Adaptive Non-uniformity Correction and Image Enhancement Processing Methods for Infrared Imaging Equipment

Abstract

This invention relates to the field of infrared imaging image processing technology, and discloses an adaptive non-uniformity correction and image enhancement processing method for infrared imaging devices. The method includes: acquiring multiple frames of original infrared images and IMU angular rate data, registering them, and calculating an effective diversity index; generating a spatiotemporal mask of a moving target based on the median absolute deviation in the time domain and estimating a reference radiation frame; updating offset correction parameters and outputting the corrected image; generating a correction confidence map; performing Laplacian pyramid decomposition on the corrected image and calculating a local signal-to-noise ratio (SNR) index; correcting the SNR index using the correction confidence map and performing pixel tri-classification; applying adaptive classification gain to the detail layer; performing piecewise linear stretching of the radiometric semantic interval on the base layer; and performing inverse pyramid reconstruction and adaptive gamma mapping to output the final enhanced image.

Description

Technical Field

[0001] This invention relates to the field of infrared imaging image processing technology, and more specifically, to a method for adaptive non-uniformity correction and image enhancement processing of infrared imaging devices. Background Technology

[0002] In practical operation, infrared focal plane array imaging equipment relies on two core sequential processing steps—non-uniformity correction and image enhancement—to determine the final image quality. In the non-uniformity correction stage, multiple consecutive frames of raw infrared images and inertial measurement unit displacement data are typically acquired. After registering each frame to a unified spatial coordinate system, the diversity of effective observation sources at each spatial location is statistically analyzed to distinguish between sufficiently moving and insufficiently moving regions. Reference frame estimation and offset correction parameter updates are then performed accordingly. In the image enhancement stage, multi-scale decomposition is performed on the corrected image. At the detail level, edge and flat regions are classified based on gradient magnitude thresholds, and corresponding enhancement or suppression coefficients are applied. At the base level, piecewise linear stretching is performed, and finally, the enhanced image is output through inverse reconstruction and gamma mapping.

[0003] However, when the device's aiming line is in a high-frequency reciprocating motion mode, the diversity of detector pixel sampling in different spatial regions varies significantly, resulting in a highly uneven spatial distribution of non-uniform correction residuals. Current image enhancement stages perform uniform processing on the corrected image, failing to perceive spatial differences in correction quality. This leads to the following problems: First, the gradient magnitude in areas with large correction residuals is inflated by non-uniform artifacts, causing noisy or weak signal pixels to be incorrectly assigned to edge sets and thus having enhancement coefficients applied, resulting in artifact amplification. Second, weak radiation targets in areas with low correction quality have their weak gradient signals masked by residual interference, being assigned to flat sets and attenuated by suppression coefficients, further reducing target visibility. Third, existing diversity counting suffers from artificially high counts under reciprocating motion; alternating observations of a few detector pixels result in high count values ​​but insufficient actual independent samples, distorting the reliability assessment of the reference frame estimation. Fourth, if signals generated by multiple moving targets in the registration coordinate system are not excluded, median estimation will introduce background-biased correction errors into the target region. The root cause of the above problems is the lack of an information transmission channel for the perception of correction quality between the non-uniformity correction stage and the image enhancement stage. The enhancement stage cannot adaptively adjust the processing strategy according to the spatial distribution of the correction quality. Summary of the Invention

[0004] This invention provides an adaptive non-uniformity correction and image enhancement processing method for infrared imaging devices, which solves the technical problems in related technologies such as insufficient accuracy of infrared focal plane array non-uniformity correction, ineffective utilization of correction confidence leading to poor image enhancement effect, and insufficient weak target detection capability.

[0005] This invention discloses an adaptive non-uniformity correction and image enhancement processing method for infrared imaging devices, comprising: Acquire a continuous sequence of multiple frames of raw images from an infrared focal plane array and angular rate data from an inertial measurement unit. Register the multiple frames of raw images to a unified spatial coordinate system. Count the number of detector pixel numbering types and the number of numbering switches between adjacent frames at each spatial location. Calculate the effective diversity index. A spatiotemporal mask for moving targets is generated based on the registered image sequence. Reference radiation is estimated for locations that are not marked by the mask and whose effective diversity index is greater than the sufficient motion threshold. For locations whose effective diversity index is less than the sufficient motion threshold, a complete reference frame is obtained by spatial interpolation. The offset correction parameters of the detector pixels are updated based on the complete reference frame, and non-uniformity correction is performed on the current frame to output the corrected image. A corrected confidence plot is generated based on the ratio of the effective diversity index to the sufficient motion threshold; The Laplacian pyramid decomposition is performed on the corrected image to obtain the detail layer and the base layer. The local signal-to-noise ratio index is calculated for the fine detail layer. The local signal-to-noise ratio index is corrected using the correction confidence map. Based on the corrected signal-to-noise ratio, the pixels are divided into three categories: high, medium, and low. Spatial connectivity analysis is performed on pixels with medium signal-to-noise ratio to extract weak target candidate regions and an adaptive enhancement gain is applied. Edge enhancement gain is applied to pixels with high signal-to-noise ratio and a noise suppression attenuation coefficient is applied to pixels with low signal-to-noise ratio. Piecewise linear stretching is performed on the base layer, and pyramid inverse reconstruction is performed on the enhanced detail layer and the base layer, followed by adaptive gamma mapping to output the final enhanced image.

[0006] Furthermore, the effective diversity index is calculated as follows: during the registration process, the physical detector pixel number corresponding to each spatial location in each frame is synchronously recorded to form a number sequence. The number sequence is used to count all the different numbers that have appeared at each spatial location to obtain the number of detector pixel number types. The frame-by-frame pixel number record of each spatial location in the number sequence is compared with the numbers of adjacent frames one by one and the number of different numbers is accumulated to obtain the number switching number. The effective diversity index is equal to the product of the number of detector pixel number types and the number of number switching divided by the total number of frames to obtain the normalized switching frequency.

[0007] Further, the generation of the moving target spatiotemporal mask includes: calculating the temporal median and temporal median absolute deviation for a multi-frame observation sequence at each spatial location in the registered image sequence, wherein the temporal median absolute deviation is the median of the absolute value of the difference between each frame observation and the temporal median in the time dimension; marking observations that deviate from the temporal median by more than three times the temporal median absolute deviation as outliers caused by the moving target, thereby generating the moving target spatiotemporal mask; the method for estimating the reference radiation is as follows: for spatial locations not marked by the moving target spatiotemporal mask and whose effective diversity index is greater than the sufficient motion threshold, after sorting the effective observations by value and removing the extreme values ​​of the first and last preset proportions, the arithmetic mean of the remaining observations is calculated to obtain the reference radiation estimate.

[0008] Furthermore, obtaining a complete reference frame through spatial interpolation includes: using the spatial location of the completed reference radiometric estimation as a known constraint point, and performing spatial interpolation using a joint bilateral-guided filtering. The joint bilateral-guided filtering uses the latest corrected frame that has completed offset correction as the guide image. The interpolated value of the position to be interpolated is obtained by a weighted average of the surrounding known constraint points. The weights consider both spatial distance and gray-level similarity on the guide image. Constraint points that are closer in spatial distance have higher weights, and constraint points whose gray-level values ​​on the guide image are closer to the position to be interpolated have higher weights. When there are no available correction frames during system initialization, the guide image degenerates into a bilateral filter that only depends on spatial distance. After the first frame is corrected, it switches to the joint bilateral-guided filtering mode.

[0009] Further, the method of updating the offset correction parameters of the detector pixels includes: for each detector pixel, calculating the offset correction increment as the average of the difference between the original output value of the detector pixel in the current frame and the estimated spatial position of the corresponding position in the complete reference frame across all observation positions of the detector pixel; updating the offset correction parameters using an exponentially weighted moving average method, wherein the updated offset correction parameters are equal to the sum of the difference between the original offset correction parameters multiplied by a smoothing coefficient and the offset correction increment multiplied by a smoothing coefficient, where the smoothing coefficient is greater than zero and less than one; and subtracting the offset correction parameters of the detector pixel corresponding to each spatial position in the original image of the current frame from the updated offset correction parameters to obtain the corrected image.

