Image processing methods, apparatus, computing devices, and computer-readable storage media

Abstract

This application discloses an image processing method, apparatus, computing device, and computer-readable storage medium, comprising: extracting a multi-scale feature map of an image to be processed; determining the acquisition scene corresponding to the image to be processed based on the brightness distribution information corresponding to the multi-scale feature map; if the acquisition scene is a preset scene, generating a signal probability map corresponding to the image to be processed based on a preset image processing network and the multi-scale feature map; performing discrete transformations on the image to be processed and the signal probability map respectively, and extracting a signal layer and a noise layer from the image to be processed based on the discrete transformation results and the signal probability map; and processing the image to be processed based on the signal layer and the noise layer. The image processing scheme of this application can solve the problem that traditional techniques cannot balance the preservation of highlight details and the suppression of noise in dark areas.

Description

Technical Field

[0001] This application relates to the field of data processing technology, specifically to an image processing method, apparatus, computing device, and computer-readable storage medium. Background Technology

[0002] With the rapid development of smart cities and intelligent security, outdoor surveillance cameras are widely used in key areas such as traffic intersections, parking lots, and park entrances. However, image quality in nighttime scenes remains a pain point for the industry. Strong light sources such as vehicle headlights, streetlights, and billboards can cause local overexposure, halos, and glare in images, while areas far from the light source suffer from severe noise and loss of detail due to insufficient lighting.

[0003] Currently, there are two traditional solutions: (1) using high dynamic range (HDR) sensors or multiple exposure fusion technology, but the hardware cost is high and artifacts are easily generated for moving targets; (2) using wide dynamic range (WDR) or high light compensation (HLC) algorithms based on traditional image processing, but these methods often process globally and cannot achieve a balance between preserving texture details in the highlight area and suppressing noise in the dark area. It can be seen that in strong light scenes at night, traditional solutions generally have technical defects such as high cost, poor processing effect, and inability to balance the preservation of strong light details and the suppression of noise in the dark area, which are difficult to meet the actual use needs of outdoor monitoring. Summary of the Invention

[0004] This application provides an image processing method, apparatus, computing device, and computer-readable storage medium, which can solve the technical defects of current image processing solutions, such as high cost, poor processing effect, and inability to balance the preservation of strong light details and the suppression of noise in dark areas.

[0005] In a first aspect, embodiments of this application provide an image processing method, including: Extract multi-scale feature maps from the image to be processed; Based on the brightness distribution information corresponding to the multi-scale feature map, the acquisition scene corresponding to the image to be processed is determined; If the acquisition scenario is a preset scenario, then a signal probability map corresponding to the image to be processed is generated based on the preset image processing network and the multi-scale feature map. Discrete transformations are performed on the image to be processed and the signal probability spectrum, respectively, and the signal layer and noise layer are extracted from the image to be processed based on the discrete transformation results and the signal probability spectrum. The image to be processed is processed according to the signal layer and the noise layer.

[0006] Secondly, embodiments of this application provide an image processing apparatus, including: The extraction module is used to extract multi-scale feature maps from the image to be processed; The determination module is used to determine the acquisition scene corresponding to the image to be processed based on the brightness distribution information corresponding to the multi-scale feature map; The generation module is used to generate a signal probability map corresponding to the image to be processed based on a preset image processing network and the multi-scale feature map if the acquisition scene is a preset scene. The transformation module is used to perform discrete transformations on the image to be processed and the signal probability spectrum respectively, and extract the signal layer and noise layer from the image to be processed based on the discrete transformation results and the signal probability spectrum. The processing module is used to process the image to be processed based on the signal layer and the noise layer.

[0007] Thirdly, embodiments of this application also provide a computing device, including: At least one processor; and A memory communicatively connected to the at least one processor; wherein: The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the steps of any of the image processing methods described above.

[0008] Fourthly, embodiments of this application also provide a computer-readable storage medium storing computer instructions, which, when executed by a processor, implement the steps of the image processing method as described in any of the preceding claims.

[0009] This application provides an image processing method, apparatus, computing device, and computer-readable storage medium. After extracting a multi-scale feature map of the image to be processed, the acquisition scene corresponding to the image to be processed is determined based on the brightness distribution information corresponding to the multi-scale feature map. If the acquisition scene is a preset scene, a signal probability map corresponding to the image to be processed is generated based on a preset image processing network and the multi-scale feature map. Then, discrete transformations are performed on the image to be processed and the signal probability map, respectively. Based on the discrete transformation results and the signal probability map, a signal layer and a noise layer are extracted from the image to be processed. Finally, the image to be processed is processed based on the signal layer and the noise layer. In the image processing scheme provided in this application, based on the multi-scale feature map... The brightness distribution information determines the image acquisition scene, specifically identifying preset scenes such as strong nighttime light. This eliminates the need for high-cost HDR sensors and multi-exposure fusion hardware solutions. Simultaneously, a signal probability map is generated by combining a preset image processing network with multi-scale feature maps. Then, discrete transformations are performed on the image to be processed and the signal probability map, accurately separating the signal layer and noise layer of the image based on the transformation results. This achieves decoupling and differentiation of signal and noise. Subsequent differentiated processing is then performed based on the signal layer and noise layer. This effectively improves the local overexposure and glare problems caused by strong nighttime light, thus perfectly preserving the texture details in the highlight areas, while also suppressing noise in the low-light areas and repairing lost image details. This effectively solves the problem that traditional technologies cannot balance the preservation of highlight details and the suppression of noise in dark areas. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic flowchart of the image processing method provided in the embodiments of this application; Figure 2 This is a schematic diagram of the structure of the image processing apparatus provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the computing device provided in the embodiments of this application. Detailed Implementation

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] In the description of this application, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application. Furthermore, the terms "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first," "second," "third," and "fourth" may explicitly or implicitly include one or more of the stated features. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly specified.

[0014] This application provides an image processing method, apparatus, computing device, and computer-readable storage medium, which will be described in detail below.

