Infrared detector dead pixel detection method

Abstract

The invention relates to the technical field of infrared detection, and discloses an infrared detector dead pixel detection method, which comprises the following steps: acquiring a current scene image during the working period of an infrared detector and a plurality of frames of isothermal uniform surface background images acquired by the infrared detector at different working temperatures to form a background image sequence; performing singular point detection on each frame of image in the current scene image and the background image sequence by adopting an improved wavelet transform algorithm, screening out effective singular points, and calculating a singularity value of each effective singular point; based on the singular point detection result, the singularity value difference and historical background image data, static dead pixels, dynamic dead pixels and noisy points are distinguished through a multi-layer judgment rule, and a dead pixel detection result is obtained; and periodically evaluating the accuracy of the defective pixel detection result, and dynamically calibrating the detection parameters according to the evaluation result. Through collaborative analysis of the current working image and the multi-temperature background image sequence, dead pixels can be accurately detected in real time in the normal working process of the detector.

Description

Technical Field

[0001] This invention relates to the field of infrared detection technology, and specifically to a method for detecting defects in infrared detectors. Background Technology

[0002] Infrared detectors are the core sensors of infrared imaging systems, and their performance directly affects image quality. Due to manufacturing defects, material inhomogeneity, and long-term operation in complex temperature-changing environments, dead pixels can develop on the detector's focal plane array. Dead pixels are classified into two types: static dead pixels (fixed locations that do not change with time or temperature) and dynamic dead pixels (appearing or disappearing intermittently depending on the detector's operating temperature). Dead pixels manifest as abnormally bright / dark pixels in the image, severely reducing the image's signal-to-noise ratio and usability.

[0003] The current mainstream methods for detecting dead pixels have the following shortcomings:

[0004] 1. Low detection accuracy: Most methods rely on spatial domain filtering (such as median filtering) to reduce noise in advance, but the filtering process can easily smooth out real bad points (especially bad points with weak response), leading to missed detections; at the same time, strong random noise may be misjudged as bad points, leading to false alarms.

[0005] 2. Inability to effectively detect dynamic defects: Traditional methods often calibrate defects at a single temperature, failing to capture dynamic defects that change with temperature. Dynamic defects are widespread in practical applications and are a significant factor contributing to unstable image quality.

[0006] 3. Reliance on dedicated testing environment: High-precision defect detection typically requires uniform irradiation calibration using a large-area blackbody source in a laboratory environment. This method cannot be integrated into the normal operation of the detector and cannot achieve real-time, online defect monitoring and compensation.

[0007] 4. Disconnect between detection and compensation: Existing methods mostly focus on the detection stage and use a single interpolation compensation strategy for detected bad pixels. They do not consider the different requirements of bad pixel type (static / dynamic) and image region (edge / non-edge) for compensation effect, which may introduce new image distortion after compensation, especially in edge regions.

[0008] Therefore, there is an urgent need for an infrared detector defect detection method that can distinguish between static / dynamic defects and noise online, in real time, and with high precision, and can adaptively perform differentiated compensation. Summary of the Invention

[0009] This invention aims to overcome the shortcomings of existing technologies and provide a method for detecting defective pixels in infrared detectors. This method can acquire multi-temperature background images using the detector's own shutter without relying on an external blackbody calibration environment. Combined with improved wavelet transform and multi-layer judgment logic, it accurately distinguishes between static defective pixels, dynamic defective pixels, and random noise. Furthermore, this invention integrates adaptive parameter calibration and targeted defective pixel compensation strategies, achieving self-optimization of detection accuracy and intelligent improvement of image quality.

[0010] This invention is achieved through the following technical solution:

[0011] A method for detecting dead pixels in an infrared detector, comprising:

[0012] S1. Acquire the current scene image A during the operation of the infrared detector, as well as multiple frames of isothermal uniform surface background images collected by the infrared detector at different operating temperatures, to form a background image sequence. Where i is the collection sequence number and ;

[0013] S2. An improved wavelet transform algorithm is used to process the current scene image A and the background image sequence, respectively. The improved wavelet transform algorithm includes wavelet basis adaptive selection, two-dimensional wavelet transform and high-frequency detail component extraction, adaptive threshold setting and singularity filtering, and singularity value calculation.

[0014] S3, Based on the current scene image A and the background image sequence The results of singularity detection at the corresponding pixel position, the difference in singularity values, and historical background image data are used to distinguish static bad points, dynamic bad points, and noise through multi-layer judgment rules, resulting in bad point detection results that include bad point type, location, and corresponding temperature parameters.

[0015] S4. Periodically evaluate the accuracy of the defect detection results, dynamically calibrate the detection parameters based on the evaluation results, and feed the defect detection results back to the detector control unit.

[0016] As an optimization, the specific process of S1 is as follows:

[0017] S1.1 Acquire images of the detector's working scene according to the preset frame rate, and after pixel grayscale value calibration and image size normalization preprocessing, form the current scene image A and store it;

[0018] S1.2. Through a linkage control mechanism, the shutter is triggered to close when the detector's operating temperature changes to a preset temperature step size, thus acquiring a single frame of background image. The images are preprocessed, correlated with temperature parameters, and then stored in a background image database to form the background image sequence. The background image database adopts a cyclic overlay mechanism, retaining the most recent N frames of background images, where N is a preset positive integer.

[0019] As an optimization, the preset frame rate is 10-30 frames per second; and / or,

[0020] The preset temperature step size is: ; and / or,

[0021] The temperature parameter acquisition accuracy is ; and / or,

[0022] .