[0010] Furthermore, the method for generating the corrected confidence map is as follows: the corrected confidence value at each pixel location is equal to the minimum of the ratio of the effective diversity index to the sufficient motion threshold at that location and one; the range of the corrected confidence value is a closed interval from zero to one; the corrected signal-to-noise ratio is equal to the product of the local signal-to-noise ratio index and the corrected confidence value; the division of pixels into high, medium, and low categories includes: calculating the global median of the corrected signal-to-noise ratio across the entire image; pixels with a corrected signal-to-noise ratio greater than twice the global median are classified as high signal-to-noise ratio pixels; pixels with a corrected signal-to-noise ratio between the global median and twice the global median are classified as medium signal-to-noise ratio pixels; and pixels with a corrected signal-to-noise ratio less than the global median are classified as low signal-to-noise ratio pixels.

[0011] Furthermore, the step of performing spatial connectivity analysis on the mid-signal-to-noise ratio pixels to extract weak target candidate regions includes: marking the connected components of the binary mask image of the mid-signal-to-noise ratio pixels; counting the total number of pixels contained in each connected component as its area; retaining the connected components whose area is within the weak target size range to form a weak target candidate region mask; the upper and lower limits of the weak target size range are predetermined based on the optical system parameters of the infrared imaging device, the typical target size, and the detection distance; the adaptive enhancement gain is equal to the product of a weak target enhancement intensity control parameter plus a modified signal-to-noise ratio divided by twice the global median; the enhanced detail layer value of the pixels within the weak target candidate region mask is the original detail layer value multiplied by the corresponding adaptive enhancement gain.

[0012] Further, the step of performing piecewise linear stretching on the base layer includes: smoothing the gray-level histogram of the base layer and detecting local minima; identifying local minima that satisfy the minimum depth condition as effective valleys; dividing the gray-level range into several intervals using the gray-level values ​​of the effective valleys as the segmentation boundaries of the radiation semantic intervals; applying a linear stretching coefficient to each interval; extracting the gray-level mean of the weak target candidate region mask at the corresponding position in the base layer; determining the gray-level interval where the gray-level mean is located; defining a target gray-level neighborhood sub-interval centered on the gray-level mean within the gray-level interval; applying an increased stretching coefficient to the target gray-level neighborhood sub-interval to generate an enhanced base layer; and degrading the application of a single linear stretching to the entire gray-level range when the number of detected effective valleys is zero.

[0013] Further, the adaptive gamma mapping includes: calculating the global grayscale mean of the full-resolution enhanced image; the adaptive gamma parameter value is equal to the natural logarithm of the intermediate reference value of the grayscale range divided by the natural logarithm of the global grayscale mean plus a minimum positive constant; the output value of each pixel in the final enhanced image is equal to the gamma parameter value raised to the power of the quotient of the maximum allowable grayscale value multiplied by the pixel's grayscale value divided by the maximum allowable grayscale value; when the global grayscale mean is lower than the intermediate reference value of the grayscale range, the gamma parameter value is less than one, causing the grayscale value of the dark area to increase; when the global grayscale mean is higher than the intermediate reference value of the grayscale range, the gamma parameter value is greater than one, causing the grayscale value of the bright area to compress downward.

[0014] This invention provides an adaptive non-uniformity correction and image enhancement processing system for infrared imaging devices, comprising: The data acquisition and registration module is used to acquire a continuous multi-frame raw image sequence of the infrared focal plane array and the angular rate data of the inertial measurement unit, register the multi-frame raw images to a unified spatial coordinate system, count the number of detector pixel number types and the number of number switching between adjacent frames at each spatial location, and calculate the effective diversity index. The reference frame estimation module is used to generate a spatiotemporal mask for moving targets based on the registered image sequence, estimate reference radiation for positions that are not marked by the mask and whose effective diversity index is greater than the sufficient motion threshold, and obtain a complete reference frame by spatial interpolation for positions whose effective diversity index is less than the sufficient motion threshold. The non-uniformity correction module is used to update the offset correction parameters of the detector pixels based on the complete reference frame, perform non-uniformity correction on the current frame, and output the corrected image. The confidence generation module is used to generate a corrected confidence map based on the ratio of the effective diversity index to the sufficient motion threshold. The confidence-aware classification module is used to perform Laplacian pyramid decomposition on the corrected image to obtain the detail layer and the base layer. It calculates the local signal-to-noise ratio index for the fine detail layer, corrects the local signal-to-noise ratio index using the corrected confidence map, and classifies the pixels into three categories: high, medium, and low based on the corrected signal-to-noise ratio. The detail layer adaptive enhancement module is used to perform spatial connectivity analysis on pixels with medium signal-to-noise ratio to extract weak target candidate regions and apply adaptive enhancement gain, apply edge enhancement gain to pixels with high signal-to-noise ratio, and apply noise suppression attenuation coefficient to pixels with low signal-to-noise ratio. The base layer stretching and reconstruction output module is used to perform piecewise linear stretching on the base layer, perform inverse pyramid reconstruction on the enhanced detail layer and base layer, and perform adaptive gamma mapping to output the final enhanced image.

[0015] This invention addresses the technical problem in existing methods where the two processing stages are independent and the enhancement stage cannot perceive spatial differences in correction quality by introducing a correction confidence map as an information transmission channel for correction quality perception between the non-uniformity correction stage and the image enhancement stage. The invention achieves the following technical effects: It employs an effective diversity index that multiplies the number of detector types by the normalized switching frequency, avoiding inflated diversity assessments caused by alternating observations of a few pixels in reciprocating motion modes, thus making the reliability assessment of reference radiometric estimation more accurate; it uses robust outlier detection based on the median absolute deviation in the time domain to generate a spatiotemporal mask for moving targets, combined with truncated mean estimation, reducing the interference of moving target signals on reference radiometric estimation; it introduces the correction confidence map into the fine detail layer's signal-to-noise ratio correction, reasonably lowering the signal-to-noise ratio index in areas with large correction residuals that are inflated by artifacts, reducing the situation where non-uniformity artifacts are incorrectly classified into edge classes and thus subject to enhancement coefficients; and it performs spatial connectivity analysis and area constraints on mid-signal-to-noise ratio pixels to screen weak target candidate regions, combined with adaptive enhancement gain and base layer target grayscale neighborhood stretching, improving the contrast and visibility of weak targets while suppressing noise and artifacts. Attached Figure Description

[0016] Figure 1 This is a flowchart of the adaptive non-uniformity correction and image enhancement processing method for infrared imaging equipment provided in the embodiments of the present invention; Figure 2 This is a schematic diagram comparing four typical spatial location effective diversity indicators provided in the embodiments of the present invention; Figure 3 This is a schematic diagram illustrating the relationship between the effective diversity index and the corrected confidence level provided in the embodiments of the present invention; Figure 4 This is a schematic diagram comparing the reference radiometric estimate and the corrected pixel value provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the detector pixel offset correction parameter update process provided in an embodiment of the present invention; Figure 6 This is a schematic diagram comparing the original signal-to-noise ratio and the corrected signal-to-noise ratio at a typical location provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the adaptive gain calculation results for candidate pixels of weak target C provided in an embodiment of the present invention; Figure 8 This is a schematic diagram illustrating the division of the basic layer grayscale histogram into semantic intervals, as provided in an embodiment of the present invention. Detailed Implementation

[0017] In practical operation, infrared focal plane array imaging equipment relies on two core sequential processing steps—non-uniformity correction and image enhancement—to determine the final image quality. In the non-uniformity correction stage, multiple consecutive frames of raw infrared images and inertial measurement unit (IMU) displacement data are typically acquired. After registering each frame to a unified spatial coordinate system, the diversity of effective observation sources at each spatial location is statistically analyzed to distinguish between sufficiently moving and insufficiently moving regions. Reference frame estimation and offset correction parameter updates are then performed accordingly. In the image enhancement stage, multi-scale decomposition is performed on the corrected image. At the detail level, edge and flat regions are classified based on gradient magnitude thresholds, and corresponding enhancement or suppression coefficients are applied. At the base level, piecewise linear stretching is performed, and finally, the enhanced image is output through inverse reconstruction and gamma mapping.

