A multi-dimensional feature cross-fusion double-layer dynamic dimming and color fidelity method

Abstract

The application discloses a kind of multi-dimensional feature cross fusion's double-layer dynamic dimming and color fidelity method, it is related to liquid crystal display technical field, the method is converted to RGB image to HSV space and locks luminosity channel, is separated basic layer and detail layer by bilateral filter empty frequency decoupling;S curve and CLAHE double-track weighted fusion reconstruction are used to basic layer, and the gain matrix of three-dimensional cross fusion of physical edge mask, cognitive saliency mask and physiological visual mask is constructed to carry out spatial heterogeneity enhancement to detail layer;fusion reconstruction luminosity is output after soft knee cut-off, and according to the power function color protection compensation of saturation to luminosity amplification ratio, simultaneously solve dynamic backlight control factor adjustment LED drive flux.The application realizes the unity of contrast enhancement and color fidelity, effectively inhibits dark state noise, halo artifact and hue drift, and whole process is based on analytical expression calculation, adapts low algorithm power hardware real-time processing.

Description

Technical Field

[0001] This invention relates to the field of liquid crystal display technology, and in particular to a dual-layer dynamic dimming and color fidelity method that integrates multi-dimensional features. Background Technology

[0002] With the rapid development of Mini-LED backlight technology and multi-zone driving architecture, LCD displays have achieved significant improvements in peak brightness, contrast ratio, and overall power consumption by independently controlling the brightness of backlight areas. However, liquid crystal molecules themselves only act as passive photoelectric switches to modulate backlight flux and do not possess pixel-level self-emissive characteristics, thus exhibiting an inherent physical bottleneck of light leakage in dark states. When presenting dark scenes, strong backlight, or high dynamic range scenes, issues such as washed-out dark states, highlight overflow, and loss of detail in low-light conditions severely restrict the high dynamic range reconstruction quality of LCD displays.

[0003] Current technologies typically employ a collaborative control strategy combining regional dynamic backlight modulation and panel pixel compensation. This reduces power consumption by lowering the backlight drive current or duty cycle, while simultaneously performing reverse gain compensation on pixel grayscale at the panel end to improve visual brightness. However, this mechanism is affected by the diffusion effect of the LED dot spread function, allowing backlight energy to easily penetrate into adjacent dark pixels, resulting in severe edge halo artifacts. Furthermore, photon shot noise amplification, grayscale collapse, and color degradation in extremely low-light scenarios further reduce the subjective visual fidelity of the displayed image.

[0004] In current low-light enhancement and dynamic dimming algorithms, traditional global physical mapping operators are prone to causing hard truncation of highlights, swallowing of dark levels, and hue shift. While algorithms based on deep neural networks have strong fitting capabilities, their computational demands are enormous, failing to meet the hardware deployment requirements of display driver chips for microsecond-level real-time processing. Furthermore, existing detail enhancement algorithms often employ a global, indiscriminate brightening strategy, which can easily lead to excessive amplification of invalid background noise, making it impossible to achieve targeted enhancement of semantic subjects. Moreover, they are generally accompanied by problems such as saturation reduction and color distortion during brightness stretching.

[0005] Therefore, there is an urgent need to develop a dynamic dimming method that is low in complexity, high in image quality, and low in power consumption, which can improve the dynamic range of the display, effectively suppress halos and noise, ensure color fidelity, and adapt to the real-time deployment requirements at the chip level. Summary of the Invention

[0006] The purpose of this invention is to provide a two-layer dynamic dimming and color fidelity method that integrates multi-dimensional features. This method solves the problems of severe halo artifacts, malicious amplification of dark field noise, hard truncation of highlights, color saturation degradation, excessively high algorithm computing power making chip deployment difficult, and background redundancy gain caused by indiscriminate detail enhancement in current dynamic dimming technologies. This invention achieves the unity of high dynamic range reconstruction and high color fidelity.

[0007] To achieve the above objectives, this invention provides a two-layer dynamic dimming and color fidelity method based on multi-dimensional feature cross-fusion, comprising the following steps: Obtain the RGB color image of the image to be displayed; The RGB color image is converted to the HSV color space, and the hue channel, saturation channel and lightness channel are separated. The lightness channel is then locked as the only dimming channel. A bilateral filtering operator with spatial and value domain dual Gaussian weighting is used to decouple the lightness channel spatially and frequency-frequency, separating the low-frequency base layer that carries the macroscopic illuminance distribution and the high-frequency detail layer that carries the microscopic texture features. The reconstructed base layer is obtained by performing a dual-track nonlinear reconstruction on the low-frequency base layer. The dual-track nonlinear reconstruction includes: the first track uses an adaptive anchor point S-curve to stretch the global contrast, and the second track uses a limited contrast adaptive histogram equalization to wake up local dark details. A three-dimensional cross-joint gain matrix is ​​constructed to apply spatial heterogeneity gain to the high-frequency detail layer, resulting in an enhanced detail layer. The three-dimensional cross-joint gain matrix is ​​generated by fusing a physical edge mask representing the physical geometric boundary, a cognitive saliency mask representing the cognitive semantic weight, and a physiological visual mask representing the visual perception threshold. The reconstructed base layer and the enhanced detail layer are summed and merged pixel by pixel to obtain the reconstructed brightness channel; The reconstructed luminance channel is subjected to a nonlinear exponential decay piecewise soft knee mapping based on the McLaurin series expansion to obtain the output luminance components. Based on the pixel-level magnification ratio before and after the brightness channel enhancement, adaptive power function color compensation is applied to the original saturation channel to obtain the compensated saturation components. The dynamic backlight control factor is calculated based on the global statistical mean of the reconstructed brightness channel, and the physical driving flux of the LED backlight is adaptively adjusted according to the average brightness of the screen. Convert the hue channel, compensated saturation component, and output brightness component back to the RGB color space to output a color image after dimming compensation.

[0008] Preferably, the joint weight matrix of the bilateral filter operator is composed of the spatial domain Gaussian distance weights and the value domain Gaussian brightness weights, as expressed below: in, For spatial domain Gaussian distance weights, The horizontal coordinate of the currently processed pixel. The spatial ordinate of the currently processed pixel. m The x-coordinate of the sliding pixel within the neighborhood. nThe vertical coordinate of the sliding pixel within the neighborhood. For spatial standard deviation parameters, The value range is a Gaussian brightness weight. For the brightness channel in Pixel value at that location, For the brightness channel in Pixel value at that location, The parameter is the standard deviation of the range. This is the joint weight matrix; The expression for the low-frequency base layer is: in, As a low-frequency base layer, For the current pixel being processed during bilateral filtering ( i , j ) and sliding pixels in the neighborhood ( m , n The joint weights between ) The expression for the high-frequency detail layer is: in, This is a high-frequency detail layer.

[0009] Preferably, by performing dual-track nonlinear reconstruction on the low-frequency foundation layer, the specific content of the reconstructed foundation layer includes: The first reconstruction component is obtained by globally stretching the low-frequency base layer using an adaptive anchor point S-curve. A limiting contrast adaptive histogram equalization is used to locally awaken dark details in the first reconstructed component to obtain the second reconstructed component; The first and second reconstruction components are weighted and fused to obtain the reconstruction base layer.

[0010] Preferably, the first reconstructed component is obtained by using an adaptive anchor point S-curve to perform global contrast stretching on the low-frequency base layer, including: The pixel grayscale values ​​of the low-frequency base layer are sorted in ascending order. The 25% and 50% quantiles of the sorted sequence are extracted, and an empirical convergence factor is introduced to calculate the stretching anchor point. The calculation formula is as follows: in, For adaptive photometric anchor points, At the 25th percentile, 50% quantile; Based on the stretching anchor points, an S-curve mapping function is constructed, and nonlinear contrast stretching mapping is performed on the low-frequency base layer to obtain the first reconstructed component. The expression for the S-curve mapping function is: in, This is the illumination gain multiplier mapping matrix. The horizontal coordinate of the currently processed pixel. The spatial ordinate of the currently processed pixel. To control the slope factor of the tensile tension curve. As a low-frequency base layer, To prevent the minimum constant with a denominator of zero, This is the first reconstructed component.

[0011] Preferably, the first reconstructed component is locally reconstructed to enhance shadow detail by employing amplitude-limited contrast adaptive histogram equalization, resulting in the following specific content for the second reconstructed component: The gray-level histogram of the first reconstructed component is statistically analyzed, and a limiting factor is preset to calculate the global peak-shaving threshold. The peak values ​​of the histogram that exceed the peak-shaving threshold are truncated, the total overflow is calculated and evenly distributed to each gray level, and the histogram after peak-shaving and valley-filling is obtained. The cumulative distribution function is calculated based on the histogram after peak shaving and valley filling, and gray-level remapping is performed to obtain the second reconstructed component.

[0012] Preferably, the expression for reconstructing the base layer is: in, To reconstruct the base layer, The horizontal coordinate of the currently processed pixel. The spatial ordinate of the currently processed pixel. The weighting coefficients for the first reconstructed component. , For the first reconstructed component For the second reconstructed component Preferably, the three-dimensional cross-joint gain matrix is ​​generated by fusing a physical edge mask representing the physical geometric boundary, a cognitive saliency mask representing the cognitive semantic weight, and a physiological visual mask representing the visual perception threshold in the following manner: The Sobel operator is used to perform horizontal and vertical convolution operations on the original brightness channels, calculate the gradient magnitude and normalize it to the [0,1] interval to obtain the physical edge mask; The RGB color image is converted to the CIELAB perceptual uniform color space. Color quantization and dimensionality reduction are performed based on global histogram statistics. The weighted sum of the three-dimensional color difference distance between each representative color level and all pixels in the image is calculated. The calculation results are then mapped back to pixel coordinates and normalized to obtain a cognitive saliency mask. Constructing a physiological visual mask based on the Weber-Fechner law; The physical edge mask, cognitive saliency mask, and physiological visual mask are spatially cross-fused to generate a three-dimensional cross-joint gain matrix. Preferably, the expression for the physiological visual mask is: in, For physiological visual mask weights, For the brightness channel in Pixel value at that location, The horizontal coordinate of the currently processed pixel. represents the spatial ordinate of the currently processed pixel, 255 is the physical maximum constant of the luminance channel, and 0.35 is the preset minimum value to prevent loss of detail in extreme brightness areas.