[0023] As an optimization, the improved wavelet transform algorithm specifically includes the following sub-steps:

[0024] S2.1 Adaptive selection of wavelet basis: Bioorthogonal wavelets are selected as the wavelet basis, and the wavelet decomposition scale of 3-5 levels is dynamically determined based on the gray mean and variance of the image.

[0025] S2.2 Two-dimensional wavelet transform and detail component extraction: Perform two-dimensional wavelet transform on the image to decompose it into a low-frequency detail component LL, and three high-frequency detail components: horizontal detail component LH, vertical detail component HL, and diagonal detail component HH.

[0026] S2.3 Adaptive Threshold Setting and Singularity Filtering: Calculate the standard deviation of grayscale values ​​in non-edge regions of each high-frequency detail component. As a noise intensity indicator, an adaptive threshold is set. , where k is an empirical coefficient of 2.5-3.5; the first neighborhood window is used to traverse each of the high-frequency detail components, and candidate singularities whose gray value of the center pixel of the first neighborhood window is a local maximum and exceeds the corresponding threshold T are selected, and the intersection of the candidate singularities in the three high-frequency detail components is taken as the effective singularity, thereby obtaining the effective singularity set;

[0027] S2.4 Singularity value calculation: Extract the wavelet coefficient modulus values ​​at the locations of the effective singular points under different wavelet decomposition scales, obtain the attenuation index by fitting the attenuation curve of the modulus value as a function of scale, and use the attenuation index as the singularity value.

[0028] As an optimization, step S2, after obtaining the effective singularity and before executing step S3, further includes:

[0029] For each of the aforementioned valid singular points, the valid singular point is designated as a point to be verified.

[0030] At least two different sizes of second neighborhood windows are used to perform secondary singularity determination at the location of the point to be verified.

[0031] The number of times the point to be verified is identified as a singular point in all second neighborhood windows is counted, and the confidence value C of the point to be verified is calculated based on the number of times.

[0032] The confidence value C is compared with a preset confidence threshold C0. If... If the point to be verified passes verification, it is determined that the point to be verified has passed verification, and the point to be verified is retained as the final valid singular point; if If the verification of the point to be verified fails, the point to be verified will be removed from the set of valid singular points.

[0033] As an optimization, the multi-layer decision rule is a three-layer decision rule. Based on the valid singularities obtained after filtering and verification in step S2, the following three-layer decision rule is executed:

[0034] First-level judgment: If a pixel position is a valid singular point in the current scene image A, then in the background image sequence... The latest single-frame background image in China If a pixel is not a valid singularity, then that pixel location is determined to be noise.

[0035] Second-level judgment: If a pixel's position is within the current scene image A and the latest single-frame background image... If all of them are valid singular points, then the position of this pixel in the current scene image A and the latest single-frame background image is calculated respectively. The difference between the corresponding singularity values Set the difference threshold ,like If a pixel is identified as a static defect, it is determined to be a static defect; otherwise, it is determined to be a noise pixel. This represents the singularity value of the pixel location in the current scene image A. The pixel position in a single frame background image The corresponding singularity value;

[0036] The third layer of judgment: If a pixel position is not a valid singular point in the current scene image A, then in the latest single-frame background image... If the pixel in the middle is a valid singularity, then retrieve the pixel position in the historical background image. to The detection results are as follows: if the pixel location was a valid singular point in any frame of the historical background image, then the pixel location is determined to be a dynamic bad point; otherwise, the pixel location is determined to be noise.

[0037] As an optimization, the difference threshold The value range is 0.1-0.3.

[0038] As an optimization, the specific process of S4 is as follows:

[0039] S4.1 Parameter Adaptive Calibration:

[0040] S4.1.1 Periodically calculate the accuracy of the defect detection results and compare the accuracy with a preset accuracy threshold;

[0041] S4.1.2 If the accuracy is lower than the preset accuracy threshold, then based on the direction and degree of the accuracy deviation, adjust the wavelet decomposition scale, singular point detection threshold coefficient k, and bad point judgment difference threshold by a preset adjustment step size. At least one parameter in the background image acquisition temperature step size should be increased or decreased.

[0042] S4.1.3 Using the newly acquired verification images of a preset number of M frames after parameter adjustment, repeat steps S2 to S3 to calculate the new accuracy to verify the effect of parameter adjustment, until the accuracy reaches or exceeds the preset accuracy threshold, where M is a positive integer.

[0043] S4.2 Adaptive compensation for bad pixels:

[0044] Based on the type, location information, and corresponding temperature parameters in the defect detection results, the corresponding compensation strategy is invoked to perform pixel compensation.

[0045] As an optimization, S4.2 calls the corresponding compensation strategy for pixel compensation, specifically including:

[0046] S4.2.1 Compensation strategy initialization:

[0047] A pre-stored library of differentiated compensation strategies associated with bad pixel types and image regions is provided. The strategy library includes: a conventional neighborhood interpolation strategy for static bad pixels in non-edge regions, a weighted neighborhood interpolation strategy for static bad pixels in edge regions, and a temperature-prediction-based pre-compensation strategy for dynamic bad pixels.

[0048] S4.2.2 Static defect compensation execution:

[0049] For pixels classified as static defects, the image region where the static defect is located is determined. If it is located in a non-edge region, the conventional neighborhood interpolation strategy is called. Centered on the pixel of the static defect, the gray values ​​of all normal pixels within the preset compensation neighborhood window of the pixel of the static defect are taken and interpolated to obtain the compensation value. If it is located in an edge region, the weighted neighborhood interpolation strategy is called. Centered on the pixel of the static defect, the gray values ​​of each pixel within its preset compensation neighborhood window are taken and interpolated with preset weight coefficients to obtain the compensation value. Pixels belonging to the edge direction are given higher weights.