[0018] However, when the aiming line of the device operates in a high-frequency reciprocating motion mode (e.g., when the aiming line switches rapidly between multiple targets during simultaneous tracking), the diversity of detector pixel sampling in different spatial regions is significant. Regions along the path repeatedly traversed by the aiming line experience frequent detector switching and high correction quality, while regions at the edges of the field of view and between targets experience sparse detector sampling and low correction quality. This results in a highly uneven spatial distribution of the non-uniform correction residuals. However, current image enhancement stages perform uniform processing on the corrected image, failing to perceive the spatial differences in correction quality, thus leading to the following specific problems: First, in the fine detail layer, the gradient magnitude of areas with large correction residuals is amplified by non-uniform artifacts, causing pixels that originally belong to noise or weak signals to be incorrectly classified into edge sets and thus have enhancement coefficients applied, resulting in artifact amplification.

[0019] Second, weak radiation targets that occupy only a few pixels at the far end of the field of view have their weak gradient signals masked by residual interference in areas with low correction quality. They are then classified into flat sets and their suppression coefficients are attenuated, further reducing the visibility of the targets.

[0020] Third, existing diversity counts suffer from inflated values ​​under reciprocating motion—alternating observations of a few detector pixels lead to high counts but insufficient actual independent samples, which in turn distorts the reliability assessment of the reference frame estimation.

[0021] Fourth, if the signals generated by multiple moving targets in the registration coordinate system are not excluded, the median estimation will discard the target radiation as an outlier, introducing a correction error biased towards the background in the target area.

[0022] Therefore, an infrared image processing method is needed that can establish an information transmission channel for correction quality perception between the non-uniformity correction stage and the image enhancement stage, so that the enhancement stage can adaptively adjust the processing strategy according to the spatial distribution of correction quality, thereby suppressing artifacts while maintaining the visibility of weak targets.

[0023] According to an embodiment of this invention, this embodiment provides an adaptive non-uniformity correction and image enhancement processing method for an infrared imaging device. It should be understood that the execution entity of the adaptive non-uniformity correction and image enhancement processing method for an infrared imaging device is an image processor in an infrared imaging device equipped with an infrared focal plane array detector and an inertial measurement unit. The image processor is communicatively connected to the infrared focal plane array detector to acquire raw infrared image frames, and is communicatively connected to the inertial measurement unit to acquire angular rate data.

[0024] At least one embodiment of the present invention discloses an adaptive non-uniformity correction and image enhancement processing method for infrared imaging devices, such as... Figure 1 As shown, it includes the following steps: Step 1: Acquire the original infrared image sequence and IMU data, register them to a unified spatial coordinate system, and calculate the effective diversity index; Acquire continuous infrared focal plane array Original image sequence of frames and the angular rate data sequence output by the inertial measurement unit ,in The frame number, For pixel coordinates, Total number of frames. Diagonal rate data sequence. By performing time integration and combining it with the focal length of the optical system to convert the angle to a pixel, the inter-frame pixel displacement sequence is obtained. Based on inter-frame pixel displacement sequence ,Will The original frames of images are registered to a unified spatial coordinate system to obtain a sequence of registered images. .

[0025] During the registration process, each spatial location is recorded simultaneously. The corresponding physical detector pixels in each frame are numbered to form a numbering sequence. Based on number sequence Statistical analysis of each spatial location Number of different detector pixel numbers appearing at the location and the number of times the numbering of adjacent frames changes. Calculate the effective diversity index : in, Indicates spatial location exist The number of species observed by different detector pixels in a frame. This indicates the number of times the detector pixel number changes between adjacent frames. This represents the total number of frames. and All quantities are dimensionless. The normalized switching frequency is obtained by multiplying the two. It is also a dimensionless index, with consistent dimensions, and its calculations are valid.

[0026] It should be noted that the above-mentioned effective diversity indicators This refers to the number of detector types. With switching frequency The joint index of multiplication. This reflects how many different detector pixels have sampled this location, and This reflects the uniformity of sampling switching over time. When the aiming line moves back and forth, if a certain position is observed alternately by only two detector pixels, then... The value is 2. Although the number of switching may be high, the product is still limited by the low number of categories, thus avoiding an inflated assessment of diversity caused by relying on only a single counting metric.

[0027] Furthermore, The statistical method is as follows: traverse the number sequence Medium spatial position In all The frame's pixel number record is used to count and remove duplicates of all occurrences of the unique numbers; the total number of unique numbers obtained is the result. . The statistical method is as follows: for the numbered sequence Medium spatial position The frame-by-frame pixel numbering record is compared one by one. Frame and the The frame numbering is used to count a handover when two adjacent frame numbers are different. The total number of handovers occurring in all adjacent frame pairs is the total number of handovers. The effective statistical range is to common A pair of adjacent frames.

[0028] Step 2: Generate a spatiotemporal mask for the moving target and estimate the reference radiation frame; For the registered image sequence Each spatial location of Calculate the median in the time domain from a sequence of frame observations. and median absolute deviation in the time domain : in, This indicates a median operation over the time dimension. Values ​​deviating from the median by more than [a certain amount] will be considered. times The observed values ​​are marked as outliers, and a spatiotemporal mask for moving targets is generated. : in, This indicates that the observation was marked as an outlier caused by a moving target.

[0029] Spatiotemporal mask for non-moving targets Mark (i.e.) And effective diversity indicators Greater than the full motion threshold The spatial location is used as the truncated mean of the effective observations as the reference radiation estimate. The truncated mean is calculated by sorting the valid observations numerically, removing a certain percentage of the first and last extreme values, and then calculating the arithmetic mean of the remaining observations. For example, the mean is calculated after removing 10% of the first and last extreme values. The truncated mean has higher statistical efficiency than the median, and the truncating operation reduces the impact of residual outliers on the estimation results.

[0030] For effective diversity indicators Below the threshold of full exercise For spatial locations with insufficient effective sample sizes, the reference radiation estimate is used for spatial locations that have already been estimated. Given the constraint points, spatial interpolation is performed using joint bilateral-guided filtering to obtain the complete reference frame. The joint bilateral-guided filtering uses the local structural information of the corrected image as the guide image. During spatial interpolation, known constraint points that are close to the position to be interpolated and have similar structures on the guide image are given higher weights. This allows the reference radiation estimation to be filled in the motion-deficient region to maintain the continuity and consistency of the spatial structure.

[0031] Furthermore, the guide image for the joint bilateral-guided filtering is the latest corrected frame that has completed offset correction, and the local grayscale structure of the latest corrected frame is used as the basis for spatial similarity measurement. For the position to be interpolated... interpolation position The interpolated value is obtained by a weighted average of the surrounding known constraint points. The weights consider both spatial distance and gray-level similarity on the guide image: constraint points that are spatially closer have higher weights, and constraint points whose gray-level values ​​on the guide image are closer to the position to be interpolated have higher weights. When there are no available correction frames during system initialization, the guide image degenerates into bilateral filtering that relies solely on spatial distance. After the first frame is corrected, it switches to a joint bilateral-guided filtering mode.

[0032] Step 3: Update the offset correction parameters and output the corrected image; Based on the complete reference frame Calculate the offset correction increment for each detector pixel based on the detector pixel response corresponding to the current frame. For detector pixels Offset correction increment For detector pixels The original output value of the current frame and the complete reference frame The difference between the estimated spatial locations of the corresponding pixels in the current frame, i.e., the pixels detected by the detector in the current frame. The mean of the differences between the raw output values ​​and the reference radiation estimate at all observed spatial locations reflects the detector pixel value. The amount of deviation of the current offset state relative to the reference radiation.

[0033] The offset correction parameters are updated using an exponentially weighted moving average method. : in, Number the detector pixels. Smoothing coefficient and , The offset correction increment calculated for the current frame. The offset correction parameters before the update. These are the updated offset correction parameters. , and The dimensions of all values ​​are consistent with the dimensions of the detector output grayscale value. These are dimensionless coefficients; the dimensions of each term in the formula match, ensuring valid calculations. Smoothing coefficient. Control the fusion ratio of new and old parameters, a smaller ratio Value makes the offset correction parameter The update is more gradual, suitable for stable operating conditions with high calibration quality; larger... The value then makes the offset correction parameter Faster tracking of changes, suitable for offset correction parameters In practical applications, drift exists. The range of values ​​can be .

