Daytime effect night vision device low-illumination image dynamic noise reduction method based on image recognition

Abstract

The invention discloses a low-illumination image dynamic noise reduction method for a daytime effect night vision device based on image recognition, and the method comprises the following steps: obtaining a low-illumination image sequence, and determining a current frame as a target frame; performing noise level estimation and image recognition on the target frame to obtain a noise parameter and a recognition confidence map; generating an initial edge set based on the target frame, and determining an initial edge weight; performing semantic boundary constraint reconstruction on the initial edge set, and outputting a reconstructed edge set; performing Felzenszwalb graph segmentation on the reconstructed edge set, sorting the reconstructed edge set according to the reconstructed edge weight, taking edges in sequence according to the sorting sequence, and outputting a final region label graph and a segmented region set according to a merging criterion; determining a region-level noise reduction intensity parameter for each region, and generating a noise reduction control chart; and performing regional dynamic noise reduction processing on the target frame according to the noise reduction control chart to obtain a noise reduction output frame. According to the method, the target edge and detail fidelity is improved, and the excessive smoothing risk of cross-semantic boundaries is reduced.

Description

Technical Field

[0001] This invention relates to the fields of image processing and machine vision technology, and in particular to a method for dynamic noise reduction of low-light images of daytime night vision devices based on image recognition. Background Technology

[0002] Day-effect night vision devices need to work continuously day and night in scenarios such as security patrol, vehicle-mounted assisted driving, border monitoring and emergency search and rescue. At night or in low light conditions, the number of incident photons of the imaging sensor is insufficient, and the target frame often exhibits characteristics such as low signal-to-noise ratio, weak detail contrast and obvious noise particles. Existing systems usually introduce noise reduction processing at the sensor end or image processing end to suppress random noise and improve visibility. Common practices include spatial domain smoothing, edge-preserving filtering or multi-scale denoising processing on single-frame low-light images. Some solutions also introduce segmentation or superpixel concepts, dividing the image into several regions and then implementing noise reduction adjustment of the partition parameters in order to achieve a trade-off between smoothing noise and preserving structure.

[0003] Existing technologies still face several problems under low-light conditions. Many noise reduction methods use globally fixed parameters or only adapt to the overall noise level, lacking differentiated control over different semantic regions. This can easily lead to over-smoothing in the target region, resulting in loss of target edges and textures, or insufficient noise reduction in the background region, leaving residual noise interference. Some regionalization schemes rely on traditional segmentation or simple similarity graph construction. During the segmentation process, image recognition information is not introduced to constrain semantic boundaries. Region merging may cross the boundary between the target and the background, leading to "mis-partitioning and mis-controlling of intensity" in the subsequent region-level noise reduction intensity mapping. This manifests as the target contour being swallowed by the background or background noise being brought into the target region.

[0004] Therefore, how to provide a method for dynamic noise reduction of low-light images of daytime night vision devices based on image recognition is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a dynamic noise reduction method for low-light images of daytime night vision devices based on image recognition. This invention introduces noise level estimation and image recognition confidence into the graph construction and segmentation process. By reconstructing the edge set through semantic boundary constraints and combining it with Felzenszwalb segmentation to obtain stable region labels, a noise reduction control map is generated according to the region noise intensity and recognition confidence to achieve region-level adaptive noise reduction, which balances noise suppression and target edge detail preservation, and reduces excessive smoothing across semantic boundaries and region mismerging.

[0006] The image recognition-based dynamic noise reduction method for low-light images of daytime night vision devices according to embodiments of the present invention includes the following steps:

[0007] Acquire the low-light image sequence output by the daytime night vision device and determine the current frame as the target frame;

[0008] Noise level estimation and image recognition are performed on the target frame to obtain noise parameters and recognition confidence maps;

[0009] Based on the target frame, an initial edge set is generated according to a preset connection rule, and the initial edge weights are determined according to the differences in pixel features.

[0010] Based on the recognition confidence graph, the initial edge set is reconstructed using semantic boundary constraints. A confidence threshold is set and the edges are retained. The initial edge weights of the retained edges are used as the reconstructed edge weights, and the reconstructed edge set is output.

[0011] Perform Felzenszwalb graph segmentation on the reconstructed edge set, sort the reconstructed edge set according to the reconstructed edge weight, initialize each pixel as an independent region, take edges in the sorted order and perform merging or preservation on the regions to which the two ends of the edge belong according to the merging criterion, and output the final region label map and segmented region set after traversal.

[0012] Using the segmented region set as the region unit, the region-level noise reduction intensity parameters are determined for each region by combining the identification confidence map and noise parameters, and the region-level noise reduction intensity parameters of each region are mapped to the final region label map to generate a noise reduction control map.

[0013] Based on the noise reduction control chart, the target frame is subjected to regional dynamic noise reduction processing to obtain the noise-reduced output frame.

[0014] Optionally, obtaining the target frame specifically includes:

[0015] The imaging sensor driving the daytime night vision device outputs low-light image frames during continuous sampling. Each low-light image frame is written into the image buffer in chronological order to form a low-light image sequence.

[0016] Read the low-light image frame with the latest frame number and the largest sampling timestamp from the low-light image sequence, use it as the current frame, mark the current frame as the target frame, record the frame number and sampling timestamp corresponding to the target frame, and write the target frame into the target frame buffer.

[0017] Optionally, obtaining the noise parameters and the identification confidence map specifically includes:

[0018] Luminance channel extraction is performed on the target frame to obtain a luminance map, and a one-to-one correspondence between the pixel positions of the target frame is established.

[0019] A smoothed brightness map is obtained by performing a smoothing filter on the brightness map. The brightness map and the smoothed brightness map are then subtracted by pixel position to obtain a residual map. At the same time, the absolute value of each pixel in the residual map is taken to form an absolute residual set. Based on the absolute residual set, noise parameters characterizing the noise level of the target frame are calculated.