[0013] Preferably, the reconstructed luminance channel is subjected to a nonlinear exponential decay piecewise soft-knee mapping process based on the Maclaurin series expansion, resulting in the following specific contents of the output luminance components: The reconstructed brightness is normalized to the [0,1] interval to obtain the normalized reconstructed brightness; A preset transition threshold is set. When the normalized reconstructed brightness is less than or equal to the preset transition threshold, the output brightness component is the normalized reconstructed brightness. When the normalized reconstructed brightness is greater than the preset transition threshold, the expression for the output brightness component is: ; in, To output the lightness component, The horizontal coordinate of the currently processed pixel. The spatial ordinate of the currently processed pixel. For the currently processed pixel ( i , j The reconstructed brightness value at point ) is the brightness value after normalization. To preset the turning threshold, .

[0014] Preferably, based on the pixel-level magnification ratio before and after the brightness channel enhancement, adaptive power function color compensation is performed on the original saturation channel to obtain the specific content of the compensated saturation components, including: The saturation gain factor is determined based on the pixel-level ratio of the enhanced pre-brightness channel to the output brightness channel. The initial compensated saturation is obtained by performing a product-type compensation on the original saturation based on the saturation gain factor. An amplitude constraint is applied to the initial compensated saturation to obtain the compensated saturation component.

[0015] In summary, the dual-layer dynamic dimming and color fidelity method based on multi-dimensional feature cross-fusion provided by this invention has the following advantages compared to traditional technologies: (1) By strictly locking the hue channel unchanged in the HSV color space and performing nonlinear mapping only in the lightness channel, the pollution path of illuminance reconstruction to spectral information is completely cut off from the underlying algebraic logic. Combined with the saturation adaptive compensation mechanism, the hue drift artifacts derived from traditional RGB linkage stretching are effectively suppressed, making the color hue retention rate close to the ideal value and greatly improving the color reproduction accuracy. (2) The spatial frequency decoupling is achieved by using a bilateral filtering operator. The gray-scale cliff blocking mechanism based on the value range weight is used to cut off the energy fusion across the object boundary, thereby reducing the edge halo artifacts that are common in traditional dimming algorithms from the source and ensuring the sharpness of the high-frequency contour. (3) By constructing a three-dimensional cross-gain matrix of physical edge, cognitive saliency and physiological vision, the detail enhancement is upgraded from global indiscriminate brightening to semantic subject-oriented wake-up. While strengthening the texture of the core area, it effectively suppresses the amplification of the background noise in the darkest area and improves the signal-to-noise ratio of the picture. (4) The low-frequency base layer adopts a dual-track nonlinear reconstruction of global contrast stretching using an adaptive anchor point S-curve and local dark detail awakening using a limited contrast adaptive histogram equalization, which takes into account both macroscopic large-span contrast and microscopic dark detail performance. (5) Nonlinear exponential decay piecewise soft knee mapping solves the problem of hard truncation of highlights and preserves the tonal transition details of the highlight area.

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a two-layer dynamic dimming and color fidelity method for cross-fusion of multi-dimensional features in this invention. Figure 2 This is a qualitative visual comparison of different saliency feature extraction algorithms in a typical Kodim 13 scenario in this invention: Figure 2 (a) in the figure is the original image; Figure 2 (b) in the figure is a qualitative visual image of the GBVS algorithm; Figure 2 (c) in the figure is the qualitative visual image of the AC algorithm; Figure 2 In the figure, (d) is the qualitative visual image of the FT algorithm; Figure 2 (e) in the figure represents the qualitative visual representation of the LC algorithm. Figure 2 (f) in the figure represents the qualitative visual image of the HC algorithm; Figure 3This is a comparison chart of the mean values ​​of the five saliency extraction algorithms (GBVS algorithm, AC algorithm, FT algorithm, LC algorithm and HC algorithm) in this invention under objective evaluation indicators; Figure 3 (a) in the figure is the average gradient (GMG) value plot; Figure 3 (b) in the image shows the entropy values ​​of the mask information. Figure 3 (c) in the figure represents the global spatial frequency (SF) value map; Figure 3 (d) in the figure represents the global standard deviation (STD) values. Figure 3 (e) in the figure is the Enhancement Metrics (EME) plot; Figure 4 This is a visual comparison diagram of the kodim05 decoupled multi-feature perception scale and high-frequency detail gain matrix in this invention; Figure 4 (a) in the figure is the visual representation of the target saliency prior mask β; Figure 4 (b) in the diagram represents the physical edge mask. W edge Visual image; Figure 4 (c) in the diagram represents the physiological visual mask. W hvs Visual image; Figure 4 (d) in the figure represents the detail gain matrix. G matrix Visual image; Figure 5 This is the normalized radar image of multidimensional features in this invention; Figure 6 This is a spatial distribution curve of the brightness histogram before and after dimming enhancement in a typical low-contrast scene of Kodim15 in this invention; Figure 7 This is a dynamic distribution diagram of the broken-line game between the dual Y-axis S-curve method and the CLAHE method in this invention; Figure 8 This is a subjective visual comparison of the Kodim01 typical low-light scene reconstruction and dimming algorithm in this invention; Figure 8 (a) in the figure is the original image; Figure 8 (b) in the figure is the graph after processing by the S-curve algorithm; Figure 8 (c) in the figure is the graph after processing by the algorithm in the literature; Figure 8 (d) in the figure represents the image processed by the algorithm of this invention; Figure 9 This is a comprehensive quantitative comparison chart of objective evaluation indicators for different enhancement algorithms in this invention; Figure 9 (a) in the image is a comparison chart of GMG. Figure 9 (b) in the figure is an Entropy comparison chart; Figure 9 (c) in the figure is the SF comparison chart; Figure 9 (d) in the figure is a comparison chart of STD; Figure 9(e) in the figure is an EME comparison chart; Figure 9 (f) in the figure is a CHPR comparison chart; Figure 10 This is a discrete fluctuation trend diagram of the six main robustness indicators of the dimming reconstruction algorithm in this invention; Figure 10 (a) in the figure is the discrete fluctuation trend chart of the GMG indicator; Figure 10 (b) in the figure is a discrete fluctuation trend chart of the Entropy indicator; Figure 10 (c) in the figure is the discrete fluctuation trend chart of the SF indicator; Figure 10 (d) in the figure represents the discrete fluctuation trend of the STD indicator; Figure 10 (e) in the figure represents the discrete fluctuation trend of the EME indicator; Figure 10 (f) in the figure represents the discrete fluctuation trend of the CHPR indicator. Detailed Implementation

[0018] The technical method of the present invention will be further described below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application.

[0019] The following description of at least one exemplary embodiment is merely illustrative and is not intended to limit the scope of this application or its application or use.

[0020] Techniques, systems, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, they should be considered part of the instruction manual.

[0021] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0022] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0023] To overcome the problems of microscopic geometric low-pass blurring, local extreme tone hard truncation, and accompanying chromatographic degradation in traditional dynamic dimming and enhancement operators when dealing with complex dark fields or extreme high-tolerance conflicting light and shadow fields, this invention provides a two-layer dynamic dimming reconstruction architecture that combines multi-dimensional feature cross-sensory scales. It severs the energy coupling between physical illuminance components and high-frequency texture structures at the algebraic level, and simultaneously reshapes macroscopic large-span contrast while relying on a closed-loop compensation mechanism to rigorously determine the topological fidelity of the source color space.

[0024] This invention provides a two-layer dynamic dimming and color fidelity method based on the cross-fusion of multi-dimensional features, such as... Figure 1 As shown, it includes the following steps: S1. Obtain the RGB color image of the image to be displayed.

[0025] S2. Convert the RGB color image to the HSV color space, separate the hue channel, saturation channel and lightness channel, and lock the lightness channel as the only dimming channel.

[0026] S3. A bilateral filter operator with spatial and value domain Gaussian weighting is used to decouple the lightness channel spatially and frequency-frequency, separating the low-frequency base layer carrying the macroscopic illuminance distribution and the high-frequency detail layer carrying the microscopic texture features. The joint weight matrix of the bilateral filter operator is composed of the product of the spatial domain Gaussian distance weight and the value domain Gaussian brightness weight, as expressed below: in, For spatial domain Gaussian distance weights, The horizontal coordinate of the currently processed pixel. The spatial ordinate of the currently processed pixel. m The x-coordinate of the sliding pixel within the neighborhood. n The vertical coordinate of the sliding pixel within the neighborhood. For spatial standard deviation parameters, The value range is a Gaussian brightness weight. For the brightness channel in Pixel value at that location, For the brightness channel in Pixel value at that location, The parameter is the standard deviation of the range. This is the joint weight matrix.

[0027] The expression for the low-frequency base layer is: in, As a low-frequency base layer, For the current pixel being processed during bilateral filtering ( i , j ) and sliding pixels in the neighborhood ( m , n The joint weights among the sliding pixels (determined by the spatial domain Gaussian distance weight and the value domain Gaussian brightness weight) are used to measure the contribution of the sliding pixels in the neighborhood to the filtering result of the current pixel; the closer the spatial distance and the more similar the brightness, the better. The larger the value, the stronger the influence of the sliding pixels within that neighborhood on the low-frequency base layer computation, while the denominator contains... This is used to normalize all weights, preventing the overall output brightness from being too high or too low due to different weight sums.

[0028] The expression for the high-frequency detail layer is: in, This is a high-frequency detail layer.