[0050] S4.2.3, Dynamic dead pixel compensation execution:

[0051] For pixels classified as dynamic defects, the pre-compensation strategy is invoked. Based on the temperature parameters associated with the dynamic defect, the temperature-defect mapping relationship established based on historical data is queried. When the absolute value of the difference between the detector's real-time operating temperature and the temperature threshold indicated in the mapping relationship where the dynamic defect is prone to occur is less than or equal to a preset warning temperature difference... When the temperature threshold is reached, the compensation parameters pre-stored for that temperature threshold are automatically invoked to pre-compensate the pixel of the dynamic defect.

[0052] As an optimization, the conventional neighborhood interpolation strategy is as follows:

[0053] Centered on the static bad pixel to be compensated, take the gray values ​​of all normal pixels within the preset compensation neighborhood window of the static bad pixel, calculate the arithmetic mean of the gray values, and use the arithmetic mean as the compensation value to replace the original gray value of the bad pixel.

[0054] The weighted neighborhood interpolation strategy is as follows:

[0055] Centered on the pixel of the static defect to be compensated, select the pixels within its preset compensation neighborhood window;

[0056] After assigning a weight coefficient to the grayscale value of each pixel within the compensation neighborhood window, the weighted average value is calculated as the compensation value; wherein, the assignment rule of the weight coefficient is: adjacent pixels in the edge tangent direction of the pixel to which the static defect belongs are assigned a higher weight coefficient than pixels in other directions.

[0057] The pre-compensation strategy is as follows:

[0058] Based on the temperature threshold at which dynamic bad pixels are likely to occur obtained from the query, compensation parameters that have been verified to be effective under this temperature threshold are retrieved from the historical compensation records; the compensation parameters include at least the size of the compensation neighborhood window, the weight coefficient distribution, or the direct optimal compensation value; when the pre-compensation trigger condition is met, the compensation parameters are directly applied to replace the grayscale value of the pixels of the dynamic bad pixels.

[0059] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0060] This invention, through collaborative analysis of the current working image and multi-temperature background image sequences, combined with an improved wavelet transform algorithm, can detect bad spots in real time and accurately during the normal operation of the detector without interrupting operation or relying on an external blackbody source. The detection accuracy is significantly higher than that of traditional filtering and denoising methods.

[0061] This invention creatively utilizes multi-temperature background image sequences and their historical data to design a three-layer judgment rule, which can effectively distinguish between static bad pixels, dynamic bad pixels, and random noise, solving the industry problem that traditional methods are not sensitive to dynamic bad pixels.

[0062] This invention enables the method to adapt to different detector models and the performance drift of detectors over time by periodically evaluating the detection accuracy and adjusting key algorithm parameters (such as wavelet decomposition scale and detection threshold). It has strong versatility and robustness.

[0063] This invention employs differentiated compensation strategies based on the type and location of defective pixels. For static defective pixels in edge regions, weighted interpolation is used to preserve details, while for dynamic defective pixels, pre-compensation is performed based on temperature prediction. This eliminates defective pixels while maximizing the preservation of the original image details and spatial continuity, thereby improving overall image quality.

[0064] This invention adds a multi-scale neighborhood verification step after core detection, and further filters out false singularities through confidence evaluation, making the data in the input judgment stage purer and fundamentally reducing the false judgment rate. Attached Figure Description

[0065] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0066] Figure 1 This is a flowchart of a method for detecting dead pixels in an infrared detector according to the present invention. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0068] This embodiment 1 provides a method for detecting dead pixels in an infrared detector, such as... Figure 1 As shown, steps S1-S4 are included. Next, these four steps will be described in detail.

[0069] S1. Acquire the current scene image A during the operation of the infrared detector, as well as multiple frames of isothermal uniform surface background images collected by the infrared detector at different operating temperatures, to form a background image sequence. Where i is the collection sequence number and .

[0070] This step involves image acquisition, the core of which is to simultaneously acquire the current scene image A, which reflects the actual scene information, and the background image sequence, which reflects the detector's own background response at different temperatures. .

[0071] In some embodiments, the specific process of S1 is as follows:

[0072] S1.1 Acquire images of the detector's working scene according to the preset frame rate, and after pixel grayscale value calibration and image size normalization preprocessing, form the current scene image A and store it.

[0073] The current scene image A is the normal imaging result of the detector on the observed target scene, which is used to analyze the bad spots that may appear under real working conditions.

[0074] The infrared detector operates continuously at its nominal frame rate. The output of the image sensor (focal plane array) is converted from analog to digital to form a raw digital video stream. The image acquisition unit (typically an FPGA or a dedicated image interface chip) captures single frames from this video stream at a preset frame rate (20 frames / second in this embodiment). This frame rate is chosen to balance real-time requirements with the computational load of the processing system. The captured raw images undergo a standardized preprocessing pipeline to ensure the accuracy and consistency of subsequent algorithm processing.

[0075] 1. Non-uniformity correction and grayscale calibration: Eliminate fixed pattern noise caused by inconsistencies in the responsivity and dark current of different pixels in the detector. Apply the correction coefficients calibrated at the detector's factory. A two-point correction method is typically used, and the calculation formula is as follows: ;

[0076] Where (x,y) are pixel coordinates. The values ​​are the original grayscale values, and Gain and Offset are pre-calibrated gain and bias coefficient matrices. This is the grayscale value after pixel coordinate correction. Through this step, the response of each pixel to the same radiative input is calibrated to be consistent.