[0020] The target frame is input into the image recognition model that has been pre-trained and deployed on the daytime night vision device. Forward inference is performed to obtain a recognition confidence map with the same pixel size as the target frame.

[0021] Optionally, the process of generating an initial edge set based on the target frame and according to a preset connection rule, and determining the initial edge weights based on pixel feature differences, specifically includes:

[0022] Traverse all pixel positions of the target frame and form a set of pixel positions. For each pixel position in the set of pixel positions, establish a corresponding pixel node to obtain a set of pixel nodes.

[0023] An initial edge set is generated for the set of pixel nodes according to a preset connection rule. The preset connection rule is the eight-neighbor connection rule. For each pixel position in the set of pixel positions, the pixel positions adjacent to each other in the horizontal, vertical and diagonal directions are determined. The pixel positions are paired with the pixel nodes corresponding to each of the eight adjacent pixel positions to generate edges. All generated edges are summarized to obtain the initial edge set.

[0024] Extract pixel features from the target frame, calculate the brightness gradient of the brightness map of the target frame according to the pixel position to obtain the gradient map, and make the brightness map, gradient map and pixel position set correspond one-to-one at the pixel position;

[0025] For each edge in the initial edge set, read the brightness values ​​of the pixel positions at both ends of the current edge in the brightness map and calculate the absolute value of the brightness difference. At the same time, read the gradient values ​​of the pixel positions at both ends of the current edge in the gradient map and calculate the magnitude of the gradient difference. Multiply the absolute value of the brightness difference and the magnitude of the gradient difference by the first weight coefficient and the second weight coefficient respectively, and sum them to obtain the initial edge weight of the current edge. Combine the pixel node set, the initial edge set, and the initial edge weight to form a pixel map.

[0026] Optionally, obtaining the reconstructed edge set specifically includes:

[0027] Read the initial edge set, initial edge weight set, identify the confidence graph, set the confidence threshold, and initialize the reconstructed edge set to empty;

[0028] Traverse each edge in the initial edge set in order, read the pixel positions at both ends of the current edge, read the confidence values ​​of the pixel positions at both ends in the recognition confidence map, calculate the difference between the confidence values ​​at both ends and take the absolute value of the difference to obtain the absolute value of the confidence difference of the current edge.

[0029] The initial edge set is reconstructed using semantic boundary constraints. The absolute value of the confidence difference is compared with the confidence threshold. When the absolute value of the confidence difference is less than the confidence threshold, the current edge is retained and written into the reconstructed edge set. At the same time, the initial edge weight corresponding to the current edge is read from the initial edge weight set and used as the corresponding reconstructed edge weight. When the absolute value of the confidence difference is greater than or equal to the confidence threshold, the current edge is deleted.

[0030] After traversing the initial edge set, perform deduplication and consistency checks on the reconstructed edge set to ensure that each edge in the reconstructed edge set has a unique corresponding edge weight in the reconstructed edge weight set. Output the reconstructed edge set and the corresponding reconstructed edge weights.

[0031] Optionally, the output of the final region label map and the segmented region set specifically includes:

[0032] Traverse the set of reconstructed edges and read the pixel positions at both ends of each edge and the corresponding reconstructed edge weight. Sort all edges in ascending order of reconstructed edge weight to obtain an ordered edge sequence, and record the edge number for each edge in the ordered edge sequence.

[0033] Perform Felzenszwalb graph segmentation on the reconstructed edge set, initialize Felzenszwalb region data, initialize each pixel in the target frame as an initial region and assign a unique region identifier, generate an initial region label map, and make the label of each pixel equal to the initial region identifier. At the same time, initialize the number of region pixels, the amount of difference within the region, and the minimum spanning tree edge weight set for each initial region.

[0034] Define the Felzenszwalb merge threshold term, maintain the region threshold term for each region, and set the segmentation scale parameter for the region threshold term;

[0035] Take edges sequentially according to the ordered edge sequence and locate candidate regions. Read the labels of the pixels at both ends of the current edge in the initial region label map to determine the two candidate regions to which they belong. If the pixels at both ends belong to the same region, skip the current edge. If the pixels at both ends belong to different regions, read the internal difference of the two candidate regions, the number of pixels in the region, and the region threshold, and calculate the joint threshold of the two candidate regions.

[0036] The Felzenszwalb merging criterion is executed and an improved merging decision is made. The reconstructed edge weight of the current edge is compared with the joint threshold to obtain the original merging result. When the original merging result is allowed, the improved merging decision is further executed. The improved merging decision includes a hysteresis merging strategy, which maintains a merging suppression flag for each region and sets a tight threshold factor and a loose threshold factor. When the merging suppression flag is enabled, the joint threshold is tightened using the tight threshold factor. When the merging suppression flag is disabled, the joint threshold is relaxed using the loose threshold factor. When the improved merging decision allows merging, the two candidate regions are merged and the region identifier, region pixel count, region internal difference, region minimum spanning tree edge weight set, and merging suppression flag of the merged region are updated. The pixel labels belonging to the two candidate regions in the initial region label map are updated simultaneously.

[0037] After traversing the ordered edge sequence, the merging is completed to obtain the final region label map. The pixel positions are grouped according to the region identifier of the final region label map, and each group of pixel positions is combined into a segmented region. All segmented regions are summarized to obtain the segmented region set.

[0038] Optionally, the generation of the noise reduction control map specifically includes:

[0039] Establish a region parameter table, process each region according to the segmented region set, traverse all pixel positions belonging to the current region identifier for the current region, accumulate the confidence value corresponding to the pixel position and accumulate the number of pixels, normalize the accumulated confidence value according to the number of pixels to obtain the region-level recognition confidence aggregate value, and write the region-level recognition confidence aggregate value into the region parameter table.

[0040] For the current region, traverse all pixel positions belonging to the current region identifier, accumulate the brightness value corresponding to the pixel position and accumulate the number of pixels, normalize the accumulated brightness value according to the number of pixels to obtain the region-level brightness mean, and combine the noise parameter with the region-level brightness mean to obtain the region-level noise intensity value, and write the region-level noise intensity value into the region parameter table.