[0029] After completing the spatial basis transformation, this invention constructs a mathematical structure of spatial-frequency decoupling and hierarchical reconstruction, focusing on the independent brightness channel (V channel) carrying the global radiation intensity for refined reconstruction. The brightness channel of a natural field image is essentially a nonlinear coupling of a macroscopic low-frequency illumination prior field and a microscopic high-frequency texture perturbation term. Performing a globally indiscriminate mapping in the frequency domain would inevitably induce malignant divergence of invisible noise within smooth regions. Therefore, the algorithm introduces a nonlinear bilateral filter with dual Gaussian weighting in both the spatial and value domains to achieve precise tearing of the frequency band.

[0030] The bilateral filter overcomes the edge blurring defect of traditional Gaussian filtering by nesting a range difference kernel on top of a spatial attenuation kernel. For each center pixel of brightness V ( i , j The algorithm performs double Gaussian weighting within the sliding window Ω.

[0031] In an exemplary embodiment of the present invention, firstly, the Gaussian distance weight in the spatial domain is calculated. Let the coordinates of the neighboring pixels be ( m , n ), fixed spatial standard deviation parameter σ s =2, and its calculation formula is: .

[0032] Secondly, a Gaussian brightness weight is calculated for the edge-preserving core using bilateral filtering. When neighboring pixels cross a strong edge (i.e., a large brightness difference), this weight rapidly decays to zero, thus preventing the edge from being smoothed. A fixed standard deviation parameter is used. σ r = 0.1 × 255 = 25.5, the calculation formula is: .

[0033] Multiply the spatial domain weights by the range weights and use the normalization coefficients. W ( i , j , m , n By weighting the data, high-frequency impurities and smoothing the low-frequency global illumination topology (i.e., the low-frequency base layer) can be successfully extracted from the original image. Vb ): .

[0034] After obtaining the smoothed base layer, the zero-mean residual matrix calculated by subtracting the source brightness is the high-frequency detail layer. V d ).

[0035] .

[0036] S4. By performing dual-track nonlinear reconstruction on the low-frequency base layer, a reconstructed base layer is obtained. The dual-track nonlinear reconstruction includes: the first track uses an adaptive anchor point S-curve for global contrast stretching, and the second track uses amplitude-limited contrast adaptive histogram equalization to awaken local dark details.

[0037] For low-frequency base layers, if their brightness dynamic range exhibits significant compression, it will lead to a subjective perception of low visibility and a hazy, grayish appearance in the image. To reconstruct a deep visual perception manifold, this invention employs cascaded nonlinear adaptive mapping and statistical histogram probability convergence.

[0038] Traditional S-curve operators are often limited by a fixed center anchor point (μ≡128) as the stretching center, which can easily cause large-area highlight overflow when dealing with extremely dark or extremely bright scenes. To address this, an adaptive anchor point positioning strategy was adopted.

[0039] Furthermore, step S4 specifically includes the following steps: S401. The low-frequency base layer is subjected to global contrast stretching using an adaptive anchor point S-curve to obtain the first reconstructed component. Specifically, step S401 includes: The pixel grayscale values ​​of the low-frequency base layer are sorted in ascending order. The 25% and 50% quantiles of the sorted sequence are extracted, and an empirical convergence factor is introduced to calculate the stretching anchor point. The calculation formula is as follows: ; in, For adaptive photometric anchor points, At the 25th percentile, The 50% quantile is used. The algebraic consideration of dividing the formula by 3 is to force the stretching center down towards the low grayscale manifold, physically reserving a wide anti-overexposure redundancy band for the highlight area.

[0040] An S-curve mapping function is constructed based on the stretching anchor points. Nonlinear contrast stretching mapping is then applied to the low-frequency base layer to obtain the first reconstructed component. The expression for the S-curve mapping function is as follows: in, This is the illumination gain multiplier mapping matrix. The horizontal coordinate of the currently processed pixel. The spatial ordinate of the currently processed pixel. To control the slope factor of the tensile tension curve. As a low-frequency base layer, To prevent the minimum constant with a denominator of zero, This is the first reconstructed component.

[0041] While the S-curve significantly enhances the contrast between light and dark areas on a macroscopic scale, it can easily produce flat tonal merging in localized, microscopic dark regions, preventing details in these areas from emerging. Therefore, the algorithm uses... A second-level concatenated governance process is performed on the base map: restricted histogram equalization. Traditional histogram equalization is prone to generating background noise. In this invention, the operator limits the peak probability density by dynamically truncating a threshold.

[0042] S402. A limiting contrast adaptive histogram equalization is used to locally awaken dark details in the first reconstructed component, resulting in the second reconstructed component. Specifically, step S402 includes: The grayscale histogram of the first reconstructed component is statistically analyzed, and a limiting factor is preset to calculate the global peak-shaving threshold.

[0043] The peak values ​​of the histogram that exceed the peak-shaving threshold are truncated, the total overflow is calculated and evenly distributed to each gray level, and the histogram after peak-shaving and valley-filling is obtained.

[0044] The cumulative distribution function is calculated based on the histogram after peak shaving and valley filling, and gray-level remapping is performed to obtain the second reconstructed component.

[0045] In an exemplary embodiment of the present invention, the histogram probability count of gray level k of the first reconstructed component in the entire image is H(k). A limiting factor τ = 2.5 is set, and the global peak-shaving threshold is dynamically calculated. Limit : ; Where mean(·) is the global arithmetic mean operator for histogram counting.

[0046] Then all excess Limit Peak values ​​are truncated, the total overflow is calculated, and then evenly distributed across 256 gray levels: ; in, This is the reconstructed histogram after restriction truncation and amortization compensation. This represents the total amount of pixels that overflow beyond the limit threshold.

[0047] based on H clippedCalculate the cumulative distribution function remapping to obtain the cascaded result. V b_che .

[0048] Although V b_che It possesses extremely strong local contrast, but is prone to introducing tonal abrupt changes. Therefore, the algorithm fuses the ontology with the cascaded amplification ontology at the end, i.e.: S403. Weighted fusion of the first and second reconstruction components yields the reconstruction base layer. The expression for the reconstruction base layer is: ; in, To reconstruct the base layer, The horizontal coordinate of the currently processed pixel. The spatial ordinate of the currently processed pixel. The weighting coefficients for the first reconstructed component. , For the first reconstructed component For the second reconstructed component In an exemplary embodiment of the present invention, It is 0.55. With a weighting of 0.45, the above weighting coefficients ensure that the global physical naturalness of the S-curve dominates the threshold, while absorbing the extreme discovery capability of the histogram algorithm for microscopic dark field tonal values, thus reconstructing a high-quality lighting substrate.

[0049] In the high-frequency detail layer, not all high-frequency residuals are effective textures. Random additive noise in dark corners, if amplified, will lead to image pollution. To address this, this invention constructs a cross-gain matrix composed of three feature scales to implement highly targeted attention energy compensation.

[0050] S5. A three-dimensional cross-joint gain matrix is ​​constructed to apply spatial heterogeneity gain to the high-frequency detail layer, resulting in the enhanced detail layer. The three-dimensional cross-joint gain matrix is ​​generated by fusing a physical edge mask representing the physical geometric boundary, a cognitive saliency mask representing the cognitive semantic weight, and a physiological visual mask representing the visual perception threshold. The specific method of fusion generation is as follows: The Sobel operator is used to perform horizontal and vertical convolution operations on the original brightness channel, calculate the gradient magnitude, and normalize it to the [0,1] interval to obtain the physical edge mask. To prevent artifact magnification in smooth areas, this invention introduces a physical edge mask. This mask accurately anchors the true physical contours and sharp boundaries of the image, ensuring that subsequent gain energy is focused only on the true structural lines. Its formula is: in, Spatial coordinates ( i , j The absolute physical gradient magnitude at () The final physical edge mask weights mapped to the [0,1] interval. This is the global maximum gradient magnitude constant for the entire graph.

[0051] The RGB color image is converted to the CIELAB perceptually uniform color space. Color quantization and dimensionality reduction are performed based on global histogram statistics. The weighted sum of the three-dimensional color difference distances between each representative color level and all pixels in the image is calculated. The calculation results are then mapped back to pixel coordinates and normalized to obtain a cognitive saliency mask. This invention introduces a cognitive saliency mask, which calculates the sum of the three-dimensional color distances of each pixel in the Lab color space with other pixels in the image and normalizes it to [0,1]. This method keenly captures high-contrast main targets in the image, guides visual attention, and further eliminates the interference of invalid background redundancy on detail magnification. The formula is: ; ; in, The original saliency baseline value for the target pixel. u and v These are the index scalars for the target pixel and the traversed pixels, respectively. , , These are the lightness and chromaticity coordinate components of the target pixel in the CIELAB space. , , To iterate over the corresponding coordinate components of the pixels, The final cognitive saliency mask weights are mapped to the [0,1] interval.

[0052] A physiological visual mask is constructed based on the Weber-Fechner law. The expression for the physiological visual mask is: in, For physiological visual mask weights, For the brightness channel in Pixel value at that location, The horizontal coordinate of the currently processed pixel. represents the spatial ordinate of the currently processed pixel, 255 is the physical maximum constant of the luminance channel, and 0.35 is the preset minimum value to prevent loss of detail in extreme brightness areas.

[0053] The physiological visual mask assigns the highest weight to mid-gray areas, while forcibly converging to a minimum value of 0.35 in extremely dark blind spots. This design completely blocks the surge of useless noise in extremely dark blind spots from a physical level.

[0054] The physical edge mask, cognitive saliency mask, and physiological visual mask are spatially cross-fused to generate a three-dimensional cross-joint gain matrix. In an exemplary embodiment of the present invention, the final gain matrix is ​​synthesized by combining three feature masks and through spatial modulation of addition and multiplication. Where 1.2 represents the zero-base protection factor, and 1.4 and 2.8 represent the directional excitation coefficients: .

[0055] Based on the above formulas, the final formula for the high-frequency detail layer is derived as follows: .