[0015] In the embodiments of the image processing method of this application, an image processing device is used as the execution subject. For simplicity and ease of description, this execution subject will be omitted in subsequent method embodiments. The image processing device is applied to a computing device equipped with a camera, such as a mobile phone, tablet, PC, vehicle terminal, network camera, high-definition camera, and other terminal devices. The method includes: Multi-scale feature maps of the image to be processed are extracted. Based on the brightness distribution information corresponding to the multi-scale feature maps, the acquisition scene corresponding to the image to be processed is determined. If the acquisition scene is a preset scene, a signal probability map corresponding to the image to be processed is generated based on the preset image processing network and the multi-scale feature maps. Discrete transformations are performed on the image to be processed and the signal probability map respectively. Based on the discrete transformation results and the signal probability map, the signal layer and noise layer are extracted from the image to be processed. The image to be processed is then processed according to the signal layer and the noise layer.

[0016] The image processing scheme provided in this application determines the image acquisition scene based on the brightness distribution information of multi-scale feature maps, and specifically identifies preset scenes such as strong light at night. It does not rely on high-cost HDR sensors and hardware solutions for multiple exposure fusion. At the same time, it generates a signal probability map by combining a preset image processing network with multi-scale feature maps, and then performs discrete transformations on the image to be processed and the signal probability map separately. Based on the transformation results, it accurately separates the signal layer and noise layer of the image, realizing the decoupling and differentiation of signal and noise. In subsequent differential processing based on the signal layer and noise layer, it can effectively improve the problems of local overexposure and halo glare caused by strong light at night, thus perfectly preserving the texture details of the highlight area, and suppressing noise in the low light area and repairing lost image details. It effectively solves the problem that traditional technologies cannot balance the preservation of highlight details and the suppression of noise in the dark area.

[0017] The following sections provide detailed descriptions of each example. It should be noted that the order in which the embodiments are described is not intended to limit the priority of the embodiments.

[0018] An image processing method includes: extracting multi-scale feature maps of an image to be processed; determining the acquisition scene corresponding to the image to be processed based on the brightness distribution information corresponding to the multi-scale feature maps; if the acquisition scene is a preset scene, generating a signal probability map corresponding to the image to be processed based on a preset image processing network and the multi-scale feature maps; performing discrete transformations on the image to be processed and the signal probability map respectively; extracting a signal layer and a noise layer in the image to be processed based on the discrete transformation results and the signal probability map; and processing the image to be processed based on the signal layer and the noise layer.

[0019] Please see Figure 1 , Figure 1 This is a flowchart of an image processing method provided in an embodiment of this application. The image processing method may specifically include the following steps: 101. Extract multi-scale feature maps from the image to be processed.

[0020] The image to be processed refers to the original surveillance image captured by outdoor security cameras in a nighttime environment. This image includes overexposed, haloed, and glare areas caused by strong light sources such as vehicle headlights and streetlights, as well as original image frames with high noise and missing details in low-light areas. The multi-scale feature map is a multi-resolution feature map set obtained after multi-scale hierarchical feature extraction of the image to be processed. It integrates the brightness distribution features, edge contour features, and texture detail features of different levels of the image, and can simultaneously characterize the global brightness trend of the image and the differences between local strong and weak light areas.

[0021] For example, the raw image to be processed can be acquired from the camera, preprocessed to normalize image size and correct pixel grayscale, eliminating shooting distortion and pixel deviation interference. Then, the preprocessed image is downsampled multiple times to obtain multi-scale images at different resolutions. For each scale image, brightness distribution features, edge features, and texture features are extracted to generate single-scale feature maps corresponding to each scale. Finally, the single-scale feature maps of all levels are dimensionally aligned and fused to integrate brightness, edge, and texture information at different resolutions, resulting in a multi-scale feature map corresponding to the image to be processed.

[0022] 102. Determine the operation scenario corresponding to the image processing operation.

[0023] Brightness distribution information refers to the distribution characteristics of pixel brightness values ​​in the entire image and local areas obtained from multi-scale feature maps. This includes information such as the range of bright areas, brightness amplitude, proportion of dark areas, abrupt changes at the boundary between strong and weak light, and brightness gradient distribution. Brightness distribution information can reflect whether the image exhibits localized overexposure due to strong light or widespread dim lighting. The acquisition scene refers to the lighting environment scene corresponding to the capture of the image to be processed. In some embodiments of this application, the operation scene may include a preset scene with strong light interference (such as a scene with strong light sources like car headlights, streetlights, and billboards, accompanied by overexposure, halos, and high noise in dark areas) and a normal nighttime lighting scene (i.e., a scene without significant strong light source interference).

[0024] For example, specifically, pixel brightness data at each scale is extracted from the multi-scale feature map by partitioning, and complete brightness distribution information is obtained statistically. Then, the extracted brightness distribution information is compared and matched with a preset scene brightness feature threshold library. If the image has features such as excessive proportion of bright areas, sudden changes in local brightness, or excessive differences between bright and dark areas, it is determined to be a preset scene, i.e., a strong light interference scene. If there are no features such as excessive proportion of bright areas, sudden changes in local brightness, or excessive differences between bright and dark areas, it is determined to be a non-preset scene, i.e., a normal nighttime lighting scene. It should be noted that in some embodiments of this application, the preset scene refers to a nighttime strong light interference acquisition scene, i.e., a special imaging scene in which strong light sources such as vehicle headlights, streetlights, and billboards are present in the nighttime shooting footage of outdoor surveillance cameras, causing local overexposure, halo, and glare in the image, while the low-light areas far from the light source have severe noise and loss of detail, which is different from ordinary nighttime scenes without strong light source interference.

[0025] 103. If the scene to be collected is a preset scene, then the signal probability map corresponding to the image to be processed is generated according to the preset image processing network and multi-scale feature map.