[0077] 2. Image Size Normalization: Ensures all input images have the same spatial resolution, facilitating subsequent pixel-level comparison and indexing. Determine the original image resolution. If it matches the preset resolution (640×512 in this example), it passes directly. If not, a bilinear interpolation algorithm is used to scale the image to the preset resolution. This preset resolution must match the effective number of pixels of the detector or be an integer multiple thereof.

[0078] The preprocessed image is officially labeled as the current scene image A. This image is then sent to a first-in, first-out (FIFO) temporary image buffer. This buffer is typically designed to hold 1-5 frames, and its update cycle is synchronized with or slightly faster than the acquisition cycle of subsequent background images, ensuring that the latest scene image is always available when analysis is needed.

[0079] S1.2. Through a linkage control mechanism, the shutter is triggered to close when the detector's operating temperature changes to a preset temperature step size, thus acquiring a single frame of background image. The images are preprocessed, correlated with temperature parameters, and then stored in a background image database to form the background image sequence. The background image database adopts a cyclic overlay mechanism, retaining the most recent N frames of background images, where N is a preset positive integer.

[0080] Background image sequence This is the key innovative data source of this invention. It acquires images of a uniform radiation background at different operating temperature points through the detector's own structure, which are used to separate the detector's inherent defects (bad pixels) from the scene content.

[0081] Using platinum resistance thermometers (PT100 / PT1000) or integrated digital temperature sensors (such as the TI TMP117), Accurate real-time monitoring of the core temperature of the infrared detector chip package The sensor mounting location requires thermal design to ensure that its temperature measurement point can quickly and accurately reflect the actual temperature of the focal plane array. A normally open mechanical shutter is integrated into the detector's optical path. The shutter blades are coated with high-emissivity black paint, allowing them to function as an isothermal radiation surface approaching that of a blackbody when closed.

[0082] During system initialization, the initial temperature of the detector is recorded. And set a temperature change step size. (This embodiment is set as) ). System continuous monitoring. When detected hour( (based on the temperature at the time of the last successful acquisition of the background image), immediately trigger the following ordered operations:

[0083] 1. Shutter control: Sends a closing command to the shutter driver. The shutter closes completely within ≤50ms, blocking external scene light so that the detector's focal plane receives only infrared radiation from the closed shutter blades.

[0084] 2. Steady-state waiting: After the shutter closes, wait for a short period of time to stabilize (e.g., 100ms) to eliminate mechanical vibrations caused by the shutter action and transient response of the detector output.

[0085] 3. Image Acquisition: Under steady-state conditions, acquire one frame of the detector output image with the shutter closed. This image represents the detector output at the current detector temperature. Background image of an isothermal homogeneous surface.

[0086] 4. Shutter Reset: After data acquisition, the shutter is immediately reopened to restore the detector's normal observation function. The entire shutter operation cycle (closing-acquisition-opening) is controlled within 200-300ms, minimizing its impact on normal observation.

[0087] 5. Background image preprocessing and information association:

[0088] The acquired raw background image undergoes the exact same non-uniformity correction and size normalization preprocessing steps as the current scene image A. After processing, a usable single-frame background image is generated. .

[0089] Will The detector temperature value at the time this data acquisition was triggered Establish a strong association to form a data pair The linked data pairs The images are stored in a dedicated background image database. This database can physically be a circular buffer in memory or a circular storage area in flash memory. Each entry contains: image data. Temperature value timestamp And the acquisition sequence number i. i is used to identify the background image acquisition event triggered by the i-th temperature change reaching the step size.

[0090] The database has a fixed capacity, meaning it retains the most recent N frames of background images. In this embodiment, the lower limit is set to N=10. When a new background image... When data needs to be stored: if the current number of stored entries is less than N, it is appended directly. If N entries have already been stored, the oldest record (i.e., the record with acquisition sequence number iN) is overwritten. This mechanism ensures that the database always dynamically maintains the detector's background response samples during the most recent temperature change period. For example, if... If N=10, then the database covers approximately The temperature range is sufficient to capture the "appearance-disappearance" temperature characteristics of most dynamic defects.

[0091] The system maintains a logical pointer that always points to the latest background image in the database. (i.e., the frame with the largest value of i). When determining bad pixels, the current scene image A will be compared with this latest frame. Real-time comparison is performed. Simultaneously, the entire historical sequence of the database is analyzed. This will be used for tracing and identifying dynamic defects. This design ensures that the background data used for analysis is highly correlated with the detector's current operating state (temperature).

[0092] Through the detailed implementation process described above, this invention, without introducing an external blackbody source, cleverly utilizes the detector's own shutter to automatically construct a high-quality, uniform background image library that is updated synchronously with the operating temperature, laying an irreplaceable data foundation for subsequent high-precision defect identification.

[0093] S2. An improved wavelet transform algorithm is used to process the current scene image A and the background image sequence, respectively. The improved wavelet transform algorithm includes wavelet basis adaptive selection, two-dimensional wavelet transform and high-frequency detail component extraction, adaptive threshold setting and singularity filtering, and singularity value calculation.

[0094] Bad pixels appear as local gray-level abrupt changes in an image. The core idea of ​​this method is to use an improved wavelet transform to enhance the extraction of such abrupt features and to improve the reliability of detection through multi-scale spatial verification.

[0095] In some embodiments, the improved wavelet transform algorithm specifically includes the following sub-steps:

[0096] S2.1 Adaptive selection of wavelet basis: Bioorthogonal wavelets are selected as the wavelet basis, and the wavelet decomposition scale of 3-5 levels is dynamically determined based on the gray mean and variance of the image.