[0056] S6. The reconstructed base layer and the enhanced detail layer are summed and merged pixel by pixel to obtain the reconstructed brightness channel. After the base layer and detail layer have been fully unified and converged, to take into account the physical power consumption characteristics of the terminal display device, a brightness average based on the entire image is introduced. μ v Dynamic backlight control. Through non-linear adjustment of the backlight factor, the physical power consumption of the LCD backlight panel can be adaptively reduced according to the average brightness of the screen. The specific formula is as follows: in, The dynamic backlight power-saving constraint factor is 0.72, which is the lower limit of the backlight factor. The elastic threshold coefficient for dynamic adjustment is 0.18. (Both of these factors strictly lock the global dimming constraint within the physical safety range of [0.72, 0.90], reserving a 10% defense against highlight overflow caused by detail superposition). This refers to the reconstructed brightness component after normalization.

[0057] S7. Apply nonlinear exponential decay piecewise soft knee mapping based on McLaurin series expansion to the reconstructed luminance channel to obtain the output luminance component.

[0058] Furthermore, step S7 specifically includes the following steps: S701. Normalize the reconstructed brightness to the [0,1] interval to obtain the normalized reconstructed brightness.

[0059] S702, Preset transition threshold: When the normalized reconstructed brightness is less than or equal to the preset transition threshold, the output brightness component is the normalized reconstructed brightness.

[0060] S703. When the normalized reconstructed brightness is greater than the preset transition threshold, the expression for the output brightness component is: ; in, To output the lightness component, The horizontal coordinate of the currently processed pixel. The spatial ordinate of the currently processed pixel. To preset the turning threshold, , For the currently processed pixel ( i , j The reconstructed brightness value at point ) is normalized and serves as the input parameter for subsequent segmented soft knee mapping. It is used to characterize the current brightness level of the pixel and determine whether it exceeds the preset highlight transition threshold. If it does not exceed the threshold, the original value is maintained. If it exceeds the threshold, nonlinear soft knee compression is entered to suppress hard truncation and overexposure in the highlight area.

[0061] In an exemplary embodiment of the present invention, after experiencing the gain of the low-frequency base layer and the high-frequency detail layer, the reconstructed... V norm A very small number of pixels in the highlight area (such as the sky and incandescent lamps) are prone to exceeding the physical limit of 1.0, resulting in saturation hard truncation artifacts and causing the highlight areas to appear white. To address this, this invention introduces a soft knee protection mechanism based on the McLaurin series, setting a preset turning threshold of 0.88: .

[0062] When pixel brightness exceeds a threshold of 0.88, the linear truncation abrupt change is forcibly transformed into a smooth exponential asymptote. As brightness increases infinitely... V out It smoothly and losslessly approaches the physical upper limit of 1.0 with continuous first derivatives. This mechanism completely defends against pixel tone collapse at the underlying topology level, preserving cloud and micro-texture transitions in extremely bright areas.

[0063] S8. Based on the pixel-level magnification ratio before and after the brightness channel enhancement, perform adaptive power function color compensation on the original saturation channel to obtain the compensated saturation components.

[0064] Furthermore, step S8 specifically includes the following steps: S801. Determine the saturation gain factor based on the pixel-level ratio of the enhanced brightness channel to the output brightness channel.

[0065] S802. Perform product-type compensation on the original saturation based on the saturation gain factor to obtain the initial compensated saturation.

[0066] S803. Apply amplitude constraints to the initial compensated saturation to obtain the compensated saturation components.

[0067] S9. Calculate the dynamic backlight control factor based on the global statistical mean of the reconstructed brightness channel, and adaptively adjust the physical driving flux of the LED backlight panel according to the average brightness of the screen.

[0068] S10. Convert the hue channel, the compensated saturation component, and the output brightness component back to the RGB color space, and output the color image after dimming compensation.

[0069] In the HSV color space, drastically increasing only the luminance channel can lead to visual desaturation or bleaching of the image. This is because the luminance vector is stretched while the saturation fails to compensate accordingly. To achieve a closed-loop unity between extremely high contrast reconstruction and high color fidelity, this invention designs an adaptive saturation color protection compensator, where... For experience-based color-protecting factors, ε To prevent the denominator from being zero, the min operator strictly prevents physical out-of-bounds errors. ; in, This is to compensate for the saturation component.

[0070] Ultimately, the system will fix the original hue H and the compensated saturation. S out With reconstructed brightness V out By performing inverse color reconstruction and outputting true color images through the nonlinear inverse mapping matrix HSV2RGB, the problem of derivative color shift that is easily induced by dynamic dimming is completely solved.

[0071] To objectively and rigorously demonstrate the mathematical validity and engineering robustness of the multi-dimensional feature cross-fusion dual-layer dynamic dimming and color fidelity method of this invention, the discretization verification platform of the core algorithm is uniformly deployed in the MATLAB 2022b runtime environment. To verify the underlying mapping logic of the algorithm, all core feature operators covered in this patent architecture (including color space homeomorphism mapping, bilateral edge-preserving frequency division, nonlinear high-dimensional feature extraction, and histogram probability reconstruction, etc.) are strictly based on the underlying continuous discrete mathematical formulas and independently programmed and solidified.

[0072] The experimental benchmark uses the internationally recognized high-fidelity image quality evaluation standard – the Kodak24 dataset (comprising 24 uncompressed high-resolution natural scene color tensor matrices from kodim01 to kodim24). This dataset's spatial manifold extensively interweaves extremely dark physical blind zones, extremely wide dynamic lighting conflicts, large areas of smooth isomorphic regions, complex high-frequency microtextures, and portraits, among other demanding optical and physical scenes, making it sufficient to conduct extreme stress tests on the multidimensional generalization performance of dynamic dimming and illumination reconstruction algorithms.

[0073] The quantitative experimental system rigorously incorporates five evaluation metrics: mean gradient magnitude (GMG), entropy, spatial frequency (SF), standard deviation (STD), and enhancement metric (EME). These five metrics enable self-consistent and rigorous mathematical evaluation, establishing a fundamental scientific basis for the selection of optimal spatial cross-masks.

[0074] The spatial distribution accuracy and structural fidelity of the saliency mask play a pivotal role in the dual-layer dynamic dimming topology of this invention. Its extraction quality directly determines whether the subsequent detail gain matrix, when releasing high-frequency compensating energy, will cause abnormal divergence of background noise or halo spillover at structural boundaries. To rigorously examine the spatial feature decoupling and semantic focusing performance of various comparative algorithms (traditional GBVS algorithm, FT algorithm, AC algorithm, LC algorithm, and this invention) in dealing with unstructured, complex textures and high-frequency ambient light interference, this invention extracts a natural landscape node (kodim13) encompassing high-frequency micro-textures for qualitative visual analysis. The scene's field of view is interwoven with dynamic water waves in the foreground, high-frequency gravel beaches, and high-density dark forests and distant mountains in the background; due to the lack of a high-contrast macroscopic subject throughout, its interior is filled with false edge interference, constituting a limit stress test on the robustness of the saliency operator against high-frequency disturbances and multi-scale extraction, such as... Figure 2 As shown.

[0075] according to Figure 2 (b) and Figure 2 As shown in (d) of the visual mapping results, traditional GBVS and FT algorithms heavily rely on large-scale spatial Gaussian smoothing kernels and multi-scale pooling mechanisms in their underlying logic. While this low-pass filtering property smooths out some microscopic noise, it inevitably induces a precipitous degradation in spatial resolution. Visually, the physical contours of the foreground shrubs and gravel exhibit spatial energy leakage, becoming severely topologically bound to the dark forest in the background. If such low-pass blurring masks are cascaded to subsequent cross-modulation, they will inevitably induce blind divergence of gain weights. According to Figure 2 In (c), when dealing with highly heterogeneous manifolds in the background, the AC algorithm's local sampling window fails to effectively converge to a globally consistent prior, causing the saliency response to encounter extreme value decay, and the mask almost degenerates into a completely dark state. According to Figure 2 In (e), although the LC algorithm relies on the global brightness histogram to establish a contrast basis, its algebraic design completely removes the spatial distance prior. Visually, the LC operator suffers from severe semantic inversion, incorrectly activating low-frequency, high-brightness backgrounds such as distant clouds and the sky. This inverse mapping not only deviates from true visual attention but also misleads subsequent networks into performing ineffective high-frequency sharpening on distant scenes. Figure 2Compared to the rigid limitations of the aforementioned mechanisms, the present invention (HC algorithm) exhibits overwhelming feature isolation and edge topology fidelity. It abandons resolution-degrading spatial smoothing, directly calculating the three-dimensional Euclidean chromatic aberration distance of all pixels in the Lab perceptual uniform space, and performs dimensionality reduction through histogram-based color-gradient clustering. This underlying mathematical architecture enables it to extremely accurately anchor the spectral gradient transition between the foreground stream and the background dark forest. Visually, the HC algorithm not only accurately replicates the physical boundaries of the foreground pebbles with high response extrema, but also, through rigid clustering truncation, forcibly suppresses low-frequency redundant background elements such as distant mountains and the sky to zero-point responses. This high-fidelity mask, combining absolutely rigid boundaries and a pure background, completely severs the energy coupling between the foreground and background at the physical level, laying the foundation for cross-gain in subsequent detail layers.

[0076] Figure 3 The study presents a comparison of the mean values ​​of five saliency extraction algorithms (GBVS, AC, FT, LC, and HC) under objective evaluation metrics.

[0077] Figure 3 (a) quantifies the GMG scale of the masks extracted by each algorithm. This metric directly reflects the physical contrast and texture change rate of the microscopic details of the image. Statistical analysis confirms that the gradient capture performance of GBVS, FT, and AC algorithms is extremely limited, achieving only 1.776, 2.702, and 4.012, respectively; the GMG of the LC algorithm surges to 10.915, while the HC algorithm, which this invention is anchored to, approaches the peak at 10.191. As can be seen from the visual analysis above, the high gradient of LC essentially stems from the erroneous high-frequency activation of low-frequency background areas such as the distant sky; in contrast, the gradient gain of the HC algorithm converges to the physical contour of the real foreground target, confirming its extremely high feature focusing accuracy and effective gradient purity.