[0041] For the current region, calculate the regional noise reduction intensity parameter based on the regional noise intensity value and the aggregated value of regional identification confidence, and perform upper and lower limit pruning on the regional noise reduction intensity parameter, and write the pruned regional noise reduction intensity parameter into the regional parameter table;

[0042] Traverse all pixel positions in the final region label map, read the region identifier corresponding to the pixel position, query the region-level noise reduction intensity parameter of the region identifier in the region parameter table, write it to the corresponding pixel position of the noise reduction control map, and output the noise reduction control map after completing the traversal.

[0043] Optionally, obtaining the noise-reduced output frame specifically includes:

[0044] Read the target frame and the noise reduction control map, generate a noise reduction output frame with the same pixel size as the target frame and complete pixel position alignment, and read the noise reduction control map pixel by pixel to obtain the noise reduction intensity parameter corresponding to each pixel position;

[0045] Pixel-level filtering configurations are established based on noise reduction intensity parameters. For each pixel location, the filtering window size, brightness similarity scale, and intensity gain are determined from the noise reduction intensity parameters, and the filtering configuration corresponding to each pixel location is written into the configuration mapping.

[0046] Perform regional dynamic noise reduction processing, traverse the target frame pixel by pixel and read the filter configuration in the configuration mapping for the current pixel position. Enumerate the neighboring pixel positions within the filter window of the current pixel position and read the brightness values ​​of the neighboring pixels respectively. Determine the spatial weight based on the spatial distance between the neighboring pixels and the current pixel. Determine the similarity weight based on the difference between the brightness of the neighboring pixels and the brightness of the current pixel and the brightness similarity scale. Combine the spatial weight and the similarity weight to obtain the neighboring pixel combination weight. Accumulate the brightness of the neighboring pixels according to the combination weight and normalize the accumulated value of the combination weight to obtain the noise-reduced brightness value of the current pixel. Write the noise-reduced brightness value of the current pixel to the corresponding pixel position of the noise-reduced output frame and output the noise-reduced output frame.

[0047] The beneficial effects of this invention are:

[0048] First, this invention introduces a recognition confidence map obtained from noise level estimation and image recognition at the target frame level, and incorporates low-light noise intensity and semantic information into subsequent processing at the same time, so that noise reduction control no longer depends on a single global parameter, thereby improving the adaptive capability under different scenes and different noise levels from the source, and avoiding the loss of detail or residual noise interference caused by noise reduction intensity mismatch when noise fluctuates.

[0049] This invention reconstructs the semantic boundary constraints of the initial edge set after the pixel map is constructed. It uses confidence differences to filter out connections across semantic boundaries and then combines Felzenszwalb graph segmentation to obtain region labels and segmentation region sets that better conform to semantic boundaries. This significantly reduces the probability of the target and background being incorrectly merged, thereby reducing excessive smoothing and structural adhesion across semantic boundaries.

[0050] This invention uses segmented regions as regional units, jointly identifies confidence levels and noise parameters to determine regional-level noise reduction intensity, and maps them to generate a noise reduction control map. This enables the target region and the background region to form a differentiated noise reduction strategy, which can enhance noise suppression in the background region and preserve edge and texture details in the target region, ultimately improving the recognizability and visual quality of low-light daytime nighttime images.

[0051] The noise reduction control map output by this invention maintains a consistent mapping relationship with the noise reduction output frame at the pixel position, which facilitates deployment in the real-time processing link of daytime night vision devices and achieves stable regional dynamic noise reduction output. Attached Figure Description

[0052] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0053] Figure 1 This is a flowchart of the low-light image dynamic noise reduction method for daytime night vision devices based on image recognition proposed in this invention;

[0054] Figure 2 This is a schematic diagram of the Felzenszwalb graph segmentation algorithm structure for the dynamic noise reduction method for low-light images of daytime night vision devices based on image recognition proposed in this invention. Detailed Implementation

[0055] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0056] refer to Figures 1-2 A dynamic noise reduction method for low-light images of daytime night vision devices based on image recognition includes the following steps:

[0057] Acquire the low-light image sequence output by the daytime night vision device and determine the current frame as the target frame;

[0058] Noise level estimation and image recognition are performed on the target frame to obtain noise parameters and recognition confidence maps;

[0059] Based on the target frame, an initial edge set is generated according to a preset connection rule, and the initial edge weights are determined according to the differences in pixel features.

[0060] Based on the recognition confidence graph, the initial edge set is reconstructed using semantic boundary constraints. A confidence threshold is set and the edges are retained. The initial edge weights of the retained edges are used as the reconstructed edge weights, and the reconstructed edge set is output.

[0061] Perform Felzenszwalb graph segmentation on the reconstructed edge set, sort the reconstructed edge set according to the reconstructed edge weight, initialize each pixel as an independent region, take edges in the sorted order and perform merging or preservation on the regions to which the two ends of the edge belong according to the merging criterion, and output the final region label map and segmented region set after traversal.

[0062] Using the segmented region set as the region unit, the region-level noise reduction intensity parameters are determined for each region by combining the identification confidence map and noise parameters, and the region-level noise reduction intensity parameters of each region are mapped to the final region label map to generate a noise reduction control map.

[0063] Based on the noise reduction control chart, the target frame is subjected to regional dynamic noise reduction processing to obtain the noise-reduced output frame.

[0064] In this embodiment, obtaining the target frame specifically includes:

[0065] The imaging sensor driving the daytime night vision device outputs low-light image frames during continuous sampling. Each low-light image frame is written into the image buffer in chronological order to form a low-light image sequence. Each frame in the low-light image sequence is associated with a frame number, sampling timestamp, and image data.

[0066] Read the low-light image frame with the latest frame number and the largest sampling timestamp from the low-light image sequence, use it as the current frame, mark the current frame as the target frame, record the frame number and sampling timestamp corresponding to the target frame, and write the target frame into the target frame buffer.