[0097] The purpose of this step is to select suitable time-frequency analysis basis functions for the wavelet transform operation and dynamically configure the depth of its multi-resolution analysis.

[0098] Wavelet basis selection principle: Infrared images typically have a flat background, low signal-to-noise ratio, and bad pixels appear as pulse-like singularities. Bioorthogonal wavelets (such as Bior 2.4, Bior 3.5, etc.) are chosen as wavelet bases because of their linear phase characteristics and good regularity, which allow for more accurate location of singularities and reduced reconstruction distortion. In this embodiment, the Bior 3.5 wavelet is selected by default.

[0099] Decomposition scale adaptive determination (refining "based on the gray-level mean and variance of the image"): Calculate the global gray-level mean of the input image I. and variance The determination of the decomposition scale J follows these rules:

[0100] If the overall contrast of the image is low and the details are flat ( A deeper decomposition is needed to capture subtle bad spot features that may be overwhelmed by noise at a higher scale (coarser resolution), so J=5 is set in this case.

[0101] If the image contrast is moderate ( ), set J=4. This is typical for most outdoor scenarios.

[0102] If the image has high contrast and rich detail ( A shallower decomposition can effectively extract significant singular features while avoiding computational redundancy; in this case, J=3 is set.

[0103] This dynamic mechanism ensures that the algorithm can adapt to different scene content and always analyzes at an appropriate scale.

[0104] S2.2 Two-dimensional wavelet transform and detail component extraction: Perform two-dimensional wavelet transform on the image to decompose it into a low-frequency detail component LL, and three high-frequency detail components: horizontal detail component LH, vertical detail component HL, and diagonal detail component HH.

[0105] This step breaks the image down into different frequency subbands to separate the background and details.

[0106] Transformation operation: Perform J-level two-dimensional discrete wavelet transform on the preprocessed image. Each level of decomposition divides the image into four sub-bands:

[0107] Low-frequency approximate components ( (): It carries the main energy of the image and reflects a large area of ​​background and slowly changing information.

[0108] Horizontal high-frequency detail components ( ): Highlights horizontal edges and abrupt changes in the image.

[0109] Vertical high-frequency detail components ( ): Highlights the vertical edges and abrupt changes in the image.

[0110] Diagonal high-frequency detail components ( ): Highlights edges and abrupt changes in the image along the diagonal direction.

[0111] Where j represents the decomposition level, from 1 (most refined) to J (most coarse).

[0112] Outlier pixels, as locally isolated gray-level abrupt changes, primarily contain energy in the high-frequency detail components (LH, HL, HH). The low-frequency component (LL) mainly contains background and uniform region information, where outlier pixel features are severely diluted. Therefore, subsequent outlier detection (S2.3 and S2.4) will be performed entirely and solely based on these three high-frequency detail components, discarding the low-frequency component (LL). This strategy significantly reduces data processing volume and focuses on target features.

[0113] S2.3 Adaptive Threshold Setting and Singularity Filtering: Calculate the standard deviation of grayscale values ​​in non-edge regions of each high-frequency detail component. As a noise intensity indicator, an adaptive threshold is set. Where k is an empirical coefficient of 2.5-3.5; the first neighborhood window is used to traverse each of the high-frequency detail components, and candidate singularities whose gray value of the center pixel of the first neighborhood window is a local maximum and exceeds the corresponding threshold T are selected, and the intersection of the candidate singularities in the three high-frequency detail components is taken as the effective singularity, thereby obtaining the effective singularity set.

[0114] This step aims to identify genuine singularity candidates from high-frequency noise.

[0115] Noise intensity estimation: Directly calculating the standard deviation of the entire detail component is susceptible to interference from real edges and structures in the image, leading to an overestimation of noise levels. First, for each high-frequency detail component image (e.g., ... The gradient magnitude of the edge is calculated using the Sobel edge detection operator, and a low threshold is set to mark obvious edge pixels, generating a binary edge mask. Pixel regions marked as edges by the mask are ignored, and the standard deviation of the gray values ​​of the remaining non-edge pixels is calculated, denoted as . , , (For each scale j). This The value more purely reflects the intensity of image noise at that scale. For simplification, values ​​from various scales are often used. The average value is used as the final noise index for this component. , , .

[0116] For each high-frequency detail component (LH, HL, HH), based on its noise intensity Calculate the detection threshold for this component: Where k is an empirical coefficient, and in this embodiment, the initial value is set to 3.0. This threshold is dynamically adjusted according to the image noise level, ensuring the stability of detection sensitivity under different imaging conditions.

[0117] For each of the three high-frequency detail component images, a 3×3 first-neighborhood window is used for full-image sliding traversal. For the window center pixel (x,y) and its grayscale value I(x,y), determine whether they simultaneously satisfy:

[0118] Local extremum condition: I(x,y) is the maximum gray value of the 9 pixels within the current 3×3 window.

[0119] Significance condition: (T is the adaptive threshold corresponding to the detail component where the pixel is located).

[0120] If satisfied, then in a binary labeling graph of the current component, mark position (x,y) as 1 (candidate singularity).

[0121] After obtaining the candidate singularity marker images for each of the three detail components (LH, HL, and HH), a logical AND operation is performed on these three images. That is, a pixel is only confirmed as a valid singularity if it is marked as a candidate point in all three images (LH, HL, and HH). This strict intersection operation requires that singularities exhibit significant abrupt changes in the horizontal, vertical, and diagonal directions, thus greatly suppressing false alarms caused by noise stripes or textures in a single direction. All points selected through this step constitute the initial set of valid singularities, and their pixel coordinates are recorded.