[0026] Among them, the image processing network is a pre-built and trained dedicated feature parsing network structure that is adapted to the feature representation of strong light images at night. It can receive input multi-scale feature maps, automatically learn deep features such as image brightness distribution, edge texture, and regional illumination differences, and has the ability to distinguish between pixel-level signals and noise. It is a fixed processing architecture for generating probability maps.

[0027] The signal probability map is a pixel-level probability mapping map that corresponds one-to-one with the size of the image to be processed. Each pixel in the signal probability map corresponds to a probability value, which is used to characterize the confidence level that the pixel belongs to the real and effective signal of the image. Specifically, the higher the probability value, the more effective the signal is, such as real scene, edge texture, etc., while the lower the probability value, the more invalid interference information such as halo, glare, dark area noise, etc.

[0028] For example, specifically, after determining that the scene being collected is a preset scene, the extracted multi-scale feature map is input into a preset image processing network. Subsequently, the image processing network performs multi-scale deep feature analysis on the multi-scale feature map to mine latent features such as local illumination differences, edge textures, and boundaries between bright and dark areas. Based on the discrimination rules learned during network training, the confidence level of each pixel belonging to a real and valid signal is evaluated pixel by pixel. Finally, based on the pixel-by-pixel confidence mapping results, a signal probability map with the same resolution as the image to be processed is generated.

[0029] It should be noted that the image processing network provided in this application embodiment is a lightweight multi-scale feature fusion pixel-level probability estimation network specially designed for outdoor security nighttime strong light scenarios. It belongs to a lightweight convolutional neural network architecture, which is different from general image classification and recognition networks. Its core positioning is to perform deep analysis on the input multi-scale feature map, realize the confidence judgment of pixel-level signal and interference noise, and finally output the signal probability map with the same resolution as the image to be processed, which is suitable for the low computing power deployment requirements of embedded edge surveillance cameras.

[0030] The image processing network consists of three hierarchically connected parts: a lightweight backbone feature extraction module, a multi-scale feature fusion module, and a pixel-level probabilistic decoding module. The backbone feature extraction module abandons the standard convolutional structure of traditional large convolutional networks, employing depthwise separable convolutions to build the base layers, significantly compressing the number of network parameters and computational load to meet the real-time processing requirements of security equipment at the edge. The backbone feature extraction module takes the extracted multi-scale feature maps of the image as input and performs deep feature mining layer by layer. Shallow convolutions focus on capturing fine-grained features such as image edge contours, texture details, and strong light boundaries; mid-level convolutions focus on statistically analyzing local brightness distribution, changes in brightness gradients, and the range of halo diffusion; and deep convolutions learn global scene semantics, distinguishing between ordinary scene areas and strong interference areas formed by vehicle lights and streetlights, completing the refined encoding of basic features.

[0031] The multi-scale feature fusion module is the core design of this network. Addressing the significant differences in features between high and low light regions in nighttime strong light images, it incorporates cross-scale feature stitching and adaptive weighted fusion branches. This module spatially aligns the features from different resolution levels output by the backbone network, organically fusing high-resolution detail features, medium-resolution texture features, and low-resolution global illumination features. This preserves subtle edge details of local halos and glare while also considering the global constraints of the overall image brightness distribution. It overcomes the limitation of single-scale features in simultaneously characterizing local interference and the global scene, enhancing the differentiated representation of features between effective signal areas and noisy interference areas.

[0032] The pixel-level probabilistic decoding module uses 1×1 pointwise convolution combined with a sigmoid activation function to construct the decoding output layer, mapping the fused multidimensional features to pixel-level probability values ​​in the range of 0 to 1. The image processing network undergoes supervised training on a massive amount of nighttime strong light monitoring samples, and iteratively optimizes the weights based on pixel-level labeled signal and noise labels. It can determine the confidence level of the current pixel belonging to the real scene and effective texture, and finally generate a pixel-level confidence distribution map, i.e., a signal probability map.

[0033] This image processing network features a lightweight structure adapted to hardware deployment. Through multi-scale feature mining and pixel-level probabilistic regression, it accurately quantifies the signal reliability of each pixel, providing core data support for subsequent adaptive threshold calculation and accurate separation of the signal layer and noise layer. It solves the problem from the algorithm level that traditional fixed image processing algorithms cannot adapt to complex and strong light scenes at night.

[0034] 104. Perform discrete transformations on the image to be processed and the signal probability map respectively, and extract the signal layer and noise layer from the image to be processed based on the discrete transformation results and the signal probability map.

[0035] Discrete transformation refers to multi-scale hierarchical transformation operations performed on two-dimensional images, which can decompose the entire image information into two types of information components: low-frequency sub-bands and high-frequency sub-bands. The low-frequency sub-bands carry basic information such as the overall image contour, global brightness, and main structure, while the high-frequency sub-bands carry image edges, texture details, areas of abrupt changes in brightness, and various noise interference information. The signal layer is a layer of effective visual information precisely separated from the image to be processed. It contains real and useful image content such as real scene contours, object textures, and normal lighting, fully preserving the original texture details of strong light areas and removing invalid interference such as glare and stray light. The noise layer is a layer of invalid interference information stripped from the image to be processed. It mainly includes interference content with no actual visual value, such as halos caused by strong light sources at night, glare and stray light, random noise in low light environments, and inherent dark noise of sensors. It only needs to be suppressed and weakened, without preserving detailed features.

[0036] For example, after performing a discrete transformation on the image to be processed to obtain a first low-frequency sub-band and a first high-frequency sub-band, the signal probability spectrum is then subjected to a discrete transformation using the same transformation rule to obtain a second low-frequency sub-band and a second high-frequency sub-band. Next, the correspondence between the two types of high-frequency sub-bands is established, and the image adaptation threshold is calculated.