[0067] In this embodiment, obtaining the noise parameters and the identification confidence map specifically includes:

[0068] Luminance channel extraction is performed on the target frame to obtain a luminance map, and a one-to-one correspondence between the pixel positions of the target frame is established.

[0069] A smoothed brightness map is obtained by performing a smoothing filter on the brightness map. The brightness map and the smoothed brightness map are then subtracted by pixel position to obtain a residual map. At the same time, the absolute value of each pixel in the residual map is taken to form an absolute residual set. Based on the absolute residual set, noise parameters characterizing the noise level of the target frame are calculated.

[0070] The target frame is input into a pre-trained image recognition model deployed on a daytime night vision device. Forward inference is performed to obtain a recognition confidence map with the same pixel size as the target frame. The recognition confidence map provides the target confidence value at each pixel position, and the target confidence value is limited to the range of zero to one. The image recognition model includes an input layer, a feature extraction network, a confidence prediction head, and an output layer. The input layer receives the pixel data of the target frame and forms the model input. The feature extraction network performs multi-layer convolutional feature extraction on the model input and outputs a feature map. The confidence prediction head performs pixel-level confidence regression or classification on the feature map and generates a recognition confidence map. The output layer outputs a recognition confidence map with the same pixel size as the target frame and a value range of zero to one.

[0071] In this embodiment, the process of generating an initial edge set based on the target frame and according to a preset connection rule, and determining the initial edge weights based on pixel feature differences, specifically includes:

[0072] Traverse all pixel positions of the target frame and form a set of pixel positions. For each pixel position in the set of pixel positions, establish a corresponding pixel node to obtain a set of pixel nodes.

[0073] An initial edge set is generated for the set of pixel nodes according to the preset connection rules. The preset connection rules are the eight-neighbor connection rules. For each pixel position in the set of pixel positions, the pixel positions adjacent to each other in the horizontal, vertical and diagonal directions are determined. The pixel positions are paired with the pixel nodes corresponding to each of the eight neighboring pixel positions to generate edges. All generated edges are summarized to obtain the initial edge set.

[0074] Pixel features are extracted from the target frame, including brightness features and gradient features. The brightness gradient of the brightness map of the target frame is calculated according to the pixel position to obtain the gradient map, and the brightness map, gradient map and pixel position set are matched one-to-one at the pixel position.

[0075] For each edge in the initial edge set, read the brightness values ​​of the pixel positions at both ends of the current edge in the brightness map and calculate the absolute value of the brightness difference. At the same time, read the gradient values ​​of the pixel positions at both ends of the current edge in the gradient map and calculate the magnitude of the gradient difference. Multiply the absolute value of the brightness difference and the magnitude of the gradient difference by the first weight coefficient and the second weight coefficient respectively, and sum them to obtain the initial edge weight of the current edge. The first weight coefficient and the second weight coefficient are both preset non-zero real numbers. The first weight coefficient is used to weight the contribution of the brightness difference, and the second weight coefficient is used to weight the contribution of the gradient difference. The first weight coefficient and the second weight coefficient remain unchanged in the same target frame and are updated at the beginning of each target frame processing. The pixel node set, the initial edge set, and the initial edge weight are combined to form a pixel map.

[0076] In this embodiment, the specific steps for obtaining the reconstructed edge set include:

[0077] Read the initial edge set, initial edge weight set, identify the confidence graph, set the confidence threshold, and initialize the reconstructed edge set to empty;

[0078] Traverse each edge in the initial edge set in order, read the pixel positions at both ends of the current edge, read the confidence values ​​of the pixel positions at both ends in the recognition confidence map, calculate the difference between the confidence values ​​at both ends and take the absolute value of the difference to obtain the absolute value of the confidence difference of the current edge.

[0079] The initial edge set is reconstructed using semantic boundary constraints. The absolute value of the confidence difference is compared with the confidence threshold. When the absolute value of the confidence difference is less than the confidence threshold, the current edge is retained and written into the reconstructed edge set. At the same time, the initial edge weight corresponding to the current edge is read from the initial edge weight set and used as the corresponding reconstructed edge weight. When the absolute value of the confidence difference is greater than or equal to the confidence threshold, the current edge is deleted.

[0080] After traversing the initial edge set, perform deduplication and consistency checks on the reconstructed edge set to ensure that each edge in the reconstructed edge set has a unique corresponding edge weight in the reconstructed edge weight set. Output the reconstructed edge set and the corresponding reconstructed edge weights.

[0081] This invention utilizes a confidence graph to reconstruct the semantic boundary constraints of the initial edge set. By filtering out connections across semantic boundaries using a confidence threshold and retaining semantically consistent pixel connections, the invention retains edge edges and uses the initial edge weights to form reconstructed edge weights that correspond one-to-one with the reconstructed edge set. The edge weights are then verified for consistency through deduplication to ensure traceability. This reduces the risk of subsequent segmentation errors and oversmoothing caused by incorrect connectivity between the target and the background, and improves the stability and reliability of region segmentation and regional noise reduction control.

[0082] In this embodiment, the output of the final region label map and the segmented region set specifically includes:

[0083] Traverse the set of reconstructed edges and read the pixel positions at both ends of each edge and the corresponding reconstructed edge weight. Sort all edges in ascending order of reconstructed edge weight to obtain an ordered edge sequence, and record the edge number for each edge in the ordered edge sequence.

[0084] Perform Felzenszwalb graph segmentation on the reconstructed edge set, initialize Felzenszwalb region data, initialize each pixel in the target frame as an initial region and assign a unique region identifier, generate an initial region label map, and make the label of each pixel equal to the initial region identifier. At the same time, initialize the number of region pixels, the amount of difference within the region, and the minimum spanning tree edge weight set for each initial region.

