of the imaging device's working environment, such as 0.01 lux, calibrate the linear enhancement coefficient required at this illumination, and record the image brightness mean; then increase the illumination in turn, calibrate the corresponding linear enhancement coefficient, and record the corresponding image brightness mean until there are no color bands in the image when the linear enhancement coefficient is 1; the calibration method of the linear enhancement coefficient is as follows:
[0081] 5) Normalize the denoised image and calculate the 99th percentile, denoted as percentile_99;
[0082] 6) To ensure that 99% of the pixels in the image do not overexpose after linear enhancement, the method for solving the linear enhancement coefficient amplifier is shown in formula (1):
[0083]
[0084] In the above formula, floor represents the floor operation; after calibration, multiple groups of image average brightness intervals and corresponding linear enhancement coefficients can be obtained. Input a frame of denoised image and select the corresponding linear enhancement coefficient according to its brightness mean;
[0085] Step S2.2, Perform a moving average on the linear enhancement coefficient to avoid image brightness jumps caused by frequent changes of the linear enhancement coefficient at the interval boundaries; cache the current linear enhancement coefficient amplifier(t0) at time t0 and the historical N moments of amplifier(t0 - 1), amplifier(t0 - 2),..., amplifier(t0 - N) in the storage device. The calculation method of the amplifier_curr used for the current frame is as follows:
[0086]
[0087] Step S2.3, Perform linear enhancement on the denoised image; the denoised image is denoted as I dn, with a bit depth of d0 bits. After solving the linear enhancement coefficient amplifier through moving average, the image is enhanced, and the calculation method is as follows:
[0088]
[0089] In the above formula, clip represents the value range limiting operation. For example, clip(x, a, b) means limiting the value range of x to the interval (a, b);
[0090] Step S3. Quantization error diffusion; First, calculate the quantization error of the denoised image, and then use the Atkinson dithering image dithering algorithm to diffuse the quantization error;
[0091] Step S3.1, Calculate the quantization error; Assume that the bit depth of the quantized image is d1, and the calculation method of the quantization error quant_error is shown in formula (4):
[0092]
[0093] In formula (4), round represents the rounding operation;
[0094] Step S3.2, Error diffusion; Use the Atkinson dithering algorithm to diffuse the quantization error, and traverse all the pixels in the image in the order from left to right and top to bottom. As Figure 2 shown, the shaded area in the figure represents the current pixel I(x, y). The quantization error is diffused to the 6 neighboring pixels I(x, y + 1), I(x, y + 2), I(x + 1, y - 2), I(x + 1, y - 1), I(x + 1, y), and I(x + 2, y - 1) with the same weight, and the calculation method is as follows:
[0095]
[0096]
[0097]
[0098]
[0099]
[0100]
[0101] After completing the traversal and error diffusion of all the pixels in the image in order, round and limit the value range of the image, as shown in formula (6),
[0102]
[0103] Step S4. Quantize the image after error diffusion from the high bit depth d1 to the low bit depth d0, as shown in formula (7).
[0104]
[0105] The comparison of the effects between the original noise-reduced image and the noise-reduced image after algorithm processing is shown in FIGS. 3(1) and 3(2), where FIG. 3(1) is the original noise-reduced image and FIG. 3(2) is the noise-reduced image after color band elimination.
[0106] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, various changes and modifications can be made to the embodiments of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.
Claims
1. A method for eliminating color bands in low-light noise-reduced images, characterized in that, The method includes the following steps: S1. Perform noise reduction processing on the low-light image and output a noise-reduced image with the same bit depth as the RAW image; S2. Perform linear enhancement on the noise-reduced image to reduce the density of color bands; in a low-light environment, the pixel values after image normalization mostly concentrate in the interval less than 0.
1. Performing a linear enhancement operation on it can amplify the color change in the smooth gradient region and improve the density of color bands; it includes: S2.1, Calculate the image mean value and select the linear enhancement coefficient; S2.2, Solve the linear enhancement coefficient of the current frame by moving average; S2.3, Perform image linear enhancement; S3. Quantization error diffusion; first calculate the quantization error of the noise-reduced image, and then use the Atkinson dithering image dithering algorithm to diffuse the quantization error; S4. Output the quantized noise-reduced image: Quantize the image after error diffusion from the high bit depth d1 to the low bit depth d0, as shown in formula (7), to obtain the processed noise-reduced image.