[0037] Optionally, in some embodiments of this application, the step "perform discrete transformations on the image to be processed and the signal probability map respectively, and extract the signal layer and noise layer from the image to be processed based on the discrete transformation results and the signal probability map" may specifically include: A discrete transformation is performed on the image to be processed to obtain the low-frequency sub-band and the first high-frequency sub-band corresponding to the image to be processed; Discrete transformation is performed on the signal probability spectrum to obtain the second high-frequency sub-band corresponding to the signal probability spectrum; The image adaptation threshold is calculated based on the first and second high-frequency sub-bands. Based on the first high-frequency sub-band and the image adaptation threshold, the signal layer is extracted from the image to be processed; Based on the signal layer and low-frequency subband, a noise layer is extracted from the image to be processed.

[0038] The low-frequency sub-band refers to the component obtained by discrete transformation of the image to be processed, carrying information about the overall scene outline, global brightness, and subject layout of the image. The first high-frequency sub-band is the detail component obtained after the image to be processed is transformed, containing details such as image texture, edges, strong light glare, and dark area noise. Similarly, the second high-frequency sub-band is the detail component obtained by discrete transformation of the signal probability spectrum, reflecting the confidence level of different areas of the image as valid real-world signals. The image adaptation threshold is a judgment threshold adaptively calculated by combining the features of the two sets of high-frequency sub-bands, which can distinguish between valid image details and strong light interference information.

[0039] For example, a discrete transformation operation is performed on the acquired image to be processed, splitting it into a low-frequency sub-band and a first high-frequency sub-band. Next, a unified transformation method is used to perform a discrete transformation on the generated signal probability map, yielding a second high-frequency sub-band. Then, the spatial location information of the first and second high-frequency sub-bands is matched, and an image adaptation threshold is calculated by combining their feature data. Finally, the first high-frequency sub-band is optimized based on the image adaptation threshold, and then combined with the low-frequency sub-band to complete image reconstruction, extracting the signal layer from the image to be processed; and, based on the extracted signal layer combined with the low-frequency sub-band, the noise layer is extracted from the image to be processed.

[0040] Specifically, taking nighttime surveillance images of park entrances and exits as an example, the images show strong glare from the headlights of passing vehicles, and significant image noise in dark areas near the entrances and exits. First, the original surveillance image of the park is discretized to obtain a low-frequency sub-band that reflects the overall layout of the park's buildings and roads, and a first high-frequency sub-band containing vehicle headlight shadows, wall textures, and image noise. Second, the signal probability spectrum corresponding to this scene is discretized to obtain a second high-frequency sub-band that distinguishes between the real-world area and the area with strong light interference. Then, by comparing the data characteristics of the two sets of high-frequency sub-bands, an image adaptation threshold suitable for nighttime strong vehicle headlight scenes is calculated. This threshold is used to filter out the strong light interference components in the first high-frequency sub-band, and combined with the low-frequency sub-band, a clear, complete signal layer without glare interference is reconstructed. Finally, by combining the signal layer and the low-frequency sub-band, a noise layer composed of vehicle headlight halos and noise in dark areas is separated, completing the image information extraction.

[0041] Optionally, in some embodiments of this application, the step "extracting the signal layer from the image to be processed based on the first high-frequency sub-band and the image adaptation threshold" may specifically include: Obtain the first preset formula; The first high-frequency sub-band is processed according to the first preset formula and the image adaptation threshold to obtain the target high-frequency sub-band; Based on the target high-frequency subband and the first low-frequency subband, the signal layer is extracted from the image to be processed.

[0042] The first preset formula is a standardized calculation rule pre-set and adapted to the high-frequency component optimization processing of the image. It is used to filter and correct the high-frequency sub-band coefficients. It can distinguish effective texture details from strong light interference components by combining the image adaptation threshold, and is the fixed calculation basis for realizing adaptive correction of high-frequency information. The target high-frequency sub-band refers to the optimized high-frequency data component obtained after the original first high-frequency sub-band is processed by the first preset formula and the image adaptation threshold. Invalid interference high-frequency information such as glare, halo, and stray light in the image is eliminated, and only effective high-frequency features with practical value such as object edges and physical textures are retained.

[0043] For example, the first preset formula is: C(x, y) is the original high-frequency coefficient of the first high-frequency subband at pixel coordinates (x, y), and sign(·) is the sign function, which can preserve the positive and negative attributes of the high-frequency coefficients to ensure that the texture edge direction remains unchanged; T signal It is the calculated image adaptation threshold, P. signal C is the signal probability value corresponding to the same coordinate position in the signal probability map. signal (x, y) are the target high-frequency subband coefficients. Specifically, the suppression level can be adaptively adjusted using the signal probability value, i.e., the signal probability P at the pixel location. signal The higher the value, the better. signalThe smaller the value, the smaller the threshold reduction, and the more effective texture details are preserved; pixel position signal probability P signal The lower the level (in areas of strong light, halo, and dark noise), the more 1-P signal The larger the value, the greater the threshold reduction, and the more the interference components are suppressed; ultimately, only the effective high-frequency information above the adaptive correction threshold is retained, and negative values ​​are directly set to zero to eliminate invalid disturbances.

[0044] Furthermore, the image adaptation threshold and the corresponding probability value of the signal probability spectrum are substituted into the formula; the original coefficients of the first high-frequency sub-band are traversed pixel by pixel to obtain the target high-frequency sub-band; the target high-frequency sub-band and the first low-frequency sub-band are subjected to inverse discrete transformation to reconstruct and extract the complete signal layer. For example, in a nighttime intersection monitoring image, the vehicle body P signal Road marking area P signal Approaching 1, the correction threshold is small, and the original high-frequency coefficients of the texture are basically preserved; vehicle headlight glare P signal , Noise area in dark road surface P signal Approaching 0, the correction threshold increases significantly, effectively weakening and suppressing the high-frequency coefficient of such interference; after processing by the first preset formula, a pure target high-frequency sub-band is obtained, which, combined with low-frequency contour information, synthesizes an image signal layer without strong light interference and with intact details.