[0085] Define the Felzenszwalb merging threshold term, maintain a region threshold term for each region, the region threshold term is determined by the intra-regional difference of the current region and the number of region pixels of the current region, and set the segmentation scale parameter for the region threshold term;

[0086] Take edges sequentially according to the ordered edge sequence and locate candidate regions. Read the labels of the pixels at both ends of the current edge in the initial region label map to determine the two candidate regions to which they belong. If the pixels at both ends belong to the same region, skip the current edge. If the pixels at both ends belong to different regions, read the internal difference of the two candidate regions, the number of pixels in the region, and the region threshold, and calculate the joint threshold of the two candidate regions.

[0087] The Felzenszwalb merging criterion is executed and an improved merging criterion is applied. The reconstructed edge weight of the current edge is compared with the joint threshold to obtain the original merging result. When the original merging result allows merging, the improved merging criterion is further executed. The improved merging criterion includes a hysteresis merging strategy, which maintains a merging suppression flag for each region and sets a tight threshold factor and a loose threshold factor. When the merging suppression flag is enabled, the tight threshold factor is used to tighten the joint threshold. When the merging suppression flag is disabled, the loose threshold factor is used to relax the joint threshold. When the improved merging criterion allows merging, the two candidate regions are merged and the region identifier, region pixel count, region internal difference, region minimum spanning tree edge weight set, and merging suppression flag of the merged region are updated. The pixel labels of the two candidate regions in the initial region label map are also updated synchronously.

[0088] After traversing the ordered edge sequence, the merging is completed to obtain the final region label map. The pixel positions are grouped according to the region identifier of the final region label map, and each group of pixel positions is combined into a segmented region. All segmented regions are summarized to obtain the segmented region set.

[0089] This invention obtains the final region label map and outputs a set of segmented regions based on reconstructed edge weight sorting and Felzenszwalb graph segmentation. By maintaining the internal difference of the region and the number of pixels in the region to form a region threshold term and introducing a hysteresis merging strategy, it achieves stable control of region merging, reduces false merging and merging jitter caused by low-light noise disturbance, improves the consistency of segmentation boundaries and the integrity of region structure, and provides a reliable region unit and label basis for subsequent region-level parameter calculation and noise reduction control map mapping.

[0090] In this embodiment, the generation of the noise reduction control chart specifically includes:

[0091] Establish a region parameter table, process each region according to the segmented region set, traverse all pixel positions belonging to the current region identifier for the current region, accumulate the confidence value corresponding to the pixel position and accumulate the number of pixels, normalize the accumulated confidence value according to the number of pixels to obtain the region-level recognition confidence aggregate value, and write the region-level recognition confidence aggregate value into the region parameter table.

[0092] For the current region, traverse all pixel positions belonging to the current region identifier, accumulate the brightness value corresponding to the pixel position and accumulate the number of pixels, normalize the accumulated brightness value according to the number of pixels to obtain the region-level brightness mean, and combine the noise parameter and the region-level brightness mean to obtain the region-level noise intensity value. The region-level noise intensity value increases with the increase of the noise parameter and decreases with the increase of the region-level brightness mean, and write the region-level noise intensity value into the region parameter table.

[0093] For the current region, the regional noise reduction intensity parameter is calculated based on the regional noise intensity value and the aggregated regional identification confidence value. The regional noise reduction intensity parameter increases with the increase of the regional noise intensity value and decreases with the increase of the aggregated regional identification confidence value. The upper and lower limits of the regional noise reduction intensity parameter are clipped, and the clipped regional noise reduction intensity parameter is written into the regional parameter table.

[0094] Traverse all pixel positions in the final region label map, read the region identifier corresponding to the pixel position, query the region-level noise reduction intensity parameter of the region identifier in the region parameter table, write it to the corresponding pixel position of the noise reduction control map, and output the noise reduction control map after completing the traversal.

[0095] This invention determines regional denoising intensity parameters by fusing regional-level recognition confidence aggregated values ​​and regional-level noise intensity values, and maps them to generate a denoising control map. This allows the denoising intensity to be constrained by semantic reliability and noise level simultaneously and controlled by upper and lower limit clipping. As a result, it enables fine-grained adaptive adjustment of different segmented regions in low-light scenes, suppresses background residual noise, reduces the risk of over-smoothing of the target region, improves regional consistency and control stability, and provides directly callable pixel-level intensity guidance for subsequent regional dynamic denoising.

[0096] In this embodiment, obtaining the noise-reduced output frame specifically includes:

[0097] Read the target frame and the noise reduction control map, generate a noise reduction output frame with the same pixel size as the target frame and complete pixel position alignment, and read the noise reduction control map pixel by pixel to obtain the noise reduction intensity parameter corresponding to each pixel position;

[0098] Pixel-level filtering configurations are established based on noise reduction intensity parameters. For each pixel location, the filtering window size, brightness similarity scale, and intensity gain are determined from the noise reduction intensity parameters, and the filtering configuration corresponding to each pixel location is written into the configuration mapping.

[0099] The process performs regionalized dynamic denoising, traversing the target frame pixel by pixel and reading the filtering configuration from the configuration mapping at the current pixel position. Within the filtering window of the current pixel position, it enumerates the neighboring pixel positions and reads the brightness values ​​of each neighboring pixel. Spatial weights are determined based on the spatial distance between the neighboring pixels and the current pixel. Similarity weights are determined based on the difference between the brightness of the neighboring pixels and the brightness of the current pixel, combined with a brightness similarity scale. The spatial weights and similarity weights are combined to obtain the neighboring pixel combination weights. The brightness values ​​of the neighboring pixels are weighted and accumulated according to the combination weights, and then normalized according to the accumulated value to obtain the denoised brightness value of the current pixel. The denoised brightness value of the current pixel is written to the corresponding pixel position of the denoised output frame, and the denoised output frame is output. The denoised output frame stores the image result after regionalized dynamic denoising of the target frame according to the denoising control map. The denoised output frame and the target frame correspond one-to-one in pixel size and pixel position. The denoised output frame records the denoised brightness value or denoised pixel value of the corresponding pixel at each pixel position.