[0078] Figure 3 (b) reveals the statistical distribution of mask information entropy. In this evaluation domain, information entropy inversely measures the algorithm's ability to suppress the background and the purity of the mask. The GBVS algorithm recorded 6.594, confirming severe spatial energy dispersion within its mask; the entropy values ​​of LC, FT, and AC fluctuated between 5.1 and 6.0. Notably, the HC algorithm drastically compressed the information entropy to 2.851. This low polarization objectively demonstrates the absolute role of the HC operator in stripping away redundant background and purifying the mask's signal-to-noise ratio dimension.

[0079] Figure 3(c) compares the global spatial frequency (SF). SF maps the fluctuation intensity of the spatial domain feature topology, and its increase closely corresponds to the enrichment of high-frequency information and the sharpening of physical boundaries. The distribution curves show that, constrained by low-pass properties, the SFs of GBVS, FT, and AC are only 3.671, 7.533, and 12.266, respectively, revealing that traditional smoothing operations will affect the structural priors; LC jumps to 28.532 based on brightness contrast; while the HC algorithm obtains a global extreme value of 31.291. Objective data corroborate that the HC operator, while stripping away the semantic subject, obtains the microscopic physical high-frequency skeleton of the target intrinsics at a lossless level.

[0080] Figure 3 (d) in the figure presents the global standard deviation (STD) scale, which quantifies the macroscopic contrast span of the mask grayscale distribution. The data shows that AC has the weakest global polarization ability (21.911); FT and GBVS are approximately 34.008 and 36.564, respectively; LC and HC both exceed the 47 threshold (47.738 and 47.333, respectively). The extremely high values ​​of both confirm their algebraic ability to construct a very strong macroscopic contrast level between the bright foreground and the dark background.

[0081] Figure 3 (e) illustrates the comparison of Enhanced Estimation Metrics (EME). EME accurately captures extreme tone transitions within the local mesh. The GBVS, FT, and AC algorithms exhibit extremely low EMEs, revealing the blurred transition states of their edge topologies; LC climbs to 252.934; while the HC operator shows an exponential leap, reaching 424.624. The prior of the lowest coupling domain and minimal information entropy rigorously infers that the HC algorithm's extremely high EME is not due to the uncontrolled amplification of pseudo-random noise, but rather to maintaining an absolutely pure substrate, achieving a physical transition rate at local high-frequency boundaries. This provides mathematical support for its use as a high-confidence energy assignment mask.

[0082] In constructing a two-layer dynamic dimming architecture, this invention establishes two core underlying mechanisms: "multi-dimensional feature space cross-modulation" and "dual-track nonlinear weighted fusion." To clarify the irreplaceable and mathematically consistent nature of these mechanisms in physical reconstruction, this invention deeply integrates microscopic visual decoupling (qualitative state) and leave-one-out ablation of control variables (quantitative state), and conducts in-depth logical verification of the core algebraic modules.

[0083] After stripping away the detail layer that carries high-frequency perturbations, blindly applying a constant gain equal to the global mean will inevitably trigger the manifestation of smooth domain artifacts and a malignant increase in the background noise of extremely dark physical blind spots. To achieve energy compensation with highly targeted attention weights, the algorithm extracts multi-dimensional physiological and physical perception scales and orthogonally fuses them into the final spatial detail gain matrix. Figure 4 The qualitative evolution map of the cross-modulation of this multidimensional feature manifold was deconstructed.

[0084] Figure 4 (a) in the figure represents a target saliency semantic mask based on global histogram feature clustering. As can be seen from the mapped view, the motorcycle team and riders at the front of the field of view are assigned extremely high response extrema (highlighted mapping) as high-weight semantic targets; conversely, the response weights of non-focused redundant fields such as the background ramp and grass are forcibly attenuated. This mask blocks the spatial channels for gain to pour into invalid background clutter at the source in subsequent modulation.

[0085] Figure 4 Mapping (b) in the diagram relies on a physical edge space mask calculated using the normalized Sobel operator. Visual analysis shows that high-frequency geometric textures such as the motorcycle's spoke frame, frame lines, and fabric folds are extracted using polarization; conversely, smooth paint surfaces with uniform illumination and homogeneous background areas exhibit zero-point dark convergence. This mask completely achieves spatial separation between high-frequency transition edges and low-frequency flat regions, ensuring that subsequent high-energy compensation strictly converges to the true physical topology and reducing the irrational amplification of secondary perturbations in smooth regions.

[0086] Figure 4 (c) illustrates a nonlinear physiological mask constructed based on the HVS extreme brightness masking prior. Its topological mechanism strictly follows the Weber-Fechner law: the retinal center exhibits a sharp peak for the mid-range grayscale, which attenuates towards the bright and dark poles. The scale exhibits a nonlinear spatial weight field: the uniformly lit mid-tone vehicle body is anchored to a high-brightness peak weight; while the gain weight of the undercarriage shadow and the extremely dark blind zone experiences a precipitous drop (forced to converge to a minimum constant of 0.35). This mechanism fundamentally severs the mathematical path of the erroneous amplification of invisible additive noise in the extremely dark domain from a physiological perspective.

[0087] Figure 4 (d) in the diagram presents the ultimate detail gain matrix (Gmatrix) after cross-concatenation of spatial dot product and addition. Looking at its energy distribution, we can see that this matrix represents the high-order mathematical intersection of the heterogeneous features of the first three dimensions: it achieves both... Figure 4 (b) The sharp geometric framework in the middle can also achieve the effect of being subjected to Figure 4 In (a) the target saliency semantic mask and Figure 4 The physical edge space mask in (c) is characterized by: extremely bright main outlines; and significant reduction in blind spots or background clutter.

[0088] A global ablation experiment was conducted to demonstrate the indispensable boundary value of each mask from an algebraic perspective. The experiment used the global average gain extracted from the fully matched model (Proposed) as a static baseline, forcing each ablation group to redistribute spatial energy under the condition of absolute conservation of total gain. The quantization scale and normalized multidimensional mapping of each variable stripping group are detailed in Table 1 and... Figure 5 .

[0089] Table 1. Comparison of objective evaluation indicators for ablation experiments using multidimensional spatial feature modules.

[0090] Given the differences in polarity and dimensions of heterogeneous evaluation scales (Entropy exhibits inverse polarity, while the others are positive polarity), this invention performs Min-Max normalization (linearly mapping to the [0.2, 1.0] domain) on the data matrix in Table 1, and performs center-reversed topology on Entropy to calculate a multidimensional feature radar map (e.g. Figure 5 (As shown). The ablation deconstruction of the physical mechanism of the space module is as follows: (1) Equal baseline mapping manifold This experimental group removed all multidimensional spatial constraints, allowing the constant high-frequency gain energy to be uniformly and diffusely distributed across the two-dimensional manifold. While this mean-gain method can macroscopically improve the global frequency response, it indiscriminately amplifies the weak background thermal noise in smooth background regions (such as the sky, wall base, etc.). Quantization results show that the EME index of this group is only 11.999, which is at a moderate level, verifying that the simple mean-gain strategy cannot achieve spatial focusing on core high-order semantic regions.

[0091] (2) Geometric anchoring effect of physical mask (w / o Edge) This experimental group removed the physical gradient constraint (Wedge) within the total energy closed-loop constraint, allowing the gain energy to be controlled solely by semantic and physiological guidance mechanisms. Due to the lack of pixel-level anchoring of the rigid geometric contour by the underlying gradient operator, the high-frequency compensation energy is prone to diffusion and spillover within the main body region. Macroscopic quantitative data show that the GMG index of this group dropped to the lowest value of 11.316 across the entire field, and the SF index simultaneously dropped to 26.700, exhibiting a deep dip in the corresponding quadrant of the radar manifold. This verifies that the physical edge operator has an irreplaceable physical foundation role in anchoring the high-frequency skeleton, suppressing edge low-pass blurring, and ensuring structural fidelity.

[0092] (3) Semantic gating effect of saliency prior (w / o Saliency) This experimental group removes the saliency prior (β) mapping mechanism, causing the algorithm to degenerate into a blind gain mode based on a single underlying gradient and brightness features. Under this boundary condition, high-contrast secondary textures in the background region (such as distant tree canopies, disordered gravel interference, etc.) will receive the same weight as the core foreground region, leading to an imbalance in the spatial hierarchy of the gain matrix. Quantitative results show that the EME index of this group is 11.642, showing a significant downward trend, indicating that the increase in invalid high-frequency redundancy in the background offsets and weakens the visual tension of the core subject. This ablation experiment verifies the necessity of introducing saliency semantic priors.

[0093] (4) Background noise suppression effect of physiological visual mask (w / o HVS) This experimental group removed the HVS physiological visual scale mechanism. Experimental data showed that the GMG index of this group was 12.035 and the SF index was 28.262, exhibiting peak values ​​across the entire field. However, this phenomenon can be physically disproven by combining the information entropy index: in discrete signal systems, differential operators and frequency operators have a natural sensitive amplification effect on disordered additive noise; when the nonlinear suppression threshold of the HVS mask for extremely dark regions is missing, the physical thermal noise hidden in the dark field background will be misidentified as high-frequency structures during the brightening and reconstruction process, and thus linearly amplified. The information entropy of this ablation group reached the highest value of 7.352, fully exposing the significant deterioration of the substrate signal-to-noise ratio. This confirms that the increase in the GMG and SF indices of this group is not an improvement in effective structural information, but rather an artificially high increment mixed with a large amount of malicious background noise in the dark field; the role of the HVS mask is to effectively suppress the excessive amplification of dark noise within the range that conforms to the human visual perception threshold.

[0094] (5) Global envelope of three-dimensional cross-modulated manifold (Proposed) This experimental group fully integrates a three-dimensional heterogeneous prior mechanism, and the proposed architecture achieves Pareto optimality among multi-dimensional polarization indices. Compared with the group without HVS, the complete architecture maximizes the extraction of the contrast transition rate of real physical textures while strongly suppressing background noise (constraining and stabilizing the information entropy at 7.342), achieving the highest local microscopic contrast ratio (EME) of 12.232. Mapped to a radar image, the manifold of this invention forms the most complete and balanced outer envelope field within the five-dimensional polygonal domain.