[0122] S2.4 Singularity value calculation: Extract the wavelet coefficient modulus values ​​at the locations of the effective singular points under different wavelet decomposition scales, obtain the attenuation index by fitting the attenuation curve of the modulus value as a function of scale, and use the attenuation index as the singularity value.

[0123] This step calculates a quantitative index for each valid singularity, characterizing the significance and stability of its "bad point" properties.

[0124] For each point P (with coordinates ) in the set of valid singular points On all J decomposition scales obtained by the previous wavelet transform, locate their corresponding positions in the three high-frequency subbands of LH, HL, and HH, respectively. The wavelet coefficients (due to downsampling, the coordinates need to be scaled). The coefficient with the largest absolute value among the three is taken as the wavelet coefficient modulus at that point at that scale j, denoted as . (j=1,2,...,J).

[0125] Theoretically, for an ideal singularity (such as a bad point), the modulus of its wavelet coefficients is... It exhibits power-law decay as the decomposition scale j increases (i.e., the observation becomes "coarser"). (Plotting) A scatter plot of scale j was generated, and a linear fit was performed using the least squares method to obtain the slope of the fitted line. .

[0126] The quantization value of the singularity at this point is defined as The larger the S-value, the slower the wavelet coefficient magnitude decays with scale, meaning the singular feature remains significant across multiple scales, and is more likely to be a real, stable bad point. Conversely, points with small S-values ​​may only exhibit singular features at the finest scales, and are more likely to be random noise.

[0127] In some embodiments, after obtaining the valid singularity and before executing step S3, step S2 further includes:

[0128] For each of the aforementioned valid singular points, the valid singular point is designated as a point to be verified.

[0129] At least two different sizes of second neighborhood windows are used to perform secondary singularity determination at the location of the point to be verified.

[0130] The number of times the point to be verified is identified as a singular point in all second neighborhood windows is counted, and the confidence value C of the point to be verified is calculated based on the number of times.

[0131] The confidence value C is compared with a preset confidence threshold C0. If... If the point to be verified passes verification, it is determined that the point to be verified has passed verification, and the point to be verified is retained as the final valid singular point; if If the verification of the point to be verified fails, the point to be verified will be removed from the set of valid singular points.

[0132] This step enhances and removes false positives from the core detection algorithm. It is performed after obtaining the initial set of valid singular points and before formally entering the bad point determination step S3. Its purpose is to utilize spatial multi-scale consistency to eliminate accidental and unstable detection results.

[0133] For example, a set of second neighborhood windows is predefined, with a size larger than the first neighborhood window used in the detection phase. In this embodiment, three sizes are used: 3×3, 5×5, and 7×7. Simultaneously, a confidence threshold C0 = 0.67 (approximately 2 / 3) is set.

[0134] Point-by-point verification process:

[0135] Initialization: Traverse the initial set of valid singular points and mark each point P as a point to be verified.

[0136] Multi-scale spatial verification: On the original grayscale image I (i.e., the image after preprocessing but without wavelet transform), verification is performed sequentially using 3×3, 5×5, and 7×7 windows centered on the point to be verified:

[0137] Validation criterion: Calculate the average grayscale value of all pixels within the current window. and standard deviation Determine whether the gray value I(P) of the center point P satisfies: .in, This is a lenient threshold, for example, set to 2.0. This criterion aims to examine whether point P is a statistical outlier within the local window.

[0138] Record: If the criteria are met, record the decision result of this window as "Yes" (singularity); otherwise, record it as "No".

[0139] Confidence score calculation: Count the number of "yes" results in the three windows, denoted as n. Calculate the confidence score C = n / 3 for the point P to be verified.

[0140] like Point P is considered to exhibit singularity across multiple spatial scales and is a high-confidence candidate point. It is retained and added to the final set of valid singular points.

[0141] like The singularity of point P is considered to be unstable or only appears occasionally at a specific scale, and is likely to be noise or texture interference, so it is removed.

[0142] By requiring singularities to not only be significant in the frequency domain (wavelet transform) but also maintain statistical anomalies in multi-scale neighborhoods in the spatial domain, their authenticity is doubly confirmed. Experiments show that this step can filter out more than 90% of false singularities, significantly improving the quality of input data for subsequent bad point detection and is a key step in reducing the overall false detection rate.

[0143] At this point, for image A and each frame of the background image... Each of these methods yields a final set of valid singular points, containing pixel locations and corresponding singularity values ​​S. This data will serve as direct input for the next stage, "bad pixel determination."

[0144] S3, Based on the current scene image A and the background image sequence The system uses the singularity detection results, singularity value differences, and historical background image data at the corresponding pixel positions to distinguish between static bad points, dynamic bad points, and noise through multi-layer judgment rules, resulting in bad point detection results that include bad point type, location, and corresponding temperature parameters.

[0145] This invention fully utilizes the differences between the working image and the multi-temperature background image, as well as historical information in the time dimension, by designing a three-layer judgment rule.

[0146] Before starting pixel-by-pixel determination, the system needs to complete the following data preparation:

[0147] 1. Set of valid singular points in the current scene image A: Obtain the final set of valid singular points in the current scene image A from step S2 and the output after verification through the second neighborhood window, denoted as... Each element in this set contains pixel coordinates (x, y) and its corresponding singularity value. .

[0148] 2. Latest single-frame background images Valid set of singularities: From the background image database, obtain the background image with the latest timestamp (i.e., the largest acquisition sequence number i). (For example The final set of valid singular points corresponding to ) is denoted as Similarly, the set contains coordinates and singularity values. .

[0149] 3. Record of valid singular points in historical background images: For each frame of historical image in the background image database (e.g., ... This maintains a record of the corresponding set of valid singularities. These records constitute a historical singularity occurrence archive indexed by temperature and / or time.