[0100] This invention implements regional dynamic noise reduction on the target frame based on a noise reduction control map. By establishing a one-to-one correspondence between pixel position and regional noise reduction intensity parameters and driving the adaptive configuration of the filter window, brightness similarity scale, and intensity gain, different regions can achieve differentiated control in terms of noise suppression intensity and detail preservation. This effectively reduces noise particles and residual noise interference under low illumination conditions while maintaining the clarity of target edges and texture details. At the same time, it ensures the consistency of the noise reduction output frame and the target frame in terms of pixel size and pixel position for easy identification and display.

[0101] Example 1: To verify the feasibility of this invention in practice, it was applied to the real-time imaging processing link of a daytime night vision device in a low-light surveillance and evidence collection scenario. The scenario included people walking, vehicles passing slowly, and reflective signs and text signs appearing simultaneously. The image also contained large areas of dark background and local bright reflection points. Under low-light conditions, the sensor output image showed obvious grain noise and color jitter. Common global noise reduction schemes often increase the smoothing intensity to suppress noise, resulting in the smoothing of personnel outlines, license plate characters, and sign edges. Graying shadows that cross semantic boundaries are also prone to appear near the edges, significantly reducing the confidence of subsequent target detection and recognition. This invention aims to solve the contradiction of strong noise in low light, weak target details, and semantic boundaries that are easily over-smoothed, maintaining the stability of target edges and texture structures while ensuring noise suppression.

[0102] At the system integration level, the daytime night vision device uses a common low-light imaging module with an output resolution of 1920*1080 and a frame rate of 30 frames per second. The processing unit employs an embedded computing unit to meet real-time requirements. After the image enters the processing chain, noise level estimation is performed and noise parameters are output. Simultaneously, a pixel-level image recognition model outputs a recognition confidence map, which expresses the reliability of each pixel location's semantic correlation with the target. Then, a pixel map is constructed based on the target frame, generating an initial edge set and initial edge weights. The initial edge weights are obtained by weighting the brightness difference and gradient difference. The semantic boundary constraint reconstruction module uses the recognition confidence map to perform edge deletion and retention on the initial edge set. Connections that cross semantic boundaries are suppressed, while connections that are preserved retain their initial edge weights to form a reconstructed edge set and reconstructed edge weights. The reconstructed edge set is then used in Felzenszwalb graph segmentation to obtain a final region label map and a segmented region set. The region set is used to aggregate the identification confidence and brightness statistics within the region and combine them with noise parameters to generate region-level noise reduction intensity parameters. The region-level noise reduction intensity parameters are mapped back to the final region label map to obtain a noise reduction control map. The noise reduction control map drives pixel-level filtering configuration. The filtering window, brightness similarity scale, and intensity gain change with the region intensity parameters to achieve differentiated control of stronger noise reduction in the background region and stronger edge preservation in the target region. The output noise-reduced frame is then provided for display and intelligent analysis modules to use.

[0103] To generate repeatable and quantifiable validation data, the test data consisted of multiple low-light video sequences. The illumination levels were obtained by combining neutral density and exposure, covering both extremely low and low-light conditions. The evaluation employed both subjective discernibility and objective metrics. The objective metrics included peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), edge preservation index (EPI), target detection average accuracy (mAP), end-to-end processing latency, and the standard deviation of each metric. Peak Signal-to-Noise Ratio (PSNR) measures the error level between the denoising result and the reference image at the pixel intensity level; a higher value indicates less overall distortion. Structural Similarity Index (SSIM) measures the consistency of brightness, contrast, and structural information; a higher value indicates better structural fidelity. Edge Preservation Index (EPI) measures the proportion of gradient magnitude preserved near the target boundary; a higher value indicates more complete preservation of edge and texture details. Mean Accuracy Detection (mAP) measures the detection availability of the detection network on the denoised output frame; a higher value indicates stronger target detectability. Processing latency measures real-time performance; lower average latency and P95 quantile latency indicate more stable real-time performance. The comparison methods selected three common engineering schemes: global bilateral filtering, non-local averaging, and lightweight deep denoising. These were compared with the method of this invention under the same output resolution and frame rate conditions to ensure fairness. Specific comparison data are shown in Table 1.

[0104] Table 1. Comparison of Objective Quality Indicators and Real-Time Performance of Different Noise Reduction Methods on Low-Light Sequences

[0105] Illuminance levels method Peak signal-to-noise ratio (PSNR) (dB) Structural Similarity (SSIM) Edge Preservation Index (EPI) Target detection mAP(%) Average latency (ms / frame) Latency P95 (ms / frame) Extremely low light Global bilateral filtering 24.18(0.62) 0.701(0.018) 0.59(0.04) 57.6(2.1) 7.4(0.5) 9.2 Extremely low light Nonlocal mean 25.36(0.71) 0.726(0.020) 0.63(0.05) 60.9(2.4) 18.9(1.6) 24.7 Extremely low light Lightweight deep noise reduction 26.48(0.75) 0.758(0.021) 0.67(0.05) 64.3(2.6) 12.6(1.1) 16.9 Extremely low light Method of the present invention 27.31(0.64) 0.789(0.017) 0.74(0.04) 70.8(2.2) 11.3(0.9) 14.8 low light Global bilateral filtering 27.05(0.55) 0.781(0.014) 0.67(0.03) 72.4(1.8) 7.2(0.4) 8.8 low light Nonlocal mean 28.14(0.58) 0.804(0.015) 0.70(0.03) 74.9(2.0) 18.2(1.4) 23.6 low light Lightweight deep noise reduction 28.93(0.63) 0.828(0.016) 0.73(0.03) 77.6(1.9) 12.1(0.9) 15.8 low light Method of the present invention 29.54(0.56) 0.848(0.013) 0.78(0.03) 81.9(1.6) 10.8(0.7) 13.9