2. The method for eliminating color bands in a low-light noise reduction image according to claim 1, characterized in that, The noise reduction in step S1 can be separate spatial domain denoising, or it can be first denoised in the time domain and then in the spatial domain. Regardless of the form of noise reduction, the output is a noise-reduced image with the same bit depth as the RAW image.
3. A method for eliminating color bands in a low-light noise reduction image according to claim 1, characterized in that, Step S2 further includes: S2.1, Calibrate the linear enhancement coefficient in different illumination intervals and establish the correlation between the image brightness mean value and the linear enhancement coefficient: First, determine the lower limit of the illumination of the imaging device working environment, calibrate the linear enhancement coefficient required at this illumination, and record the image brightness mean value; Then increase the illumination in turn, calibrate the corresponding linear enhancement coefficient at each illumination, and record the corresponding image brightness mean value until there are no color bands in the image when the linear enhancement coefficient is 1; Among them, the calibration method of the linear enhancement coefficient is as follows: 1) Normalize the noise-reduced image and calculate the 99th percentile, denoted as percentile_99; 2) To ensure that 99% of the pixels in the image do not overexpose after linear enhancement, the method for solving the linear enhancement coefficient amplifier is as shown in formula (1): In the above formula, floor represents the floor operation; After calibration, multiple groups of image average brightness intervals and corresponding linear enhancement coefficients can be obtained. Input a frame of noise-reduced image, which has the same bit depth as the RAW image and has not been quantized to a low bit depth. Select the corresponding linear enhancement coefficient according to its brightness mean value; S2.2, Perform a moving average on the linear enhancement coefficient to avoid image brightness jumps caused by frequent changes of the linear enhancement coefficient at the interval boundary; Cache the linear enhancement coefficient amplifier(t0) at the current t0 moment and the historical N moments of amplifier(t0-1), amplifier(t0-2),..., amplifier(t0-N) in the storage device. The calculation method of the amplifier_curr used for the current frame is as follows: S2.3, perform linear enhancement on the denoised image; the denoised image is denoted as I dn , with a bit depth of d0 bits. After solving the linear enhancement coefficient amplifier through moving average, the image is enhanced, and the calculation method is as follows: In the above formula, clip represents the value range limiting operation. Let clip(x, a, b), then it means to limit the value range of x to the interval (a, b).
4. A method for eliminating color bands in a low-light noise reduction image according to claim 3, characterized in that, In the step S2.1, the lower limit of illuminance is selected according to the requirements for the usage environment of the imaging device, usually 0.1 lux or below, including 0.01 lux and 0.0001 lux; the illuminance is increased sequentially. Based on the lower limit of illuminance, the illuminance is increased sequentially, usually twice the lower limit of illuminance. For example, if the lower limit of illuminance is 0.001 lux, it is increased by 0.002 lux sequentially.
5. A method for eliminating color bands in a low-light noise reduction image according to claim 3, characterized in that, The step S3 further includes: S3.1, calculating the quantization error; assuming that the bit depth of the quantized image is d1, the calculation method of the quantization error quant_error is shown in formula (4): In formula (4), round represents the rounding operation; S3.2, error diffusion; the Atkinson dithering algorithm is used to diffuse the quantization error, and all pixels in the image are traversed in the order from left to right and from top to bottom; After completing the traversal and error diffusion of all pixels in the image in order, rounding and value range limitation are performed on the image.
6. A method for eliminating color bands in a low-light noise reduction image according to claim 5, characterized in that For traversing all pixels in the image in the order from left to right and from top to bottom, assuming that the current pixel traversed in the current image is I(x, y), the quantization error is diffused to the six neighboring pixels I(x, y + 1), I(x, y + 2), I(x + 1, y - 2), I(x + 1, y - 1), I(x + 1, y) and I(x + 2, y - 1) with the same weight, and the calculation method is as follows: The rounding and value range limitation of the image are as shown in formula (6). In the above formula, I ed represents the image after error diffusion.