[0034] Using the updated offset correction parameters For the original image of the current frame Perform non-uniformity offset correction and output the corrected image. : in, Indicates the spatial position in the current frame The corresponding detector pixel number. and The dimensions of both are the same as the dimensions of the detector output grayscale value, and the difference is the result. The dimensions are consistent, and the calculation is valid.

[0035] Step 4: Generate the corrected confidence plot; Based on the effective diversity index obtained in step 1 and the sufficient motion threshold used in step 2 Generate a corrected confidence plot : in, and Both are dimensionless indicators, and the ratio between them is a dimensionless quantity. The range of values ​​is The dimensions are consistent, and the calculation is valid. (Correction confidence plot) Quantitatively characterizes the reliability of non-uniformity correction at each pixel location. When the effective diversity index for a given spatial location... Reaching or exceeding the full exercise threshold At that time, the corrected confidence plot at that location A value of 1.0 indicates that the correction quality is sufficiently reliable; when the effective diversity index Below the threshold of full exercise At that time, the confidence plot was corrected. The value decreased proportionally.

[0036] Furthermore, the full motion threshold The physical meaning is: when spatial position Effective diversity indicators Reaching the full motion threshold At that time, it was assumed that a sufficient number of evenly distributed independent detector pixel observation samples had been accumulated at that location, with reference to the radiation estimate. The results are statistically reliable. (Sufficient motion threshold) Based on the number of pixels in the infrared focal plane array detector, the sampling diversity distribution under typical motion scenarios, and the reference radiation estimation The minimum number of independent samples required is predetermined and fixed during the system initialization phase.

[0037] Furthermore, the full motion threshold The determination process is as follows: given the number of detector pixels and typical motion scenarios, statistical analysis is performed on the sampling diversity distribution to determine the reference radiation estimation. The minimum effective diversity level required for the error to meet the system correction accuracy requirements, and the effective diversity index corresponding to this level. The value is set to the full motion threshold. Specifically, when the product of the number of independent detector pixel types and the switching frequency in the effective observation samples at a certain spatial location reaches the sufficient motion threshold... When the statistical error of the truncated mean estimation result at that location decreases to an acceptable range, the reference radiation estimate at that location is considered to be within acceptable limits. It possesses sufficient reliability.

[0038] Corrected confidence plot As a carrier for transmitting spatial distribution information of correction quality from the non-uniformity correction stage to the image enhancement stage, it enables subsequent enhancement processing to perceive and adapt to spatial differences in correction quality.

[0039] Step 5: Perform Laplacian pyramid decomposition on the corrected image and calculate the local signal-to-noise ratio. For the corrected image Using Laplace's pyramid decomposition, we obtain Layer of detail and 1st base layer ,in This represents the total number of levels in the pyramid. For level sequence number, Corresponding to the finest level, As the hierarchy increases, the levels gradually become coarser.

[0040] Furthermore, the specific process of Laplacian pyramid decomposition is as follows: First, the corrected image... After applying Gaussian blur and downsampling, the first layer of the Gaussian pyramid image is obtained. Image of the first level of the Gaussian pyramid. Upsampled to original resolution and the corrected image Subtraction yields the first level of detail. Image of the first level of the Gaussian pyramid Repeat the Gaussian blur and downsampling operations described above to obtain layers 2 through 3 in sequence. Layered Gaussian pyramid image and corresponding detail layers to ;No. The image of the Gaussian pyramid is the base layer. It preserves complete low-frequency radiation information.

[0041] For the two most detailed layers and Calculate each pixel In pixels Centered The original local signal-to-noise ratio within the local window : in, For the first Layer detail layer in pixels Centered Mean within a local window The standard deviation within this local window. To prevent extremely small positive numbers from being divided by zero, This represents the side length of the local window. and The dimensions of both are the grayscale values ​​of the detail layer; dividing the two yields... It is a dimensionless quantity, but the dimensions are consistent, so the calculation is valid.

[0042] Original local signal-to-noise ratio index This represents the ratio of local signal intensity to local fluctuations in the detail layer. In regions with high correction quality, the true edge signals in the detail layer can produce a high original local signal-to-noise ratio. In regions with lower correction quality, the non-uniform correction residuals themselves contribute additional local fluctuations and mean shifts, leading to a decrease in the original local signal-to-noise ratio. It cannot accurately reflect the signal quality.

[0043] Furthermore, local window size The selection should match the effective resolution of each layer of the Laplace pyramid: in The finest layer, Smaller values ​​are used to capture pixel-level detail signals; The sub-fine layer, The value is increased accordingly to adapt to the spatial distribution range of the signal at this scale. The specific value is determined in advance based on the point spread function width of the optical system and the imaging size of typical weak targets in each layer.

[0044] Step 6: Correct the local signal-to-noise ratio using the calibration confidence map and perform pixel classification; Using the corrected confidence plot For the original local signal-to-noise ratio index Make corrections and calculate the corrected signal-to-noise ratio. : in, It is a dimensionless quantity. Since they are dimensionless quantities, the product of the two is... It is also a dimensionless quantity, but its dimensions are consistent, and its calculations are valid.

[0045] Corrected signal-to-noise ratio Calculate the global median across the entire graph. Based on the global median Pixels are divided into three categories: high Pixels: This corresponds to a significant edge region.

[0046] middle Pixels: This corresponds to the area where potential weak targets are located.

[0047] Low Pixels: This corresponds to the noise-dominant region.

[0048] Corrected signal-to-noise ratio By using the original local signal-to-noise ratio index With Corrected Confidence Plot Multiplication yields the result that when a pixel location is in a region of low correction quality, even if the non-uniform residual increases the original gradient magnitude and the original local signal-to-noise ratio at that location. Value, multiplied by a lower corrected confidence plot Post-value correction signal-to-noise ratio It will still be downgraded to a level that matches the actual signal quality in that area. Therefore, artifact pixels in areas with large correction residuals are less likely to be incorrectly classified as high. Enhancement gain is applied based on category. Simultaneously, if pixels in the weak target region are located in areas with low correction quality, the signal-to-noise ratio is corrected. More likely to fall into the middle The scope, rather than being falsely inflated to a high level. This limits the scope, thus preserving the opportunity for subsequent targeted augmentation of weak targets.

[0049] Furthermore, the global median Image after correction in the current frame Corrected signal-to-noise ratio The overall distribution is the statistical basis, and the global median is the median. The value reflects the image in the first frame. The overall signal quality level at the layer detail level. Using the global median. and As a three-class classification boundary, the classification threshold adapts to the image content, avoiding the problem of classification failure in different scenarios with a fixed threshold.

[0050] Step 7: Apply classification adaptive gain to the detail layer pixels and generate an enhanced detail layer; China Perform spatial connectivity analysis on the pixel set to extract connected regions whose areas fall within the size range of the weak target. Connected components within the region are labeled as weak target candidate region masks. Weak target size range The parameters of the infrared imaging equipment's optical system, typical target size, and detection range are predetermined, among which... Minimum area threshold for weak targets. The maximum area threshold for weak targets, where the area is less than... The connected components are considered as isolated noise points, with an area greater than 1. The connected components are considered as large-area background structures and do not belong to weak target candidates.

[0051] Furthermore, spatial connectivity analysis employs four-connectivity or eight-connectivity criteria for the mid-space. The binary mask image of the pixels is labeled with connected components. For each connected component, the total number of pixels it contains is counted as the area of ​​that connected component. The area that falls within the range of pixels is retained. For each connected component within the specified range, all pixels within these connected components are set to 1, and the remaining pixels are set to 0, thus forming a weak target candidate region mask. .