[0095] Quantitative ablation experiments demonstrate that the three-dimensional cross-scale mechanism of this application is not a simple linear algebraic superposition, but rather a physical mechanism that reconstructs the energy accumulation excitation of semantic subjects and the effective suppression of dark field background through high-order cross-modulation, achieving optimal performance configuration within the gain space of limited computing power.

[0096] The luminance probability density function (PDF) is a low-level statistical scale for analyzing the global dynamic tolerance and grayscale evolution of algorithms. This is achieved by analyzing the luminance PDF distribution curves of the output images of various algorithms. Figure 6 By conducting comparative analysis, the technical effectiveness of the invention's architecture can be verified from a statistical perspective, as follows: (1) Dark grayscale range (0~50 grayscale) The brightness PDF distribution of the source image has significant defects: an impact peak with a probability value of 0.012 is formed at the grayscale extreme point 0, and dense comb-like discrete quantization distortion peaks are present in the low grayscale range of 10~50, reflecting that the original image has serious underexposure and polarization quantization defects.

[0097] To address this deficiency, traditional S-curve algorithms and existing literature algorithms employ a grayscale arithmetic right-shift strategy, which directly leads to the zeroing of the probability density of the 0-35 grayscale band, causing irrational hollowing and stripping of the dark field substrate. This results in the reconstructed image losing its three-dimensional support in the dark levels, making it prone to loss of dark details and image floating distortion.

[0098] The architecture proposed in this invention performs differentiated processing for dark areas: while accurately removing extremely dark invalid pixels in the 0-15 grayscale range, it smoothly constructs a transition probability peak within the 15-40 grayscale range, reaching a peak value of 0.009 at grayscale 25. This distribution characteristic indicates that, while improving overall brightness, this invention can achieve a smooth and gradual release of dark-field tones with high mathematical resilience, completely avoiding the floating distortion problem caused by the lack of dark area support.

[0099] (2) Mid-tone and secondary bright areas (50~230 gray levels) The brightness PDF values ​​of the source image in this grayscale range are all below the 0.002 threshold, indicating a lack of mid-tone detail. After the traditional S-curve algorithm and existing literature algorithms shift the entire pixel group to the mid-tone band, they only form isolated narrow-band peaks at grayscale 105 (PDF: 0.017) and grayscale 100 (PDF: 0.014), respectively. This reflects that the traditional mapping operator has a severely insufficient topological extension capability in the mid-tone range and cannot achieve uniform unfolding of mid-tone details.

[0100] This invention's architecture achieves a revolutionary frequency band reconstruction for this band: First, it generates a probability peak (PDF: 0.019) near grayscale 78, algebraically weighting the core saliency high-energy components. More importantly, within a wide brightness range of 100-220, the algorithm matrix maintains a highly continuous and smooth high-frequency amplitude distribution, with the PDF value stabilizing between 0.002 and 0.008. This high-bandwidth broadening probability diffusion topology provides more ample information entropy support from the information theory foundation, verifying that the algorithm can achieve full-width physical expansion of limited midtone details.

[0101] (3) Extreme highlight area (grayscale 230~255) In this extreme highlight range, all three reconstruction algorithms exhibited high probability density convergence. Among them, the existing literature algorithm reached a peak value of 0.021 at gray level 238, and the traditional S-curve algorithm climbed to 0.042 at gray level 235; while the network architecture proposed in this invention constructed the highest peak value in the entire domain near gray level 238, with a PDF value exceeding 0.053, and at the absolute extreme point of gray level 255, it preserved the highlight pole anchor that perfectly matched the original image (PDF: 0.012).

[0102] The high-density pixel aggregation characteristics in this extreme range confirm the convergence law of the synergistic effect of the dynamic margin exponential soft knee truncation and HVS physiological mask at the output end of the architecture of this invention: the algorithm uses a controllable algebraic model to uniformly constrain non-core sub-brightness pixels and redundant background pixels around the 238 grayscale anti-overshoot threshold, and stretches the macroscopic brightness contrast to the physical limit under the premise of locking the absolute highlight overflow boundary.

[0103] This invention isolates the Local Enhancement Metric (EME) and Entropy as a pair of monitoring scales with opposing polarities. By setting a weight adjustment factor ω1 for the global S-curve (corresponding to the CLAHE constraint weight 1-ω1), and performing equal-step sampling calculations within a continuous wideband of [0.3, 0.7], this invention accurately plots the dynamic response piecewise linear graph of the two-dimensional constraint index (e.g., ...). Figure 7 This transforms subjective parameter tuning into a rigorous mathematical extremum solution process.

[0104] Figure 7 This is a scatter plot analysis of two Y-axis points. The horizontal axis represents the weighting parameter ω1, and the two vertical axes represent the EME index, which characterizes the local contrast transition rate, and the information entropy index, which characterizes the global background noise divergence, respectively. By scanning the value range of the weighting parameter ω1, the balance between the two major physical mechanisms of local contrast enhancement and global noise suppression can be analyzed, as detailed below: (1) Low weight interval (ω1∈[0.3,0.45]) When the weight parameters are in the far left of the low weight range, the Local Contrast Limiting Histogram Equalization (CLAHE) probability redistribution branch dominates the network. Within this range, the EME index abnormally increases, reaching 15.750 at ω1=0.3; simultaneously, the information entropy rises significantly, reaching a maximum of 7.479, exhibiting uncontrolled growth. From an underlying mechanism perspective, the algorithm excessively enhances the tone boost in extremely dark areas within this range, polarizing and stretching the local brightness probability density function. This disrupts the smoothness of the global grayscale distribution and the coherence of the structural topology, leading to a simultaneous amplification of microscopic background noise. This enhancement method sacrifices the baseline signal-to-noise ratio, resulting in blocky artifacts and grainy noise in the image, causing a decline in subjective visual quality, and is therefore an ineffective contrast reconstruction method.

[0105] (2) High weight interval (ω1∈[0.6,0.7]) As the weight parameters slide towards the far right of the high-weight range, the global nonlinear S-shaped mapping operator gradually becomes dominant. Within this range, the information entropy index shows a smooth downward trend, reaching its lowest value at ω1=0.7, indicating that global high-frequency background noise is effectively suppressed; however, the EME index shows a significant decline, plummeting to 11.130. This phenomenon exposes the inherent defects of a single global mapping method: to maintain a smooth gradient in global brightness, the operator merges grayscale values ​​of neighboring pixels with weak differences in the dark field, producing a tone merging effect. This results in the loss of microscopic detail boundaries such as hair strands and ground shadows, ultimately causing the image to lose its sense of depth and appear flat.

[0106] (3) Optimal parameter equilibrium point (ω1=0.55) Considering the performance imbalance between the two intervals, and analyzing the convergence characteristics of the dual-index trajectory in the two-dimensional parameter space, the optimal parameter equilibrium point is determined to be ω1=0.55. Under this parameter ratio, the two performance indices achieve balanced optimization, effectively avoiding bidirectional performance degradation at both ends: First, the algorithm controls the upper limit of information entropy to 7.342 through constraint operators, suppressing the background thermal noise in the dark area to within the perceptible difference threshold for the human eye; Second, the algorithm avoids the nonlinear decay of the EME index, stabilizing the local contrast transition rate at a relatively high level of 12.232.

[0107] The above quantitative trajectory analysis verifies that the heterogeneous dual-track fusion ratio determined in this invention is a Pareto optimal solution obtained by extreme optimization of two opposing physical mechanisms: local enhancement and global suppression. This provides a low-level physical foundation with balanced gray-scale distribution and rich detail for subsequent high-frequency feature cross-excitation.

[0108] Figure 8 This demonstrates the original low-light image of the same scene and the visual reconstruction results after processing with three different algorithms. Figure 8 As shown in (a), the original input image is in a low-light state and has an underexposure problem. Due to the limited lighting conditions, the micro-contrast between the stones and the mortar joints is low and the texture details are blurred. The red door and window areas in the image are dark and lack depth. The weeds and shadow areas at the bottom of the image show detail adhesion, and large dark areas obscure part of the structural information of the scene.

[0109] like Figure 8 (b) and Figure 8As shown in (c), the processing results of the traditional S-curve algorithm and the algorithm in the comparative literature are respectively. From a visual perspective, both algorithms improve the global brightness of the image, making the main building, which was originally in the dark, visible. However, they have obvious limitations in maintaining local contrast: in the process of increasing the global brightness, they fail to effectively maintain the high-frequency contrast between adjacent pixels, resulting in varying degrees of whitening and graying of the stone texture on the wall, and a decrease in material texture; although the red door and window areas are brightened overall, the three-dimensionality of light and shadow is insufficient, the color expression is weak, and the overall image is accompanied by a slight haze.

[0110] Figure 8 (d) in the figure represents the processing result of the algorithm of this invention. Under the joint guidance of edge mask and saliency mask, the algorithm of this invention effectively overcomes the local contrast attenuation defect of existing algorithms while improving the overall brightness: the outline of the stones and the shadow details of the brick seams in the image are well preserved, and the three-dimensionality and clarity of the wall texture are significantly improved. In terms of color reproduction and dark detail representation, the reconstructed red doors and windows have appropriate color saturation and natural light and shadow transitions, effectively suppressing the color distortion phenomenon that is prone to occur during nonlinear brightening; at the same time, the detailed texture of the weed area at the bottom of the image is clearly displayed, and no obvious noise amplification is observed, reflecting the noise suppression effect of the physiological visual mask in dark area processing.

[0111] Comprehensive qualitative comparison shows that the algorithm of this invention, through a multi-dimensional spatial energy regulation mechanism, better balances the relationship between brightness enhancement and detail preservation, and the overall visual performance is significantly improved compared with existing comparison algorithms.