[0150] The system will iterate through each pixel position (x, y) of the detector and execute the following three-layer decision logic in sequence.

[0151] First-level judgment: Fast noise filtering:

[0152] Judgment criterion: The pixel position (x, y) belongs to (i.e., identified as a valid singularity in the current working image A), but not belonging to (i.e., a uniform background image in a new frame) (This is not a valid singularity).

[0153] Judgment logic: A true defect is an inherent flaw in the detector hardware, and its abnormal response characteristics should be stable. Under uniform radiation conditions ( (This excludes interference from scene content; a bad pixel should consistently behave as a singularity.) Therefore, if a point only behaves as a singularity in A containing a complex scene, but not in a uniform scene... If the object disappears from the scene, its peculiar characteristics are most likely caused by local high-contrast objects in the scene (such as point heat sources or edges) or instantaneous random noise, rather than detector defects.

[0154] Judgment result: This pixel location is determined to be noise. Its type is marked as NOISE, and the subsequent judgment process for this pixel ends.

[0155] Second-level judgment: Static bad pixel identification and noise depth discrimination:

[0156] Entry condition: Pixel position (x, y) simultaneously belongs to and (i.e., the current scene image A and the latest single-frame background image) All of them were detected as valid singularities.

[0157] Core calculation: Obtain the singularity values ​​of the point in the two sets, respectively. and Calculate the absolute value of the difference between the two. .

[0158] Judgment threshold: Set a key difference threshold. This embodiment is based on a large amount of experimental data and sets... .

[0159] Decision logic:

[0160] like This indicates a significant difference in the singularity value of this point under different scene images and uniform backgrounds. Although it behaves as a singular point under both conditions, its degree of singularity is unstable. This instability usually means that the anomalous response of this point may be modulated or superimposed by specific structures or non-uniform lighting in the scene, and it is more likely to be a background-sensitive anomalous pixel (such as a non-linear response pixel) rather than a completely failed bad pixel. To simplify the classification, this method classifies it as an anomaly that requires attention, but to distinguish it from fixed bad pixels, it is still classified as a static bad pixel, although it can be treated differently in subsequent compensation strategies (this part is scalable).

[0161] like This indicates that the quantization value of the singularity of this point is highly consistent under two drastically different imaging conditions (with and without a scene). This strongly suggests that the anomalous response of this point is inherent, stable, and independent of the pattern of external incident radiation. This is consistent with the core characteristics of static bad pixels (such as dead pixels and hot pixels), namely, that their defects are fixed.

[0162] Judgment result:

[0163] when When this occurs, it is determined to be a static bad pixel, and the type is marked as STATIC_DEAD (which can be further subdivided into bright spots or dark spots).

[0164] when When a point is detected in both images but has unstable features (which may be due to slight noise or coincidences with weak scene structure), it is identified as a noise point and categorized as NOISE. This step removes points that are detected in both images but have unstable features from the list of bad points, further reducing the false alarm rate.

[0165] Third-level judgment: Dynamic bad pixel capture

[0166] Entry condition: Pixel position (x, y) does not belong to However, it belongs to (i.e., only on the latest uniform background) The point in question is a valid singularity, but it appears normal in the current scene image A.

[0167] Historical tracing: The system automatically queries the historical frame set of the background image database for the pixel position (x, y). The detection record in the image is checked to see if it has been recorded as a valid singularity in any previous frame of the historical background image.

[0168] Decision logic:

[0169] 1. If it exists in the historical records: This indicates that the "dead pixel" characteristic of this pixel is not related to the current temperature. This is a one-off, accidental event. It also exhibited anomalies at certain temperature points in the past (corresponding to the acquisition temperature of historical background images). This intermittent behavior that occurs with temperature changes is a typical characteristic of dynamic dead pixels. Dynamic dead pixels are usually related to the physical properties of the detector material or structure that change with temperature (such as resistance, tunneling current, etc.).

[0170] If it does not exist in the historical record: then this pixel is only present at the current temperature. The following behavior exhibits a singularity, whereas it behaved normally throughout its preceding temperature history. This is more likely due to a one-off transient interference, momentary noise in the readout circuit, or random errors during background acquisition.

[0171] Judgment result:

[0172] If it exists in the historical record, it is identified as a dynamic bad pixel, marked as DYNAMIC, and its current "activated" temperature is recorded. And the associated historical temperature.

[0173] If it does not exist in the history, it is judged as noise and the type is marked as NOISE.

[0174] After completing the three-layer judgment for all pixels, the system integrates and generates the final bad pixel detection result. This result is a structured data list that contains at least the following fields:

[0175] Pixel coordinates (x, y);

[0176] Bad point types: STATIC_DEAD, DYNAMIC, NOISE, NORMAL;

[0177] Associated Temperature: For DYNAMIC type, record the current activation temperature. For all types, the current detector temperature can be added as a reference.

[0178] Singularity value: and / or (Optional, for subsequent analysis or compensation parameter calculation)

[0179] Timestamp.

[0180] This list of defect detection results serves as both the final output of this detection cycle and the input for the next "Adaptive Optimization and Compensation" module. It can also be used to update historical records, providing richer evidence for future assessments. Through this progressive, multi-dimensional (spatial, frequency, temperature, time) assessment logic, this invention achieves high-precision, low-false-error reliable identification of both static and dynamic defective pixels.

[0181] S4. Periodically evaluate the accuracy of the defect detection results, dynamically calibrate the detection parameters based on the evaluation results, and feed the defect detection results back to the detector control unit.