[0045] Optionally, in some embodiments of this application, the step "calculating the image adaptation threshold based on the first high-frequency sub-band and the second high-frequency sub-band" may specifically include: Determine the second high-frequency sub-band corresponding to the coefficients of each first high-frequency sub-band; Based on the second preset formula, the first high-frequency sub-band, the second high-frequency sub-band, and the correlation between the coefficients of each first high-frequency sub-band and the second high-frequency sub-band, the image adaptation threshold is calculated.

[0046] The coefficients of the first high-frequency sub-band refer to the numerical values ​​corresponding to the pixel coordinates (x, y) within the first high-frequency sub-band after the image to be processed undergoes discrete transformation. These coefficients characterize the strength of detail information such as edge texture, abrupt changes in brightness, strong light glare, and dark area noise at that location, and are the basic numerical units at the high-frequency level. The correlation is a one-to-one mapping and matching relationship between pixel spatial coordinates; that is, the first and second high-frequency sub-bands have the same resolution, and the same horizontal and vertical coordinate (x, y) positions correspond to each other, achieving pairing of high-frequency detail values ​​and probability confidence feature values. The second preset formula is a fixed calculation formula used to fuse the two types of high-frequency data and solve the image adaptive threshold. Based on the paired high-frequency values ​​and probability features, combined with global distribution patterns, it completes the adaptive threshold solution.

[0047] In some embodiments of this application, the second preset formula may specifically be: τ1 is a preset adjustable balance parameter, with a default value of 0.8; median(|C|) is the median of the absolute values ​​of all coefficients in the first high-frequency subband; P signal T represents the signal probability characteristic value within the second high-frequency sub-band that corresponds to the same coordinate position as the first high-frequency sub-band. signal To obtain the image adaptation threshold.

[0048] For example, based on the one-to-one correspondence of pixel coordinates, the probability feature value of the second high-frequency sub-band corresponding to each first high-frequency sub-band coefficient at the same spatial location is determined. Then, all first high-frequency sub-band coefficients are statistically analyzed, and the median(|C|) of the absolute values ​​of all coefficients is calculated. Next, a preset second preset formula is called to combine the median result, adjustable parameters, and paired P... signal Substituting into the second preset formula, the image adaptation threshold T is calculated. signal .

[0049] Taking a nighttime intersection monitoring scenario with strong light as an example, the coefficients of the first high-frequency sub-band include the vehicle headlight halo interference coefficient and the road real-scene texture coefficient; the signal probability value of the corresponding position in the second high-frequency sub-band is matched according to the coordinate correspondence, and the real-scene area P signal The value is too high, indicating strong light interference in the P region. signal The numerical value is too low; after statistically obtaining the median absolute value of the coefficients of the first high-frequency sub-band, the default parameter 0.8 is substituted to complete the calculation, generating an image adaptation threshold T suitable for the current strong light scene. signal .

[0050] Optionally, in some embodiments of this application, the step "extracting the noise layer from the image to be processed based on the signal layer and the low-frequency subband" may specifically include: The initial noise layer is obtained by performing a difference operation between the image to be processed and the signal layer; The initial noise layer is filtered based on the signal probability spectrum, and the noise layer is extracted from the image to be processed.

[0051] The initial noise layer is a layer obtained by performing pixel difference calculations between the original image to be processed and the separated signal layer. This initial noise layer mainly contains various image interference information such as headlight halo, streetlight glare, dark area grain noise, and sensor dark current noise. It also contains a small amount of weak and effective details of the real scene that have not been completely stripped away. It is an unfiltered and unpurified original noise layer.

[0052] For example, specifically, the entire image to be processed is compared pixel by pixel with the extracted signal layer using grayscale difference calculations to strip away the effective real-world content of the image. The resulting initial noise layer can be obtained through I... noise = I raw - I signal Perform calculations, where I noise It is the initial noise layer, Iraw It is the original image, I signal This is the signal layer. Since the noise obtained by the subtraction still contains a small amount of signal residue, the initial noise layer is discretized to decompose it into a low-frequency noise sub-band and a high-frequency noise sub-band. Next, the signal probability map corresponding to the same scene is retrieved, and based on the pixel confidence distribution rules of the map, adaptive weighted filtering is performed on the high-frequency noise sub-band to remove residual effective scene details in the noise layer. Finally, the filtered and optimized high-frequency and low-frequency noise sub-bands are subjected to inverse discretization to reconstruct the image, ultimately extracting a clean noise layer from the image to be processed. Optionally, in some embodiments of this application, the step "filtering the initial noise layer based on the signal probability map and extracting the noise layer from the image to be processed" may specifically include: The initial noise layer is discretized to obtain the low-frequency noise sub-band and high-frequency noise sub-band corresponding to the initial noise layer.

[0053] The high-frequency noise subband is obtained by weighted filtering using signal probability spectrum; The high-frequency and low-frequency noise subbands are subjected to inverse discrete transformation to extract the noise layer from the image to be processed.

[0054] 105. Process the image to be processed according to the signal layer and noise layer.

[0055] For example, specifically, the processing levels of different regions are divided according to the signal probability map to determine the signal layer enhancement amplitude and the noise layer suppression amplitude. Then, the separated signal layer is optimized for detail gain, preserving the original real-world outline, edge texture, and normal brightness information, thus enhancing the effective image detail. Next, the noise layer is subjected to global adaptive attenuation processing according to preset suppression rules, focusing on reducing interference information such as halos and glare caused by strong light and particle noise in low-light areas. Finally, the detail-optimized signal layer and the attenuated noise layer are fused and superimposed at the pixel level to reconstruct and integrate a complete image. Optionally, in some embodiments of this application, the fused image can be calibrated for overall brightness and grayscale equalization to eliminate the difference between light and dark areas.