[0106] As shown in Table 1, under extremely low illumination conditions, the PSNR of the method of this invention is 27.31 dB, SSIM is 0.789, EPI is 0.74, and mAP is 70.8%, which are higher than those of lightweight depth denoising (26.48 dB, 0.758, 0.67, and 64.3%). Meanwhile, the average latency is 11.3 ms and the P95 is 14.8 ms, which are lower than those of lightweight depth denoising (12.6 ms and 16.9 ms). This indicates that the present invention can simultaneously improve pixel error control, structure fidelity, and edge preservation, and improve detection even under extremely low illumination conditions with stronger noise. The method tested availability and maintained stable real-time output. Under low-light conditions, the PSNR of this invention was 29.54dB, SSIM was 0.848, EPI was 0.78, and mAP was 81.9%, which were also higher than those of lightweight depth denoising (28.93dB, 0.828, 0.73, and 77.6%). At the same time, the average latency was 10.8ms and the P95 was 13.9ms, which were better than those of lightweight depth denoising (12.1ms and 15.8ms). This demonstrates the stable gain and real-time performance advantages of this invention in terms of image quality and detectability under different low-light conditions.

[0107] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for dynamic noise reduction of low-light images of daytime night vision devices based on image recognition, characterized in that, Includes the following steps: Acquire the low-light image sequence output by the daytime night vision device and determine the current frame as the target frame; Noise level estimation and image recognition are performed on the target frame to obtain noise parameters and recognition confidence maps; Based on the target frame, an initial edge set is generated according to a preset connection rule, and the initial edge weights are determined according to the differences in pixel features. Based on the recognition confidence graph, the initial edge set is reconstructed using semantic boundary constraints. A confidence threshold is set and the edges are retained. The initial edge weights of the retained edges are used as the reconstructed edge weights, and the reconstructed edge set is output. Perform Felzenszwalb graph segmentation on the reconstructed edge set, sort the reconstructed edge set according to the reconstructed edge weight, initialize each pixel as an independent region, take edges in the sorted order and perform merging or preservation on the regions to which the two ends of the edge belong according to the merging criterion, and output the final region label map and segmented region set after traversal. Using the segmented region set as the region unit, the region-level noise reduction intensity parameters are determined for each region by combining the identification confidence map and noise parameters, and the region-level noise reduction intensity parameters of each region are mapped to the final region label map to generate a noise reduction control map. Based on the noise reduction control chart, the target frame is subjected to regional dynamic noise reduction processing to obtain the noise-reduced output frame.

2. The method for dynamic noise reduction of low-light images of daytime night vision devices based on image recognition according to claim 1, characterized in that, Obtaining the target frame specifically includes: The imaging sensor driving the daytime night vision device outputs low-light image frames during continuous sampling. Each low-light image frame is written into the image buffer in chronological order to form a low-light image sequence. Read the low-light image frame with the latest frame number and the largest sampling timestamp from the low-light image sequence, use it as the current frame, mark the current frame as the target frame, record the frame number and sampling timestamp corresponding to the target frame, and write the target frame into the target frame buffer.

3. The method for dynamic noise reduction of low-light images of daytime night vision devices based on image recognition according to claim 1, characterized in that, The acquisition of the noise parameters and the identification confidence map specifically includes: Luminance channel extraction is performed on the target frame to obtain a luminance map, and a one-to-one correspondence between the pixel positions of the target frame is established. A smoothed brightness map is obtained by performing a smoothing filter on the brightness map. The brightness map and the smoothed brightness map are then subtracted by pixel position to obtain a residual map. At the same time, the absolute value of each pixel in the residual map is taken to form an absolute residual set. Based on the absolute residual set, noise parameters characterizing the noise level of the target frame are calculated. The target frame is input into the image recognition model that has been pre-trained and deployed on the daytime night vision device. Forward inference is performed to obtain a recognition confidence map with the same pixel size as the target frame.

4. The method for dynamic noise reduction of low-light images of daytime night vision devices based on image recognition according to claim 1, characterized in that, The process of generating an initial edge set based on the target frame and according to a preset connection rule, and determining the initial edge weights based on pixel feature differences, specifically includes: Traverse all pixel positions of the target frame and form a set of pixel positions. For each pixel position in the set of pixel positions, establish a corresponding pixel node to obtain a set of pixel nodes. An initial edge set is generated for the set of pixel nodes according to a preset connection rule. The preset connection rule is the eight-neighbor connection rule. For each pixel position in the set of pixel positions, the pixel positions adjacent to each other in the horizontal, vertical and diagonal directions are determined. The pixel positions are paired with the pixel nodes corresponding to each of the eight adjacent pixel positions to generate edges. All generated edges are summarized to obtain the initial edge set. Extract pixel features from the target frame, calculate the brightness gradient of the brightness map of the target frame according to the pixel position to obtain the gradient map, and make the brightness map, gradient map and pixel position set correspond one-to-one at the pixel position; For each edge in the initial edge set, read the brightness values ​​of the pixel positions at both ends of the current edge in the brightness map and calculate the absolute value of the brightness difference. At the same time, read the gradient values ​​of the pixel positions at both ends of the current edge in the gradient map and calculate the magnitude of the gradient difference. Multiply the absolute value of the brightness difference and the magnitude of the gradient difference by the first weight coefficient and the second weight coefficient respectively, and sum them to obtain the initial edge weight of the current edge. Combine the pixel node set, the initial edge set, and the initial edge weight to form a pixel map.