[0052] Masking of weak target candidate regions Apply adaptive enhancement gain to pixels within : in, Enhance the strength control parameters for weak targets and , The corrected signal-to-noise ratio calculated in step 6. This is the global median calculated in step 6. and Both are dimensionless quantities, and their ratio is also dimensionless. These are dimensionless control parameters. The gain coefficient is dimensionless, dimensionless, and computationally valid. It is used when masking weak target candidate regions. Corrected signal-to-noise ratio of inner pixels The closer to the middle The lower bound of the range (i.e., the closer it is to the global median) When the value inside the parentheses is closer to 1, the adaptive enhancement gain is higher. The larger the value, the more it compensates for its weaker signal strength; when adjusting the signal-to-noise ratio... The closer to the middle upper limit of range (i.e., the closer to) When the value inside the parentheses is closer to 0, the adaptive enhancement gain increases. The closer the value is to 1, the more gradual the enhancement becomes. Weak target candidate region mask. Intra-pixel enhancement detail layer value From the original detail layer value Multiply by the corresponding adaptive enhancement gain get: Furthermore, the intensity control parameters for weak target enhancement The value of determines the weak target candidate region in the corrected signal-to-noise ratio. In the middle The maximum gain achievable at the lower limit of the range. When adjusting the signal-to-noise ratio... Take the middle Lower limit of range: global median At that time, adaptive enhancement gain Reaching the maximum value When adjusting the signal-to-noise ratio Take the middle upper limit of range At that time, adaptive enhancement gain Degenerates to 1, meaning no additional enhancement is applied. Weak target enhancement intensity control parameter. The specific value is determined in advance based on the trade-off between the system's need to improve the contrast of weak targets and the amplification of introduced noise.

[0053] For high Apply edge enhancement gain to pixels ,in The corresponding enhancement detail layer value is For low Apply noise suppression attenuation coefficient to pixels ,in The corresponding enhancement detail layer value is . and All are dimensionless gain coefficients. After multiplying them by the grayscale value of the detail layer, their dimensions remain unchanged, and the calculation is valid. and All are based on hierarchy Pre-defined constants, varying with level The scale is increased (i.e., the scale becomes coarser) and gradually converges to 1 to avoid introducing excessively strong enhancement or suppression effects at the coarse-scale level.

[0054] For the remaining detail layers ( Based on the gradient magnitude of each layer, a gradient magnitude threshold classification method is used to apply hierarchical adaptive gain, generating each layer of enhanced detail layers. Specifically, regarding the first Each pixel in the detail layer Calculate pixels gradient magnitude gradient magnitude It is obtained by taking the square root of the sum of the squares of the first-order difference magnitudes in the horizontal and vertical directions of the detail layer image. The gradient magnitudes are statistically analyzed across the entire layer. High percentile as edge threshold Low percentile as flatness threshold For gradient magnitude Above the edge threshold Apply edge enhancement gain to pixels For gradient magnitude Below the flat threshold Apply noise suppression attenuation coefficient to the pixel For values ​​between the edge threshold With flat threshold The pixels between the two thresholds are attenuated using a linear interpolation algorithm with a noise suppression coefficient. With edge enhancement gain Smooth transitions between layers generate corresponding enhanced detail layers. .

[0055] Furthermore, edge threshold The corresponding high percentile and flat threshold The corresponding lower percentile is based on the first The gradient magnitude statistical distribution of the detail layer is predetermined, and the selection of the high percentile should be such that the edge threshold is not affected. To distinguish the high-value portions of the gradient magnitude distribution corresponding to the edges of the real structure from the background and noise regions, the low percentile should be selected to flatten the threshold. It can cover the low-value portion of the gradient magnitude distribution corresponding to the flat background region, and the edge threshold. With flat threshold Sufficient transition intervals are maintained between the two thresholds to support smooth transitions for linear interpolation.

[0056] Based on the pixel-based three-class classification in step 6, spatial connectivity analysis is used to further classify the pixels. Further pixel filtering distinguishes between pixel regions with spatially clustered characteristics and sizes consistent with weak target features, and randomly scattered pixels with moderate signal-to-noise ratios. Masking is then applied to weak target candidate regions that satisfy connectivity and area constraints. Apply adaptive enhancement gain to pixels within Adaptive enhancement gain Values ​​and corrected signal-to-noise ratio This inverse relationship allows weaker target candidate pixels with weaker signals to receive greater enhancement.

[0057] Step 8: Perform piecewise linear stretching of the radiative semantic interval on the base layer; For the base layer The grayscale histogram is analyzed, and the grayscale range is divided into several intervals based on the radiation semantic interval division rule. A linear stretching coefficient is applied to each interval to perform piecewise linear stretching. The radiation semantic interval division rule refers to dividing the grayscale range into several continuous sub-intervals based on the trough positions in the grayscale histogram. Each sub-interval corresponds to a scene semantic with similar radiation intensity in the image (such as low-temperature background, medium-temperature ground objects, high-temperature targets, etc.). The input grayscale range and output grayscale range are determined for each sub-interval, and an independent linear stretching coefficient is applied to adjust the contrast of each radiation semantic interval independently.

[0058] Furthermore, the method for detecting the valley positions in the grayscale histogram is as follows: for the base layer After statistically analyzing the grayscale histogram of the entire image, the grayscale histogram curve is smoothed to suppress statistical noise. Local minima are then searched on the smoothed grayscale histogram curve. Local minima that satisfy the minimum depth condition (i.e., the pixel frequency of the minimum point is lower than the pixel frequency of its two adjacent maximum points by a preset ratio) are identified as valid valleys. The grayscale value of the valid valleys is used as the segmentation boundary of the radiating semantic interval. When the number of detected valid valleys is zero, the algorithm degenerates into applying a single linear stretch to the entire grayscale range.

[0059] Extracting weak target candidate region mask In the base layer Mean gray value at the corresponding position Determine the grayscale mean The gray range it is located in, at the gray average value The grayscale range is further subdivided into target grayscale neighborhood sub-ranges and non-target sub-ranges. An increased stretching coefficient is applied to the target grayscale neighborhood sub-ranges, allowing for a larger dynamic range allocation of the grayscale range where the weak target is located, thus generating an enhanced base layer. The target gray-level neighborhood sub-interval refers to the sub-interval with the average gray level. The grayscale value range centered at a predetermined grayscale neighborhood radius is selected. This grayscale neighborhood radius should be chosen so that the target grayscale neighborhood sub-interval can cover the weak target candidate region mask. The main distribution range of pixel grayscale values ​​is determined, while avoiding excessively wide ranges that could cause non-target background areas to also receive excessive stretching coefficients.

[0060] Furthermore, the ratio between the increased stretching coefficient corresponding to the target gray-level neighborhood sub-interval and the original linear stretching coefficient of the radiative semantic interval containing the target gray-level neighborhood sub-interval reflects the dynamic range allocation gain for the weak target gray-level range. The upper limit of this ratio is constrained by the ratio of the total gray-level width of the radiative semantic interval containing the target gray-level neighborhood sub-interval to the width of the target gray-level neighborhood sub-interval, ensuring that the target gray-level neighborhood sub-interval does not exceed the output gray-level range corresponding to the radiative semantic interval containing the target gray-level neighborhood sub-interval after stretching.

[0061] Step 9: Perform inverse pyramid reconstruction and adaptive gamma mapping to output the final enhanced image; Enhance the base layer With each enhancement detail layer Perform inverse Laplacian pyramid reconstruction, upsample and stack layer by layer to synthesize a full-resolution enhanced image.

[0062] Furthermore, the specific process of the reverse reconstruction of the Laplace pyramid is as follows: to strengthen the base layer Starting with the base layer, we will enhance it. Upsampling to the After the layer resolution and the enhancement detail layer Add them together to get the first one. Layer reconstructed image; the first layer The reconstructed image is upsampled and combined with the detail enhancement layer. Add them together, iterating upwards layer by layer, until you reach the layer that enhances the details. The summation yields a full-resolution enhanced image. The upsampling operation and the downsampling operation used in the Laplacian pyramid decomposition stage correspond strictly in spatial resolution.