[0112] To objectively and rigorously quantify the improvement effect of each algorithm on the original image, this invention introduces the Relative Change Ratio against Original (Rorig) as an evaluation benchmark. Based on Rorig, [the following is a more detailed description of the algorithm's performance]. Figure 9 A detailed quantitative analysis of the objective data from the six subgraphs was conducted: Figure 9 In the figure, (a) represents the mean gradient (GMG) index, which is used to characterize the sharpness of image edges and the clarity of micro-textures. The GMG of the original image is 9.069; the GMG after processing by the algorithm in the literature is 9.664, an improvement of 6.56% compared to the original image; the GMG of the algorithm in this invention reaches 11.815, an improvement of 30.28% compared to the original image. From the above quantitative data, it can be seen that the algorithm in the literature has limited ability to enhance high-frequency edge information, while the algorithm in this invention can more significantly improve the local edge gradient of the image and enhance the texture edge expression effect.

[0113] Figure 9(b) in the figure represents the information entropy index, which characterizes the average amount of information contained in an image and the richness of its grayscale distribution. The information entropy of the original image is 7.103; the information entropy of the algorithm in the literature is 7.161, an increase of 0.82% compared to the original image; the information entropy of the algorithm in this invention is 7.342, an increase of 3.36% compared to the original image. In the process of nonlinear image enhancement and reconstruction, the improvement of details and contrast generally leads to a moderate increase in information entropy. The algorithm in this invention controls the increase in information entropy to within 3.36%, indicating that while improving image clarity, the increase in global information content and background noise can be controlled within a reasonable range.

[0114] Figure 9 (c) represents the spatial frequency (SF) index, which reflects the overall activity level and the richness of spatial details in an image. The original image has an SF of 21.395; the algorithm in the literature has an SF of 21.574, a slight improvement of only 0.84% ​​compared to the original image, indicating that the algorithm is difficult to effectively increase the effective spatial details of the image; the algorithm of this invention achieves an SF of 27.785, an improvement of 29.87% compared to the original image. This data proves that the algorithm of this invention can effectively enrich the texture details of the image and improve the activity level of spatial details.

[0115] Figure 9 In the figure, (d) represents the standard deviation (STD) index. The standard deviation characterizes the dispersion of pixel gray levels relative to the mean, and is an important objective indicator for measuring the global dynamic range and macroscopic contrast of an image. The STD of the original image is 47.606; the STD of the literature algorithm decreases to 45.808, a negative increase of 3.78% relative to the original image. This data quantitatively explains the phenomenon that the literature algorithm tends to have a decrease in global contrast. The STD of the algorithm of this invention increases to 59.151, a positive increase of 24.25% relative to the original image, confirming that the algorithm of this invention can widen the pixel gray level distribution range and effectively improve the global contrast of the image.

[0116] Figure 9 In the figure, (e) represents the local contrast EME index, which characterizes the local contrast level of an image. The original dark image has an EME of 23.295. This high value originates from the large number of extremely dark pixels in the dark image that are close to zero. Due to the influence of the denominator of the logarithmic calculation model approaching zero, a mathematical singularity effect occurs. After brightness enhancement and elimination of extremely dark pixels, the EME of the literature algorithm drops to 8.947, with a change rate of -61.59% relative to the original image, indicating a severe decrease in local effective contrast. The EME of the algorithm of this invention is 12.232, with a change rate of -47.49% relative to the original image. Under the same condition of eliminating mathematical singularities, the EME decrease of the algorithm of this invention is significantly smaller than that of the literature algorithm, and it can better maintain and reconstruct the local physical contrast of the image.

[0117] Color Hue Retention Rate (CHPR) measures the consistency of color information between the enhanced image and the original image. Its ideal upper limit is 1. However, due to the inherent mathematical properties of this metric, the traditional relative growth rate method is insufficient for objectively evaluating color performance. This invention introduces Color Distortion to quantify the color deviation error of the enhanced image relative to the original image. Figure 9 In (f), the color distortion of the literature algorithm is 0.29%, while the color distortion of the algorithm of this invention is further reduced to 0.22%. Calculations show that, compared with the literature algorithm, the color distortion error of the algorithm of this invention is reduced by 24.14%. This quantitative result shows that the algorithm of this invention can effectively suppress the color shift problem caused by nonlinear mapping while completing the brightness reconstruction, and has a better color fidelity capability.

[0118] The quantitative test results of the above multiple indicators show that the global average statistical indicators verify the comprehensive performance advantages of the algorithm of this invention in contrast enhancement, detail reconstruction, and color fidelity from a macroscopic perspective. However, in the objective evaluation system of image processing, the local lighting conditions and texture complexity of natural scenes vary greatly. Relying solely on global average statistical values ​​can easily mask the risk of local failures or drastic fluctuations in indicators in extreme or special scenarios.

[0119] To further verify the generalization ability and statistical robustness of each algorithm in complex and varied scenarios, this invention introduces the Kodak dataset, a standard dataset in the field of computer vision. This dataset contains 24 lossless images with rich scene diversity and mutual independence. By calculating the objective metrics of each algorithm on these 24 independent samples, this invention plots a discrete distribution scatter plot (e.g., ...). Figure 10 (As shown). The following will conduct an in-depth quantitative analysis of this set of multi-sample discrete data from micro-dimensional perspectives such as the universality of global gain, the processing efficiency of extreme points, and the topological fidelity of data distribution.

[0120] Figure 10(a) in the figure represents the discrete fluctuation curve of the mean gradient (GMG), which is used to characterize the edge sharpness and micro-texture richness of the image. Due to the differences in the frequency domain features of each test image, the GMG scatter points of different samples show obvious fluctuations. From the overall distribution, it can be seen that the algorithm of this invention achieves stable positive gain on all 24 test samples, and the trajectory of index change is synchronized with the topological morphology of the original image, indicating that the algorithm enhances the gradient while preserving the semantic structure of the original scene of the input image. For typical extreme value samples: For the kodim01 and kodim13 samples with rich original high-frequency details, the original image GMG is 14.89 and 19.14, respectively. The algorithm in the literature only improves it to 16.55 and 18.24, while the algorithm of this invention improves it to 21.33 and 23.38, proving that the algorithm of this invention has a better ability to excite high-frequency features in complex texture scenes. For the kodim23 sample containing a large area of ​​smoothness, the original image GMG is only 4.25, while the algorithm in the literature improves it to 4.35. The algorithm of this invention adopts a restrained enhancement strategy, gradually improving it to 5.08. The above data show that the algorithm of this invention can achieve accurate enhancement of high-gradient regions and moderate constraint of low-gradient regions, and has the ability to adaptively enhance edge features.

[0121] Figure 10 (b) in the figure represents the discrete fluctuation curve of information entropy, which is used to comprehensively characterize the richness of image information and the level of background noise control under different scenes. The baseline information entropy of the 24 original images varies significantly with different scenes, ranging from 5.568 (smooth scene kodim02) to 7.824 (complex texture scene kodim08). The scatter plot trends of both the literature algorithm and the algorithm of this invention maintain topological phase synchronization with the original images, and neither algorithm destroys the underlying probability distribution structure of the original images. For extreme samples: in the original high-entropy complex scene kodim08, the original image information entropy is 7.824, the literature algorithm is 7.583, and the algorithm of this invention maintains 7.602; in the original low-entropy smooth scene kodim02, the original image information entropy is 5.568, and the algorithm of this invention gradually increases to 5.868. This result confirms that, relying on the adaptive adjustment of the physiological visual HVS mask, the algorithm of this invention can achieve moderate enhancement of smooth scenes and avoid the surge of background noise caused by over-processing in complex texture scenes when performing nonlinear brightness enhancement.

[0122] Figure 10(c) in the figure represents the spatial frequency (SF) discrete fluctuation curve, which characterizes the global activity and spatial detail dispersion of the image. From the scatter distribution of the entire sample, for 24 test images, the spatial frequency index of the algorithm of this invention is generally superior to the traditional S-curve algorithm and the algorithm compared in the literature. For the kodim08 sample with its original high-frequency, dense details, the original image's SF is as high as 39.641. After processing by the literature algorithm, details are lost, and the SF drops to 35.109. The algorithm of this invention significantly improves it to 45.641. For the kodim12 sample with scarce details, the original image's SF is only 11.050. The literature algorithm improves it to 15.578, and the algorithm of this invention further improves it to 18.700. The positive improvement effect of the index across different complexity scenarios proves that the multi-dimensional feature routing mechanism of this invention breaks through the high-frequency enhancement bottleneck of traditional algorithms and can effectively enrich the microscopic spatial details of the image.

[0123] Figure 10 In the figure, (d) represents the discrete fluctuation curve of the standard deviation (STD). The standard deviation is an important quantitative indicator for evaluating the global contrast of an image and suppressing graying and fogging distortion. Test data shows that the algorithm in the literature tends to suffer severe standard deviation collapse when processing high-contrast scenes. For example, in the kodim08 and kodim15 samples, the original image STDs were 63.272 and 82.174, respectively. After processing by the algorithm in the literature, they decreased to 50.987 and 65.903, respectively, and the dynamic range was compressed, corresponding to the visual defect of graying. The algorithm of this invention maintains the STDs of the above two samples at a relatively high level of 61.961 and 76.361, respectively; for the low-contrast scene kodim02, the original image STD was 20.022, and the algorithm of this invention positively increases it to 25.707. The above quantitative results show that the algorithm of this invention can overcome the contrast compression defect caused by traditional global dimming and ensure the macroscopic layering of the image.

[0124] Figure 10 In the figure, (e) represents the discrete fluctuation curve of the local enhancement metric (EME). The original image's dark areas contain a large number of pixels approaching zero, causing a physically meaningless artificially high EME, such as 41.720 for the kodim05 sample and 34.433 for the kodim13 sample. After brightness adjustment to eliminate mathematical singularities, the EME of the algorithm in the literature shows a precipitous drop, decreasing to 16.898 and 13.122 for the two samples mentioned above, respectively. The algorithm of this invention, while eliminating artifacts in the dark areas of the original image, effectively alleviates this metric collapse problem, increasing the EME of the two samples to 25.093 and 18.823, respectively. The superior metric under the same reconstruction conditions demonstrates that the saliency-edge joint guidance mechanism of this invention can more effectively lock and restore the true local contrast of the image.