5. The method for dynamic noise reduction of low-light images of daytime night vision devices based on image recognition according to claim 1, characterized in that, The specific steps to obtain the reconstructed edge set include: Read the initial edge set, initial edge weight set, identify the confidence graph, set the confidence threshold, and initialize the reconstructed edge set to empty; Traverse each edge in the initial edge set in order, read the pixel positions at both ends of the current edge, read the confidence values ​​of the pixel positions at both ends in the recognition confidence map, calculate the difference between the confidence values ​​at both ends and take the absolute value of the difference to obtain the absolute value of the confidence difference of the current edge. The initial edge set is reconstructed using semantic boundary constraints. The absolute value of the confidence difference is compared with the confidence threshold. When the absolute value of the confidence difference is less than the confidence threshold, the current edge is retained and written into the reconstructed edge set. At the same time, the initial edge weight corresponding to the current edge is read from the initial edge weight set and used as the corresponding reconstructed edge weight. When the absolute value of the confidence difference is greater than or equal to the confidence threshold, the current edge is deleted. After traversing the initial edge set, perform deduplication and consistency checks on the reconstructed edge set to ensure that each edge in the reconstructed edge set has a unique corresponding edge weight in the reconstructed edge weight set. Output the reconstructed edge set and the corresponding reconstructed edge weights.

6. The method for dynamic noise reduction of low-light images of daytime night vision devices based on image recognition according to claim 1, characterized in that, The output of the final region label map and the segmented region set specifically includes: Traverse the set of reconstructed edges and read the pixel positions at both ends of each edge and the corresponding reconstructed edge weight. Sort all edges in ascending order of reconstructed edge weight to obtain an ordered edge sequence, and record the edge number for each edge in the ordered edge sequence. Perform Felzenszwalb graph segmentation on the reconstructed edge set, initialize Felzenszwalb region data, initialize each pixel in the target frame as an initial region and assign a unique region identifier, generate an initial region label map, and make the label of each pixel equal to the initial region identifier. At the same time, initialize the number of region pixels, the amount of difference within the region, and the minimum spanning tree edge weight set for each initial region. Define the Felzenszwalb merge threshold term, maintain the region threshold term for each region, and set the segmentation scale parameter for the region threshold term; Take edges sequentially according to the ordered edge sequence and locate candidate regions. Read the labels of the pixels at both ends of the current edge in the initial region label map to determine the two candidate regions to which they belong. If the pixels at both ends belong to the same region, skip the current edge. If the pixels at both ends belong to different regions, read the internal difference of the two candidate regions, the number of pixels in the region, and the region threshold, and calculate the joint threshold of the two candidate regions. The Felzenszwalb merging criterion is executed and an improved merging decision is made. The reconstructed edge weight of the current edge is compared with the joint threshold to obtain the original merging result. When the original merging result is allowed, the improved merging decision is further executed. The improved merging decision includes a hysteresis merging strategy, which maintains a merging suppression flag for each region and sets a tight threshold factor and a loose threshold factor. When the merging suppression flag is enabled, the joint threshold is tightened using the tight threshold factor. When the merging suppression flag is disabled, the joint threshold is relaxed using the loose threshold factor. When the improved merging decision allows merging, the two candidate regions are merged and the region identifier, region pixel count, region internal difference, region minimum spanning tree edge weight set, and merging suppression flag of the merged region are updated. The pixel labels belonging to the two candidate regions in the initial region label map are updated simultaneously. After traversing the ordered edge sequence, the merging is completed to obtain the final region label map. The pixel positions are grouped according to the region identifier of the final region label map, and each group of pixel positions is combined into a segmented region. All segmented regions are summarized to obtain the segmented region set.

7. The method for dynamic noise reduction of low-light images of daytime night vision devices based on image recognition according to claim 1, characterized in that, The generation of the noise reduction control map specifically includes: Establish a region parameter table, process each region according to the segmented region set, traverse all pixel positions belonging to the current region identifier for the current region, accumulate the confidence value corresponding to the pixel position and accumulate the number of pixels, normalize the accumulated confidence value according to the number of pixels to obtain the region-level recognition confidence aggregate value, and write the region-level recognition confidence aggregate value into the region parameter table. For the current region, traverse all pixel positions belonging to the current region identifier, accumulate the brightness value corresponding to the pixel position and accumulate the number of pixels, normalize the accumulated brightness value according to the number of pixels to obtain the region-level brightness mean, and combine the noise parameter with the region-level brightness mean to obtain the region-level noise intensity value, and write the region-level noise intensity value into the region parameter table. For the current region, calculate the regional noise reduction intensity parameter based on the regional noise intensity value and the aggregated value of regional identification confidence, and perform upper and lower limit pruning on the regional noise reduction intensity parameter, and write the pruned regional noise reduction intensity parameter into the regional parameter table; Traverse all pixel positions in the final region label map, read the region identifier corresponding to the pixel position, query the region-level noise reduction intensity parameter of the region identifier in the region parameter table, write it to the corresponding pixel position of the noise reduction control map, and output the noise reduction control map after completing the traversal.

8. The method for dynamic noise reduction of low-light images of daytime night vision devices based on image recognition according to claim 1, characterized in that, Obtaining the noise-reduced output frame specifically includes: Read the target frame and the noise reduction control map, generate a noise reduction output frame with the same pixel size as the target frame and complete pixel position alignment, and read the noise reduction control map pixel by pixel to obtain the noise reduction intensity parameter corresponding to each pixel position; Pixel-level filtering configurations are established based on noise reduction intensity parameters. For each pixel location, the filtering window size, brightness similarity scale, and intensity gain are determined from the noise reduction intensity parameters, and the filtering configuration corresponding to each pixel location is written into the configuration mapping. Perform regional dynamic noise reduction processing, traverse the target frame pixel by pixel and read the filter configuration in the configuration mapping for the current pixel position. Enumerate the neighboring pixel positions within the filter window of the current pixel position and read the brightness values ​​of the neighboring pixels respectively. Determine the spatial weight based on the spatial distance between the neighboring pixels and the current pixel. Determine the similarity weight based on the difference between the brightness of the neighboring pixels and the brightness of the current pixel and the brightness similarity scale. Combine the spatial weight and the similarity weight to obtain the neighboring pixel combination weight. Accumulate the brightness of the neighboring pixels according to the combination weight and normalize the accumulated value of the combination weight to obtain the noise-reduced brightness value of the current pixel. Write the noise-reduced brightness value of the current pixel to the corresponding pixel position of the noise-reduced output frame and output the noise-reduced output frame.

Images