[0063] Adaptive gamma mapping is applied to the synthesized full-resolution enhanced image, based on the global gray-level mean of the full-resolution enhanced image. Adaptive determination of gamma parameter values : in, This is an intermediate reference value for the grayscale range (e.g., 128 for an 8-bit image). To enhance the global grayscale mean of the synthesized full-resolution image, To prevent extremely small positive numbers with zero values ​​from occurring in logarithmic operations. and The dimensions of all values ​​are grayscale values. and Both are dimensionless quantities obtained by taking the logarithm of the grayscale value; the ratio of the two is the gamma parameter value. The parameter is dimensionless, with consistent dimensions, making the calculation valid. When the global grayscale mean... When it is too low, the gamma parameter value Adaptive gamma mapping increases the grayscale value of dark areas; when the global grayscale mean... When it is too high, the gamma parameter value Adaptive gamma mapping compresses the grayscale values ​​of highlighted areas downwards. The output of adaptive gamma mapping... It is given by the following formula: in, To synthesize full-resolution enhanced images at location grayscale value at that location This is the maximum allowed grayscale value for the image (e.g., 255 for an 8-bit image). The adaptive gamma parameter value is determined by the above formula. The normalized grayscale value is a dimensionless quantity, expressed as a gamma parameter value. The exponentiation operation still results in a dimensionless quantity; multiplying by... The image is then restored to grayscale values, ensuring consistent dimensions and valid computation. The final enhanced infrared image is output. .

[0064] This implementation method corrects the confidence curve. An information transmission channel for correction quality perception was established between the non-uniformity correction stage and the image enhancement stage, enabling the two originally independent and sequential processing steps to form a collaborative processing relationship.

[0065] Specifically, this implementation method uses the effective diversity index. Simultaneously, the combined information of detector species count and switching frequency is considered, because relying solely on detector species count in reciprocating motion mode can lead to inflated counts due to alternating observations of a few pixels. The effective diversity index... Multiplying the number of species by the handover frequency ratio, the low species number factor helps avoid distortion in diversity assessment and improves the accuracy of reference frame estimation. The reliability assessment is more accurate.

[0066] This implementation method uses time-domain median absolute deviation. Robust outlier detection generates spatiotemporal masks for moving targets. Because of the median absolute deviation in the time domain Compared to standard deviation, it has a higher resistance to extreme values, thus accurately identifying and eliminating the radiation contribution of moving targets even when such signals are present. Combined with truncated mean estimation, it prevents moving target signals from interfering with reference radiation estimation. This avoids introducing background-biased correction errors into the target area.

[0067] This implementation method will correct the confidence graph. After being passed to the enhancement stage, the original local signal-to-noise ratio of the fine detail layer Corrected confidence plot Correction, because the area with larger correction residuals corresponds to the correction confidence map. The value is low, so the original local signal-to-noise ratio index of this region is inflated by non-uniform artifacts. The value is multiplied by a lower corrected confidence level plot. The value was then reasonably reduced, so that the artifact pixels no longer met the high requirement. The classification criteria avoid the problem of non-uniform artifacts being incorrectly classified into edge classes and thus requiring the application of enhancement coefficients. Simultaneously, the signal-to-noise ratio of weak targets in regions with low correction quality is improved. More likely to fall into the middle The range preserves the opportunity for targeted enhancement through connectivity analysis, so weak targets will not be misclassified as edge pixels or noise pixels due to residual interference and thus lose enhancement processing.

[0068] Furthermore, this embodiment achieves this by adjusting the center. Pixels undergo spatial connectivity analysis to filter candidate regions that match the size characteristics of weak targets. Because weak targets exhibit spatial clustering and their areas conform to physical constraints, the joint filtering based on connectivity and area constraints can distinguish weak target candidates from random noise distributions among pixels with moderate signal-to-noise ratios, in conjunction with adaptive gain enhancement. Stretching the grayscale neighborhood of the target in the base layer improves the contrast and visibility of weak targets while suppressing noise and artifacts.

[0069] like Figure 2-8 As shown, the method in this embodiment operates as follows: An airborne optoelectronic pod, equipped with a 640×512 pixel infrared focal plane array detector and an inertial measurement unit, performs multi-target simultaneous tracking missions. During a flight test in 20XX, the pod's line of sight rapidly switched between three ground-based heat source targets (Target A, Target B, and Target C), with a switching cycle of approximately 0.8 seconds, acquiring 40 consecutive frames (N=40) of raw infrared image sequences at a frame rate of 50 fps. The detector outputs 14-bit quantized grayscale values, operating in the 3-5 micrometer mid-wave infrared band. Target C, located in the lower right corner of the field of view, has relatively weak radiation intensity and is a small, distant heat source, making it the primary target for protection in this mission. The full motion threshold is preset to Dth = 3.0, the smoothing coefficient λ = 0.08, the total number of pyramid layers K = 4, the local window side lengths s = 7 (layer 1) and s = 11 (layer 2), the weak target size range [Amin, Amax] = [4, 36] pixels, and the weak target enhancement intensity control parameter αw = 1.2.

[0070] The image processor reads a sequence of 40 raw infrared images from the detector interface and simultaneously acquires angular rate data output from the inertial measurement unit. The angular rate data is integrated over time and converted into inter-frame pixel displacement by combining it with the optical system's focal length (equivalent focal length 180mm, pixel pitch 15μm). The 40 raw images are then registered to a unified spatial coordinate system, and the physical detector pixel number corresponding to each spatial position in each frame is recorded simultaneously. Four representative spatial positions in the field of view are selected for effective diversity index calculation: position A is located on the frequently switching path of the aiming line; position B is near target A; position C corresponds to the area where weak target C is located (edge ​​of the field of view); and position D is located in the central background area of ​​the field of view.

[0071] Taking position A as an example, the number of different detector pixel types observed in 40 frames is Ndet = 11, and the number of numbering switches between adjacent frames is Nswitch = 28. Substituting these values ​​into the formula: Taking location C (weak target area) as an example, Ndet = 3, Nswitch = 5: Table 1. Calculation results of effective diversity indices for four typical spatial locations: The Deff of positions A and B are both higher than Dth = 3.0, which means they belong to the region of full motion; the Deff of positions C and D are lower than Dth, which means they belong to the region of insufficient motion and require further interpolation.

[0072] For each spatial location in the registered image sequence, the temporal median and median absolute deviation (MAD) are calculated for 40 frames of observations. Taking location B as an example, target A passes through this location from frames 12 to 18, and the observation values ​​in the corresponding frames are significantly higher. The temporal median for this location is calculated as med = 4312 (grayscale value), MAD = 87, and the judgment threshold is 3 × MAD = 261. The observation value in frame 14 is 4891, which deviates from the median |4891 - 4312| = 579 > 261, and is marked as an outlier for moving targets. A spatiotemporal mask Mtarget(location B, 14) = 1 is generated. For spatial locations with Deff ≥ 3.0 and not marked by the mask, the effective observation values ​​are taken, and the first and last 10% extreme values ​​are removed to calculate the truncated mean, which is used as the reference radiometric estimate S(i,j). For regions with insufficient motion, such as positions C and D, spatial interpolation is performed using the estimated neighboring positions as constraint points, and a joint bilateral-guided filter is used to generate a complete reference frame Sfull(i,j).

[0073] Table 2. Reference radiation estimation results for typical locations: Based on the complete reference frame Sfull(i,j), the offset correction increment ΔO(m) for each detector pixel is calculated. Taking detector pixel M-0374 as an example, the average original output value of the spatial position observed by this pixel in the current frame (frame 40) is 4203, and the average reference radiometric estimate is 4156. Therefore, ΔO(M-0374) = 4203 - 4156 = 47. The original offset correction parameter Oold(M-0374) = 43, and the smoothing coefficient λ = 0.08, is updated as follows: Perform offset correction on the original image of the current frame (frame 40). The position (i, j) corresponds to pixel M-0374, and the original value R of frame 40 (i, j) = 4203. Table 3 shows the updated results of partial detector pixel offset correction parameters: Based on the effective diversity index Deff(i,j) from step 1 and the sufficient motion threshold Dth = 3.0, a corrected confidence map is generated. Four typical locations are used as examples: Position A: Location C (weak target area): Table 4. Calculation results of confidence maps for typical locations: The corrected confidence map Qconf(i,j) is completed. The confidence of the weak target C at location C is only 0.125. It will play a key role in suppressing the signal-to-noise ratio assessment of this area in the subsequent enhancement stage, preventing artifacts from being falsely enhanced.