[0125] Figure 10In the figure, (f) represents the discrete fluctuation curve of Hue Retention Rate (CHPR). The ideal upper limit of hue retention rate is 1, and the CHPR of the original image is always equal to 1. After nonlinear brightness reconstruction, all comparison algorithms inevitably produce a certain degree of color shift. However, in all 24 test samples, the CHPR scatter plot of the algorithm of this invention is generally higher than that of the algorithm in the literature. Especially in the kodim09 sample with severe color shift, the CHPR of the literature algorithm drops to 0.99316, while that of the algorithm of this invention increases to 0.99512; for the ordinary scene kodim01, the CHPR of the algorithm of this invention reaches 0.99814. Statistical results show that the overall index of the algorithm of this invention is closer to the ideal upper limit, has a stronger resistance to attenuation in extreme color shift scenes, and has better robustness in color reconstruction.

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multi-dimensional feature cross-fusion double-layer dynamic dimming and color fidelity method, characterized in that, Includes the following steps: Obtain the RGB color image of the image to be displayed; The RGB color image is converted to the HSV color space, and the hue channel, saturation channel and lightness channel are separated. The lightness channel is then locked as the only dimming channel. A bilateral filtering operator with spatial and value domain dual Gaussian weighting is used to decouple the lightness channel spatially and frequency-frequency, separating the low-frequency base layer that carries the macroscopic illuminance distribution and the high-frequency detail layer that carries the microscopic texture features. The reconstructed base layer is obtained by performing a dual-track nonlinear reconstruction on the low-frequency base layer. The dual-track nonlinear reconstruction includes: the first track uses an adaptive anchor point S-curve to stretch the global contrast, and the second track uses a limited contrast adaptive histogram equalization to wake up local dark details. A three-dimensional cross-joint gain matrix is ​​constructed to apply spatial heterogeneity gain to the high-frequency detail layer, resulting in an enhanced detail layer. The three-dimensional cross-joint gain matrix is ​​generated by fusing a physical edge mask representing the physical geometric boundary, a cognitive saliency mask representing the cognitive semantic weight, and a physiological visual mask representing the visual perception threshold. The reconstructed base layer and the enhanced detail layer are summed and merged pixel by pixel to obtain the reconstructed brightness channel; The reconstructed luminance channel is subjected to a nonlinear exponential decay piecewise soft knee mapping based on the McLaurin series expansion to obtain the output luminance components. Based on the pixel-level magnification ratio before and after the brightness channel enhancement, adaptive power function color compensation is applied to the original saturation channel to obtain the compensated saturation components. The dynamic backlight control factor is calculated based on the global statistical mean of the reconstructed brightness channel, and the physical driving flux of the LED backlight is adaptively adjusted according to the average brightness of the screen. Convert the hue channel, compensated saturation component, and output brightness component back to the RGB color space to output a color image after dimming compensation.

2. The method of claim 1, wherein the method is a multi-dimensional feature cross fusion based dual-layer dynamic dimming and color fidelity method, characterized in that, The joint weight matrix of the bilateral filter operator is composed of the product of the spatial domain Gaussian distance weight and the value domain Gaussian brightness weight, as expressed below: in, For spatial domain Gaussian distance weights, The horizontal coordinate of the currently processed pixel. The spatial ordinate of the currently processed pixel. m The x-coordinate of the sliding pixel within the neighborhood. n The vertical coordinate of the sliding pixel within the neighborhood. For spatial standard deviation parameters, The value range is a Gaussian brightness weight. For the brightness channel in Pixel value at that location, For the brightness channel in Pixel value at that location, The parameter is the standard deviation of the range. This is the joint weight matrix; The expression for the low-frequency base layer is: in, As a low-frequency base layer, For the current pixel being processed during bilateral filtering ( i , j ) and sliding pixels in the neighborhood ( m , n The joint weights between ) The expression for the high-frequency detail layer is: in, This is a high-frequency detail layer.

3. The method for dual-layer dynamic dimming and color fidelity through multi-dimensional feature cross-fusion according to claim 1, characterized in that, By performing dual-track nonlinear reconstruction on the low-frequency foundation layer, the specific content of the reconstructed foundation layer is obtained, including: The first reconstruction component is obtained by globally stretching the low-frequency base layer using an adaptive anchor point S-curve. A limiting contrast adaptive histogram equalization is used to locally awaken dark details in the first reconstructed component to obtain the second reconstructed component; The first and second reconstruction components are weighted and fused to obtain the reconstruction base layer.

4. The method for dual-layer dynamic dimming and color fidelity based on multi-dimensional feature cross-fusion according to claim 3, characterized in that, The first reconstructed component is obtained by using an adaptive anchor point S-curve to perform global contrast stretching on the low-frequency base layer, and includes the following specific contents: The pixel grayscale values ​​of the low-frequency base layer are sorted in ascending order. The 25% and 50% quantiles of the sorted sequence are extracted, and an empirical convergence factor is introduced to calculate the stretching anchor point. The calculation formula is as follows: in, For adaptive photometric anchor points, At the 25th percentile, 50% quantile; Based on the stretching anchor points, an S-curve mapping function is constructed, and nonlinear contrast stretching mapping is performed on the low-frequency base layer to obtain the first reconstructed component. The expression for the S-curve mapping function is: in, This is the illumination gain multiplier mapping matrix. The horizontal coordinate of the currently processed pixel. The spatial ordinate of the currently processed pixel. To control the slope factor of the tensile tension curve. As a low-frequency base layer, To prevent the minimum constant with a denominator of zero, This is the first reconstructed component.

5. The method for dual-layer dynamic dimming and color fidelity based on multi-dimensional feature cross-fusion according to claim 3, characterized in that, A limiting contrast adaptive histogram equalization is used to locally awaken dark details in the first reconstructed component, resulting in the following specific contents of the second reconstructed component: The gray-level histogram of the first reconstructed component is statistically analyzed, and a limiting factor is preset to calculate the global peak-shaving threshold. The peak values ​​of the histogram that exceed the peak-shaving threshold are truncated, the total overflow is calculated and evenly distributed to each gray level, and the histogram after peak-shaving and valley-filling is obtained. The cumulative distribution function is calculated based on the histogram after peak shaving and valley filling, and gray-level remapping is performed to obtain the second reconstructed component.

6. The method for dual-layer dynamic dimming and color fidelity through multi-dimensional feature cross-fusion according to claim 1, characterized in that, The expression for reconstructing the base layer is: in, To reconstruct the base layer, The horizontal coordinate of the currently processed pixel. The spatial ordinate of the currently processed pixel. The weighting coefficients for the first reconstructed component. , For the first reconstructed component, This is the second reconstructed component.

7. The method for dual-layer dynamic dimming and color fidelity through multi-dimensional feature cross-fusion according to claim 1, characterized in that, The three-dimensional cross-joint gain matrix is ​​generated by fusing a physical edge mask representing the physical geometric boundary, a cognitive saliency mask representing the cognitive semantic weight, and a physiological visual mask representing the visual perception threshold in the following manner: The Sobel operator is used to perform horizontal and vertical convolution operations on the original brightness channels, calculate the gradient magnitude and normalize it to the [0,1] interval to obtain the physical edge mask; The RGB color image is converted to the CIELAB perceptual uniform color space. Color quantization and dimensionality reduction are performed based on global histogram statistics. The weighted sum of the three-dimensional color difference distance between each representative color level and all pixels in the image is calculated. The calculation results are then mapped back to pixel coordinates and normalized to obtain a cognitive saliency mask. Constructing a physiological visual mask based on the Weber-Fechner law; The physical edge mask, cognitive saliency mask, and physiological visual mask are spatially cross-fused to generate a three-dimensional cross-joint gain matrix.

8. The method for dual-layer dynamic dimming and color fidelity through multi-dimensional feature cross-fusion according to claim 1, characterized in that, The expression for the physiological visual mask is: in, For physiological visual mask weights, For the brightness channel in Pixel value at that location, The horizontal coordinate of the currently processed pixel. represents the spatial ordinate of the currently processed pixel, 255 is the physical maximum constant of the luminance channel, and 0.35 is the preset minimum value to prevent loss of detail in extreme brightness areas.

9. The method for dual-layer dynamic dimming and color fidelity through multi-dimensional feature cross-fusion according to claim 1, characterized in that, Applying a nonlinear exponential decay piecewise soft-knee mapping based on the Maclaurin series expansion to the reconstructed luminance channel yields the following specific contents of the output luminance components: The reconstructed brightness is normalized to the [0,1] interval to obtain the normalized reconstructed brightness; A preset transition threshold is set. When the normalized reconstructed brightness is less than or equal to the preset transition threshold, the output brightness component is the normalized reconstructed brightness. When the normalized reconstructed brightness is greater than the preset transition threshold, the expression for the output brightness component is: ； in, To output the lightness component, The horizontal coordinate of the currently processed pixel. The spatial ordinate of the currently processed pixel. For the currently processed pixel ( i , j The reconstructed brightness value at point ) is the brightness value after normalization. To preset the turning threshold, .

10. The method for dual-layer dynamic dimming and color fidelity through multi-dimensional feature cross-fusion according to claim 1, characterized in that, Based on the pixel-level magnification ratio before and after brightness channel enhancement, adaptive power function color compensation is applied to the original saturation channel to obtain the specific contents of the compensated saturation components, including: The saturation gain factor is determined based on the pixel-level ratio of the enhanced pre-brightness channel to the output brightness channel. The initial compensated saturation is obtained by performing a product-type compensation on the original saturation based on the saturation gain factor. An amplitude constraint is applied to the initial compensated saturation to obtain the compensated saturation component.

Images