[0056] Optionally, in some embodiments of this application, the step "processing the image to be processed according to the signal layer and the noise layer" may specifically include: Edge sharpening and texture restoration are performed on the signal layer to obtain the enhanced signal layer; The adaptive gain coefficient is calculated based on the signal probability spectrum, and the noise layer is weighted and attenuated using the adaptive gain coefficient to obtain the attenuated noise layer. The target image is obtained by fusing the low-frequency information, the signal enhancement layer, and the noise attenuation layer of the image to be processed.

[0057] Edge sharpening involves gradient enhancement of object contours, scene boundaries, and line edges within the signal layer, brightening edge pixel differences and improving blurry edges and outlines in nighttime imaging, clearly delineating the boundaries of entities such as pedestrians, vehicles, and buildings. Texture restoration fills in the subtle surface textures, ground patterns, and object material details lost in the signal layer due to strong light obstruction or low-light imaging, restoring the original real-world information of the image and repairing lost details. The enhanced signal layer is an effective image layer with clear contours, complete details, and higher image quality after edge sharpening and texture restoration, serving as the core foundation layer for reconstructing high-quality images. The adaptive gain coefficient is a dynamically adjusted coefficient calculated based on the confidence level of the signal probability spectrum pixel by pixel. Its value changes in real time with scene lighting and interference levels, used to flexibly adjust the noise suppression intensity. Weighted attenuation differs from uniform intensity noise reduction; it uses the adaptive gain coefficient to differentially weaken noise in different regions of the noise layer, increasing attenuation in areas with strong light interference and moderately weakening noise in areas with slight interference. The noise attenuation layer is a low-interference noise layer that is formed after the noise layer has undergone adaptive weighted attenuation processing, effectively suppressing and weakening interference information such as headlight halo, glare, and dark area particle noise.

[0058] For example, edge sharpening is performed on the extracted raw signal layer to enhance the edge contours of various entities in the image and improve the sense of depth. Based on the sharpening, texture restoration processing is performed on the signal layer to fill in missing subtle details of the real scene, resulting in an enhanced signal layer. Next, the global signal probability map is retrieved, and the corresponding adaptive gain coefficient is calculated pixel-by-pixel based on the confidence values ​​of each pixel signal within the map. Then, the calculated adaptive gain coefficient is used to perform global weighted attenuation processing on the clean noise layer to suppress various image interference information, resulting in an attenuated noise layer. Simultaneously, the complete low-frequency global information retained after the discrete transformation of the image to be processed is extracted. Finally, the low-frequency information of the image, the enhanced signal layer, and the attenuated noise layer are pixel-level fused and superimposed to generate the target image.

[0059] Taking nighttime road monitoring footage as an example, firstly, the signal layer containing vehicles and roads is sharpened to enhance the vehicle body and road edges, and the texture of road surface gravel and markings is restored to obtain an enhanced signal layer; then, the gain coefficient is calculated based on the signal probability, and the noise layer containing headlight glare and dark area noise is differentially suppressed to obtain an attenuated noise layer; finally, the low-frequency information of the overall layout of the image is combined to complete the fusion and obtain the target image.

[0060] The multi-layer bond structure is a bond structure designed for image processing scenarios in this application embodiment. Before normalization, this application embodiment provides an image processing method. After extracting multi-scale feature maps of the image to be processed, the acquisition scene corresponding to the image to be processed is determined according to the brightness distribution information corresponding to the multi-scale feature maps. If the acquisition scene is a preset scene, a signal probability map corresponding to the image to be processed is generated according to a preset image processing network and the multi-scale feature maps. Then, discrete transformations are performed on the image to be processed and the signal probability map, respectively. Based on the discrete transformation results and the signal probability map, a signal layer and a noise layer are extracted from the image to be processed. Finally, the image to be processed is processed according to the signal layer and the noise layer. In the image processing scheme provided in this application, based on multi-scale features... The brightness distribution information of the feature map determines the image acquisition scene, specifically identifying preset scenes such as strong light at night. This eliminates the need for high-cost HDR sensors and multi-exposure fusion hardware solutions. Simultaneously, a signal probability map is generated by combining a preset image processing network with multi-scale feature maps. Then, discrete transformations are performed on the image to be processed and the signal probability map, accurately separating the signal layer and noise layer of the image based on the transformation results. This achieves decoupling and differentiation of signal and noise. Subsequent differentiated processing is then performed based on the signal layer and noise layer. This effectively improves the problems of local overexposure and glare caused by strong light at night, thus perfectly preserving the texture details in the highlight areas. At the same time, it suppresses noise in the low-light areas and restores lost image details, effectively solving the problem that traditional technologies cannot balance the preservation of highlight details and the suppression of noise in dark areas.

[0061] To facilitate better implementation of the image processing method of the embodiments of this application, the embodiments of this application also provide an image processing apparatus, wherein the meanings of the terms are the same as those in the image processing method described above, and specific implementation details can be found in the description of the system embodiments.

[0062] Please see Figure 2 , Figure 2 This is a schematic diagram of the structure of an image processing device provided in an embodiment of this application. The image processing device may specifically include an extraction module 201, a determination module 202, a generation module 203, a transformation module 204, and a processing module 205, as follows: Extraction module 201 is used to extract multi-scale feature maps of the image to be processed; The determination module 202 is used to determine the acquisition scene corresponding to the image to be processed based on the brightness distribution information corresponding to the multi-scale feature map. The generation module 203 is used to generate a signal probability map corresponding to the image to be processed based on the preset image processing network and the multi-scale feature map if the acquisition scene is a preset scene. The transformation module 204 is used to perform discrete transformations on the image to be processed and the signal probability spectrum respectively, and extract the signal layer and noise layer from the image to be processed based on the discrete transformation results and the signal probability spectrum. The processing module 205 is used to process the image to be processed according to the signal layer and the noise layer.