[0074] A 4-layer Laplacian pyramid decomposition (K=4) is performed on the 40th frame (i, j) of the corrected image C to obtain detail layers L1, L2, L3 and a base layer B4. The original local signal-to-noise ratio (SNR) is calculated for the first detail layer (s=7) and the second detail layer (s=11). Two contrast locations are selected: location C (weak target region, low correction quality) and a non-uniform artifact location in the field of view (location E, low correction quality but artifacts exist).

[0075] Taking position E in the first detail layer as an example, its mean μ1 = 18.4, standard deviation σ1 = 6.2, and ε = 0.001 within its 7×7 local window: Taking position C in the first detail layer as an example, its mean μ1 = 4.1 and standard deviation σ1 = 3.8 within its 7×7 local window: The original local signal-to-noise ratio index was corrected using a calibrated confidence plot. The Qconf at location E is 0.148 (the effective diversity index Deff for this region is 0.444). Qconf = 0.125 at position C: If the global median SNR median1 of the full-image correction signal-to-noise ratio in the first detail layer is 0.312, then the classification thresholds are SNR median1 = 0.312 and 2×SNR median1 = 0.624.

[0076] Table 5. Typical Location Corrected Signal-to-Noise Ratio Calculation and Pixel Classification Results: The original signal-to-noise ratio (SNR) of position E (artifact) was 2.967, which could have been classified as high SNR and thus enhanced. After confidence correction, it dropped to 0.439, falling into the medium SNR range. Subsequent connectivity analysis will further filter out features that do not conform to the size of weak targets. Although the weak target signal at position C dropped to low SNR after correction, some of its neighboring pixels fell into the medium SNR range, preserving candidate opportunities for connectivity analysis.

[0077] Eight-connected component labeling is performed on the SNR pixels in layers 1 and 2. The area of ​​each connected component is calculated, and connected components with an area in the range of [4, 36] pixels are retained as weak target candidate regions mask M weak target (i, j). After connectivity analysis, the core pixels (a total of 14 pixels, the area of ​​which meets the constraints) of the region where weak target C is located are included in weak target M; the area of ​​the connected component of the artifact region at position E is 173 pixels, which exceeds Amax = 36, and is therefore excluded.

[0078] An adaptive enhancement gain is applied to the core pixels of weak target C within weak target M. Taking one pixel as an example, the SNR correction 1 = 0.351, and the median SNR 1 = 0.312. The original value of the pixel's first detail layer, L1, was 4.1. After enhancement: Apply βedge(1) = 1.6 to high SNR pixels, apply βflat(1) = 0.4 to low SNR pixels, and apply hierarchical adaptive gain (βedge(3) = 1.2, βflat(3) = 0.7) to the third detail layer based on gradient magnitude threshold classification.

[0079] Table 6 Example of pixel gain calculation for candidate region of weak target C: The mean grayscale value gtarget = 4071 is extracted from the weak target candidate region mask M, corresponding to the weak target (i, j) in the base layer B4 (i, j). After smoothing the grayscale histogram of B4, two effective valleys are detected, located at grayscale values ​​of 3200 and 5100 respectively. The grayscale range is divided into three radiometric semantic intervals: low-temperature background interval [0, 3200), medium-temperature ground feature interval [3200, 5100], and high-temperature target interval [5100, 16383]. gtarget = 4071 falls within the medium-temperature ground feature interval.

[0080] Within the mid-temperature land cover region, a target gray-level neighborhood sub-interval [3821, 4321] is defined with g target = 4071 as the center and a gray-level neighborhood radius of 250 as the target gray-level neighborhood. An increased stretching coefficient of 1.85 is applied to this sub-interval (the original stretching coefficient of the mid-temperature land cover region is 1.20). The other sub-intervals retain their original stretching coefficients, thus generating an enhanced base layer B4'(i,j).

[0081] Starting with the enhanced base layer B4'(i,j), layer-by-layer upsampling and superimposed enhancement detail layers L3', L2', and L1' are used to synthesize the full-resolution enhanced image I(i,j). The global grayscale mean g of the full-resolution enhanced image is calculated as mean = 3847. For a 14-bit image, g_median = 8192, g_maximum = 16383, ε = 0.001, and the adaptive gamma parameter values ​​are: Since the mean g = 3847 < the median g = 8192, ρ = 1.092 > 1. Gamma mapping will compress the highlight area and adjust the overall grayscale distribution. Taking the grayscale value I = 4160 of the synthesized enhanced image at the location of the weak target C as an example: The final enhanced infrared image F(i,j) is output.

[0082] Table 7 Key parameters and typical pixel output for pyramid inverse reconstruction and gamma mapping: The data flow in this example demonstrates a complete logical chain: In step 1, the effective diversity index Deff (0.375 for position C, 7.700 for position A) statistically analyzes the 40 registered images, directly driving step 4 to generate the correction confidence map Qconf (0.125 for position C, 1.000 for position A); In step 2, the truncated mean estimate of the moving target spatiotemporal mask after excluding interference from target A is updated exponentially in step 3 to form the offset correction parameters, outputting the 40th frame of the corrected image C; In step 6, the correction confidence map compresses the original signal-to-noise ratio of the artifact at position E from 2.967 to 0.439, blocking its path to be enhanced in the high SNR class; In step 7, the candidate pixels of weak target C obtain the highest adaptive gain of 1.551, which, combined with the priority stretching of the target gray-level neighborhood sub-intervals in step 8, enables weak target C to obtain targeted enhancement in both the base layer and detail layer channels; Finally, in step 9, the adaptive gamma mapping adaptively determines ρ = based on the global mean. Version 1.092 completes global brightness adjustment and outputs the final enhanced image, achieving consistent data transfer throughout the entire process from the original multi-frame sampling data to the final enhanced output.

[0083] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for adaptive non-uniformity correction and image enhancement processing of an infrared imaging device, characterized in that, Includes the following steps: Acquire a continuous sequence of multiple frames of raw images from an infrared focal plane array and angular rate data from an inertial measurement unit. Register the multiple frames of raw images to a unified spatial coordinate system. Count the number of detector pixel numbering types and the number of numbering switches between adjacent frames at each spatial location. Calculate the effective diversity index. A spatiotemporal mask for moving targets is generated based on the registered image sequence. Reference radiation is estimated for locations that are not marked by the mask and whose effective diversity index is greater than the sufficient motion threshold. For locations whose effective diversity index is less than the sufficient motion threshold, a complete reference frame is obtained by spatial interpolation. The offset correction parameters of the detector pixels are updated based on the complete reference frame, and non-uniformity correction is performed on the current frame to output the corrected image. A corrected confidence plot is generated based on the ratio of the effective diversity index to the sufficient motion threshold; The Laplacian pyramid decomposition is performed on the corrected image to obtain the detail layer and the base layer. The local signal-to-noise ratio index is calculated for the fine detail layer. The local signal-to-noise ratio index is corrected using the correction confidence map. Based on the corrected signal-to-noise ratio, the pixels are divided into three categories: high, medium, and low. Spatial connectivity analysis is performed on pixels with medium signal-to-noise ratio to extract weak target candidate regions and an adaptive enhancement gain is applied. Edge enhancement gain is applied to pixels with high signal-to-noise ratio and a noise suppression attenuation coefficient is applied to pixels with low signal-to-noise ratio. Piecewise linear stretching is performed on the base layer, and pyramid inverse reconstruction is performed on the enhanced detail layer and the base layer, followed by adaptive gamma mapping to output the final enhanced image.

2. The method of claim 1, wherein, The effective diversity index is calculated as follows: during the registration process, the physical detector pixel number corresponding to each spatial location in each frame is recorded synchronously to form a number sequence. The number sequence is used to count the number of different numbers that have appeared at each spatial location to obtain the number of detector pixel number types. The number sequence is used to compare the pixel number records of each spatial location in each frame with the numbers of adjacent frames and accumulate the number of times the numbers are different to obtain the number switching number. The effective diversity index is equal to the product of the number of detector pixel number types and the number of number switching number divided by the total number of frames to obtain the normalized switching frequency.

3. The method of claim 1, wherein the method further comprises: The generation of the moving target spatiotemporal mask includes: calculating the temporal median and temporal median absolute deviation for a multi-frame observation sequence at each spatial location in the registered image sequence, wherein the temporal median absolute deviation is the median of the absolute value of the difference between each frame observation and the temporal median in the time dimension; marking observations that deviate from the temporal median by more than three times the temporal median absolute deviation as outliers caused by the moving target, thereby generating the moving target spatiotemporal mask; the estimation of reference radiation is as follows: for spatial locations not marked by the moving target spatiotemporal mask and with an effective diversity index greater than the sufficient motion threshold, the effective observations are sorted numerically, and the extreme values ​​of the first and last preset proportions are removed, and the arithmetic mean of the remaining observations is calculated to obtain the reference radiation estimate.

4. The adaptive non-uniformity correction and image enhancement processing method of infrared imaging device according to claim 1, characterized in that, The process of obtaining a complete reference frame through spatial interpolation includes: using the spatial location of the completed reference radiometric estimation as a known constraint point, and performing spatial interpolation using a joint bilateral-guided filtering method. The joint bilateral-guided filtering method uses the latest corrected frame that has completed offset correction as the guide image. The interpolated value of the position to be interpolated is obtained by a weighted average of the surrounding known constraint points. The weights consider both spatial distance and gray-level similarity on the guide image. The closer the spatial distance of the constraint point, the higher the weight. The closer the gray-level value of the constraint point on the guide image is to the position to be interpolated, the higher the weight. When there is no available correction frame during system initialization, the guide image degenerates into a bilateral filtering method that only depends on spatial distance. After the first frame is corrected, it switches to the joint bilateral-guided filtering mode.

5. The adaptive non-uniformity correction and image enhancement processing method of infrared imaging device according to claim 1, characterized in that, The method for updating the offset correction parameters of detector pixels includes: for each detector pixel, calculating the offset correction increment as the average of the difference between the original output value of the detector pixel in the current frame and the estimated spatial position of the corresponding position in the complete reference frame across all observation positions of the detector pixel; updating the offset correction parameters using an exponentially weighted moving average method, wherein the updated offset correction parameters are equal to the difference between the original offset correction parameters multiplied by a factor minus a smoothing coefficient and the offset correction increment multiplied by a smoothing coefficient, where the smoothing coefficient is greater than zero and less than one; and subtracting the offset correction parameters of the detector pixel corresponding to each spatial position in the original image of the current frame from the updated offset correction parameters to obtain the corrected image.

6. The adaptive non-uniformity correction and image enhancement processing method of infrared imaging device according to claim 1, characterized in that, The correction confidence map is generated as follows: the correction confidence value for each pixel location is equal to the minimum of the ratio of the effective diversity index to the sufficient motion threshold at that location and one; the range of the correction confidence value is a closed interval from zero to one; the corrected signal-to-noise ratio is equal to the product of the local signal-to-noise ratio index and the correction confidence value; the classification of pixels into high, medium, and low categories includes: calculating the global median of the corrected signal-to-noise ratio across the entire image; pixels with a corrected signal-to-noise ratio greater than twice the global median are classified as high signal-to-noise ratio pixels; pixels with a corrected signal-to-noise ratio between the global median and twice the global median are classified as medium signal-to-noise ratio pixels; and pixels with a corrected signal-to-noise ratio less than the global median are classified as low signal-to-noise ratio pixels.

7. The adaptive non-uniformity correction and image enhancement processing method of infrared imaging device according to claim 1, characterized in that, The step of performing spatial connectivity analysis on the mid-signal-to-noise ratio pixels to extract weak target candidate regions includes: marking connected components in the binary mask image of the mid-signal-to-noise ratio pixels; counting the total number of pixels contained in each connected component as its area; retaining connected components whose area is within the weak target size range to form a weak target candidate region mask; the upper and lower limits of the weak target size range are predetermined based on the optical system parameters of the infrared imaging device, typical target size, and detection distance; the adaptive enhancement gain is equal to the product of a weak target enhancement intensity control parameter plus a subtracted correction signal-to-noise ratio divided by twice the global median; the enhanced detail layer value of the pixels within the weak target candidate region mask is the original detail layer value multiplied by the corresponding adaptive enhancement gain.

8. The adaptive non-uniformity correction and image enhancement processing method for infrared imaging equipment according to claim 1, characterized in that, The step of performing piecewise linear stretching on the base layer includes: smoothing the gray-level histogram of the base layer and detecting local minima; identifying local minima that meet the minimum depth condition as effective valleys; dividing the gray-level range into several intervals using the gray-level values ​​of the effective valleys as the segmentation boundaries of the radiative semantic intervals; applying a linear stretching coefficient to each interval; extracting the gray-level mean of the weak target candidate region mask at the corresponding position in the base layer; determining the gray-level interval where the gray-level mean is located; defining a target gray-level neighborhood sub-interval within the gray-level interval with the gray-level mean as the center; applying an increased stretching coefficient to the target gray-level neighborhood sub-interval to generate an enhanced base layer; and degrading to apply a single linear stretching to the entire gray-level range when the number of detected effective valleys is zero.

9. The adaptive non-uniformity correction and image enhancement processing method for infrared imaging equipment according to claim 1, characterized in that, The adaptive gamma mapping includes: calculating the global grayscale mean of the full-resolution enhanced image; the adaptive gamma parameter value is equal to the natural logarithm of the intermediate reference value of the grayscale range divided by the natural logarithm of the global grayscale mean plus a minimum positive constant; the output value of each pixel in the final enhanced image is equal to the maximum allowable grayscale value multiplied by the gamma parameter value raised to the power of the quotient of the pixel's grayscale value divided by the maximum allowable grayscale value; when the global grayscale mean is lower than the intermediate reference value of the grayscale range, the gamma parameter value is less than one, causing the grayscale value of the dark area to increase; when the global grayscale mean is higher than the intermediate reference value of the grayscale range, the gamma parameter value is greater than one, causing the grayscale value of the bright area to decrease.

10. An adaptive non-uniformity correction and image enhancement processing system for an infrared imaging device, used to execute the adaptive non-uniformity correction and image enhancement processing method for an infrared imaging device according to any one of claims 1 to 9, characterized in that, include: The data acquisition and registration module is used to acquire a continuous multi-frame raw image sequence of the infrared focal plane array and the angular rate data of the inertial measurement unit, register the multi-frame raw images to a unified spatial coordinate system, count the number of detector pixel number types and the number of number switching between adjacent frames at each spatial location, and calculate the effective diversity index. The reference frame estimation module is used to generate a spatiotemporal mask for moving targets based on the registered image sequence, estimate reference radiation for positions that are not marked by the mask and whose effective diversity index is greater than the sufficient motion threshold, and obtain a complete reference frame by spatial interpolation for positions whose effective diversity index is less than the sufficient motion threshold. The non-uniformity correction module is used to update the offset correction parameters of the detector pixels based on the complete reference frame, perform non-uniformity correction on the current frame, and output the corrected image. The confidence generation module is used to generate a corrected confidence map based on the ratio of the effective diversity index to the sufficient motion threshold. The confidence-aware classification module is used to perform Laplacian pyramid decomposition on the corrected image to obtain the detail layer and the base layer. It calculates the local signal-to-noise ratio index for the fine detail layer, corrects the local signal-to-noise ratio index using the corrected confidence map, and classifies the pixels into three categories: high, medium, and low based on the corrected signal-to-noise ratio. The detail layer adaptive enhancement module is used to perform spatial connectivity analysis on pixels with medium signal-to-noise ratio to extract weak target candidate regions and apply adaptive enhancement gain, apply edge enhancement gain to pixels with high signal-to-noise ratio, and apply noise suppression attenuation coefficient to pixels with low signal-to-noise ratio. The base layer stretching and reconstruction output module is used to perform piecewise linear stretching on the base layer, perform inverse pyramid reconstruction on the enhanced detail layer and base layer, and perform adaptive gamma mapping to output the final enhanced image.

Images