[0063] Optionally, in some embodiments of this application, the transformation module 204 may specifically include: The transformation unit is used to perform discrete transformation on the image to be processed to obtain the low-frequency sub-band and the first high-frequency sub-band corresponding to the image to be processed, and to perform discrete transformation on the signal probability spectrum to obtain the second high-frequency sub-band corresponding to the signal probability spectrum. The calculation unit is used to calculate the image adaptation threshold based on the first high-frequency sub-band and the second high-frequency sub-band; The first extraction unit is used to extract the signal layer in the image to be processed based on the first high-frequency sub-band and the image adaptation threshold. The second extraction unit is used to extract the noise layer from the image to be processed based on the signal layer and the low-frequency subband.

[0064] Optionally, in some embodiments of this application, the first extraction unit may specifically be used for: Obtain the first preset formula; The first high-frequency sub-band is processed according to the first preset formula and the image adaptation threshold to obtain the target high-frequency sub-band; Based on the target high-frequency subband and the first low-frequency subband, the signal layer is extracted from the image to be processed.

[0065] Optionally, in some embodiments of this application, the computing unit may specifically be used for: Determine the second high-frequency sub-band corresponding to the coefficients of each first high-frequency sub-band; Based on the second preset formula, the first high-frequency sub-band, the second high-frequency sub-band, and the correlation between the coefficients of each first high-frequency sub-band and the second high-frequency sub-band, the image adaptation threshold is calculated.

[0066] Optionally, in some embodiments of this application, the second extraction unit may specifically be used for: The initial noise layer is obtained by performing a difference operation between the image to be processed and the signal layer; The initial noise layer is filtered based on the signal probability spectrum, and the noise layer is extracted from the image to be processed.

[0067] Optionally, in some embodiments of this application, the second extraction unit may specifically be used for: The initial noise layer is discretized to obtain the low-frequency noise sub-band and high-frequency noise sub-band corresponding to the initial noise layer. The high-frequency noise subband is obtained by weighted filtering using signal probability spectrum; The high-frequency and low-frequency noise subbands are subjected to inverse discrete transformation to extract the noise layer from the image to be processed.

[0068] Optionally, in some embodiments of this application, the processing module 205 may specifically be used for: The signal layer is subjected to edge sharpening and texture restoration processing to obtain an enhanced signal layer; The adaptive gain coefficient is calculated based on the signal probability spectrum, and the noise layer is weighted and attenuated using the adaptive gain coefficient to obtain the attenuated noise layer. The target image is obtained by fusing the low-frequency information, the signal enhancement layer, and the noise attenuation layer of the image to be processed.

[0069] This application provides an image processing apparatus. An extraction module 201 extracts a multi-scale feature map of the image to be processed. A determination module 202 determines the acquisition scene corresponding to the image to be processed based on the brightness distribution information corresponding to the multi-scale feature map. A generation module 203, if the acquisition scene is a preset scene, generates a signal probability map corresponding to the image to be processed based on a preset image processing network and the multi-scale feature map. Next, a transformation module 204 performs discrete transformations on the image to be processed and the signal probability map, and extracts a signal layer and a noise layer from the image to be processed based on the discrete transformation results and the signal probability map. Finally, a processing module 205 processes the image to be processed based on the signal layer and the noise layer. In the image processing scheme provided in this application, based on the multi-scale feature map… The brightness distribution information of the feature map determines the image acquisition scene, specifically identifying preset scenes such as strong light at night. This eliminates the need for high-cost HDR sensors and multi-exposure fusion hardware solutions. Simultaneously, a signal probability map is generated by combining a preset image processing network with multi-scale feature maps. Then, discrete transformations are performed on the image to be processed and the signal probability map, accurately separating the signal layer and noise layer of the image based on the transformation results. This achieves decoupling and differentiation of signal and noise. Subsequent differentiated processing is then performed based on the signal layer and noise layer. This effectively improves the problems of local overexposure and glare caused by strong light at night, thus perfectly preserving the texture details in the highlight areas. At the same time, it suppresses noise in the low-light areas and restores lost image details, effectively solving the problem that traditional technologies cannot balance the preservation of highlight details and the suppression of noise in dark areas.

[0070] Furthermore, embodiments of this application also provide a computing device, such as... Figure 3 As shown, it illustrates a schematic diagram of the computing device involved in the embodiments of this application, specifically: The computing device may include: a processor 402, a communications interface 404, a memory 406, and a communications bus 408.

[0071] The processor 402, communication interface 404, and memory 406 communicate with each other via communication bus 408. Communication interface 404 is used to communicate with other network elements such as clients or other servers. The processor 402 executes program 410, specifically performing the relevant steps in the above-described embodiment of the bandwidth acquisition method for computing devices.

[0072] Specifically, program 410 may include program code that includes computer operation instructions.

[0073] Processor 402 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. The computing device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0074] Memory 406 is used to store program 410. Memory 406 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0075] Specifically, program 410 can be used to cause processor 402 to execute the bandwidth acquisition method in any of the above method embodiments. The specific implementation of each step in program 410 can be found in the corresponding descriptions of the steps and units in the above bandwidth acquisition embodiments, and will not be repeated here. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the devices and modules described above can be referred to the corresponding process descriptions in the foregoing method embodiments, and will not be repeated here.

[0076] The algorithms and displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing such systems is apparent from the above description. Furthermore, the embodiments of this application are not directed to any particular programming language. It should be understood that the contents of the embodiments of this application described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best implementation of the embodiments of this application.

[0077] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination can be used to combine all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or device so disclosed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.

[0078] Furthermore, those skilled in the art will understand that although some embodiments described herein include certain features but not others included in other embodiments, combinations of features from different embodiments are meant to be within the scope of the embodiments of this application and form different embodiments. For example, in the following claims, any one of the claimed embodiments can be used in any combination.

[0079] The various component embodiments of this application can be implemented in hardware, or as software modules running on one or more processors, or a combination thereof. Those skilled in the art will understand that microprocessors or digital signal processors (DSPs) can be used in practice to implement some or all of the functions of some or all of the components according to the embodiments of this application. The embodiments of this application can also be implemented as device or apparatus programs (e.g., computer programs and computer program products) for performing part or all of the methods described herein. Such programs implementing the embodiments of this application can be stored on a computer-readable medium, or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, provided on a carrier signal, or provided in any other form: Therefore, embodiments of this application provide a computer-readable storage medium storing a plurality of instructions that can be loaded by a processor to execute steps in any of the image processing methods provided in embodiments of this application. For example, the instructions can execute the following steps: Multi-scale feature maps of the image to be processed are extracted. Based on the brightness distribution information corresponding to the multi-scale feature maps, the acquisition scene corresponding to the image to be processed is determined. If the acquisition scene is a preset scene, a signal probability map corresponding to the image to be processed is generated based on the preset image processing network and the multi-scale feature maps. Discrete transformations are performed on the image to be processed and the signal probability map respectively. Based on the discrete transformation results and the signal probability map, the signal layer and noise layer are extracted from the image to be processed. The image to be processed is then processed according to the signal layer and the noise layer.

[0080] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0081] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0082] Since the instructions stored in the computer-readable storage medium can execute the steps of any of the image processing methods provided in the embodiments of this application, the beneficial effects that any of the image processing methods provided in the embodiments of this application can achieve can be realized, as detailed in the preceding embodiments, and will not be repeated here.

[0083] The foregoing has provided a detailed description of an image processing method, apparatus, computing device, and computer-readable storage medium provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. An image processing method, characterized in that, include: Extract multi-scale feature maps from the image to be processed; Based on the brightness distribution information corresponding to the multi-scale feature map, the acquisition scene corresponding to the image to be processed is determined; If the acquisition scenario is a preset scenario, then a signal probability map corresponding to the image to be processed is generated based on the preset image processing network and the multi-scale feature map. Discrete transformations are performed on the image to be processed and the signal probability spectrum, respectively, and the signal layer and noise layer are extracted from the image to be processed based on the discrete transformation results and the signal probability spectrum. The image to be processed is processed according to the signal layer and the noise layer.

2. The image processing method according to claim 1, characterized in that, The step of performing discrete transformations on the image to be processed and the signal probability map, and extracting the signal layer and noise layer from the image to be processed based on the discrete transformation results and the signal probability map, includes: The image to be processed is subjected to a discrete transformation to obtain the low-frequency sub-band and the first high-frequency sub-band corresponding to the image to be processed; The signal probability spectrum is discretized to obtain the second high-frequency sub-band corresponding to the signal probability spectrum; Based on the first high-frequency sub-band and the second high-frequency sub-band, calculate the image adaptation threshold; Based on the first high-frequency sub-band and the image adaptation threshold, the signal layer is extracted from the image to be processed; Based on the signal layer and the low-frequency subband, a noise layer is extracted from the image to be processed.

3. The image processing method according to claim 2, characterized in that, The step of extracting a signal layer from the image to be processed based on the first high-frequency sub-band and the image adaptation threshold includes: Obtain the first preset formula; The first high-frequency sub-band is processed according to the first preset formula and the image adaptation threshold to obtain the target high-frequency sub-band; Based on the target high-frequency subband and the first low-frequency subband, a signal layer is extracted from the image to be processed.

4. The image processing method according to claim 2, characterized in that, The calculation of the image adaptation threshold based on the first high-frequency sub-band and the second high-frequency sub-band includes: Determine the second high-frequency sub-band corresponding to the coefficients of each first high-frequency sub-band; Based on the second preset formula, the first high-frequency sub-band, the second high-frequency sub-band, and the correlation between the coefficients of each first high-frequency sub-band and the second high-frequency sub-band, the image adaptation threshold is calculated.

5. The image processing method according to claim 2, characterized in that, The step of extracting a noise layer from the image to be processed based on the signal layer and the low-frequency sub-band includes: The initial noise layer is obtained by performing a difference operation between the image to be processed and the signal layer. The initial noise layer is filtered based on the signal probability spectrum, and the noise layer is extracted from the image to be processed.

6. The image processing method according to claim 5, characterized in that, The step of filtering the initial noise layer based on the signal probability map and extracting the noise layer from the image to be processed includes: The initial noise layer is subjected to discrete transformation to obtain the low-frequency noise sub-band and high-frequency noise sub-band corresponding to the initial noise layer; The high-frequency noise sub-band is obtained by weighting and filtering the signal probability spectrum. The high-frequency noise subband and the low-frequency noise subband are subjected to inverse discrete transformation to extract the noise layer from the image to be processed.

7. The image processing method according to any one of claims 1 to 6, characterized in that, The process of processing the image to be processed based on the signal layer and the noise layer includes: The signal layer is subjected to edge sharpening and texture restoration processing to obtain an enhanced signal layer; The adaptive gain coefficient is calculated based on the signal probability spectrum, and the noise layer is weighted and attenuated using the adaptive gain coefficient to obtain the attenuated noise layer. The target image is obtained by fusing the low-frequency information of the image to be processed, the enhanced signal layer, and the noise attenuation layer.

8. An image processing apparatus, characterized in that, include: The extraction module is used to extract multi-scale feature maps from the image to be processed; The determination module is used to determine the acquisition scene corresponding to the image to be processed based on the brightness distribution information corresponding to the multi-scale feature map; The generation module is used to generate a signal probability map corresponding to the image to be processed based on a preset image processing network and the multi-scale feature map if the acquisition scene is a preset scene. The transformation module is used to perform discrete transformations on the image to be processed and the signal probability spectrum respectively, and extract the signal layer and noise layer from the image to be processed based on the discrete transformation results and the signal probability spectrum. The processing module is used to process the image to be processed based on the signal layer and the noise layer.

9. A computing device, characterized in that, include: At least one processor; and A memory that is communicatively connected to the at least one processor; The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the steps of the image processing method according to any one of claims 1 to 14.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a processor, implement the steps of the image processing method as described in any one of claims 1 to 14.

Images