[0182] This step forms a complete closed loop from detection to processing to optimization, which is the key to the high practicality and long-life stability of this method.

[0183] S4.1 Parameter Adaptive Calibration:

[0184] This step aims to address the issue of needing to manually adjust algorithm parameters due to individual detector differences, aging drift, or environmental changes. By establishing a closed-loop control loop based on performance feedback, the automatic optimization of key parameters can be achieved.

[0185] S4.1.1 Accuracy Statistics and Evaluation Cycle:

[0186] Evaluation period setting: The system performs performance evaluation at fixed time or frame intervals. In this embodiment, an evaluation period is set to be every 100 frames of the current scene image A (approximately 5 seconds of processing time). This period needs to be much longer than the background image acquisition period to ensure the statistical significance of the evaluation samples.

[0187] Accuracy Definition and Statistics: Within each evaluation period, the system generates 100 bad pixel detection results. To assess its accuracy, a combination of manual calibration and automatic statistical analysis is used.

[0188] 1. Randomly select 20 frames from the detection results corresponding to these 100 frames as evaluation samples.

[0189] 2. Experienced operators will visually identify and mark all confirmed bad pixels (including static and dynamic ones) on these 20 frames of images using high-precision displays and auxiliary tools, forming a baseline truth value list.

[0190] 3. Compare the 20 test results automatically generated by the system with the baseline true value list. Calculate the precision as the accuracy metric.

[0191] Accuracy P = (TP) / (TP + FP);

[0192] Wherein, TP (True Positive) is the number of bad pixels correctly detected by the system (consistent with manual annotation), and FP (False Positive) is the number of normal pixels that the system misidentifies as bad pixels (i.e., false alarms).

[0193] Threshold comparison: The calculated accuracy P is compared with a preset accuracy threshold. Comparison. This embodiment sets... This threshold represents the minimum acceptable performance standard for the system.

[0194] S4.1.2 Parameter Adjustment Decision and Execution:

[0195] If the statistical accuracy P is lower than If this occurs, the parameter calibration process is triggered. The system needs to determine the main cause of the performance degradation (missed detection or false alarm) and decide on the adjustment direction and target accordingly.

[0196] If (TP+FN) / GT is low (where FN is the number of missed detections and GT is the total number of manually labeled data), indicating a low detection rate, it suggests that the system is too conservative and may have a large number of missed detections. The main problem is that bad pixels are not being effectively identified.

[0197] If the number of false alarms (FPs) is significantly high, indicating a high false alarm rate, it means the system is overly sensitive and misclassifies a large amount of noise or texture as bad pixels. The main problem is the excessive number of false alarms.

[0198] Adjusting the strategy and step size: Based on the diagnostic results, adjust the wavelet decomposition scale, singularity detection threshold coefficient k, and bad point judgment difference threshold. At least one key parameter in the temperature step size of the background image acquisition is adjusted in a targeted, small-amplitude iterative manner. The preset adjustment step size is as follows:

[0199] Wavelet decomposition scale J: Adjust the step size to ±1 levels. If there are many missed detections, increase J (e.g., from 4 to 5) to capture features at a coarser scale; if there are many false alarms, decrease J (e.g., from 4 to 3) to reduce noise interference.

[0200] Singularity detection threshold coefficient k: The adjustment step size is ±0.2. If there are many missed detections, decrease k (e.g., from 3.0 to 2.8) to lower the detection threshold; if there are many false alarms, increase k (e.g., from 3.0 to 3.2) to raise the detection threshold.

[0201] Defect detection difference threshold Adjust the step size to ±0.05. If many defects are missed (especially static dead pixels), it may be due to... If the value is slightly larger than the threshold and is mistakenly identified as noise, the value can be increased appropriately. (e.g., from 0.2 to 0.25); if there are many false alarms (the second layer of judgment classifies noise as static bad pixels), then reduce the value. (e.g., from 0.2 to 0.15), tighten the judgment criteria.

[0202] Background image acquisition temperature step size Adjust the step size to ±0.1℃. If dynamic defect detection is unsatisfactory, it may be due to insufficient temperature sampling density; this can be addressed by reducing the step size. (e.g., from 1.0℃ to 0.9℃) to obtain denser temperature samples; if the system load is too high or the background image changes too frequently, the range can be increased appropriately.

[0203] Perform adjustments: Each calibration typically selects only the parameter with the most significant impact on the current accuracy for adjustment, following the principle of starting with the easier parameters and progressing to the more difficult ones (e.g., prioritizing the adjustment of k and ...). The adjustment command is then sent to the corresponding algorithm module.

[0204] S4.1.3 Validation and Iteration of Adjustment Effects:

[0205] Parameter adjustments should not be made blindly; they must be verified online to ensure their effectiveness.

[0206] Validation Set Acquisition: After parameter adjustment, the system immediately begins acquiring a new, independent set of validation images. This validation set consists of M=10 consecutive new scene images A. These images were acquired after the new parameters took effect, ensuring the correlation between the validation data and the adjusted state.

[0207] Re-execution and evaluation: For these M verification images, completely re-execute steps S2 (singularity detection) and S3 (bad pixel determination) to generate new detection results. Similarly, calculate the accuracy of these M frame results. .

[0208] Iteration termination condition:

[0209] like This indicates that the parameter adjustments were effective and the system performance has returned to normal. The calibration process is now complete, and the new parameters are applied until the next periodic evaluation.

[0210] like This indicates that a single adjustment failed to resolve the issue. The system will then... and The discrepancy and the new TP / FP distribution are identified, and the diagnostic and adjustment process in S4.1.2 is performed again, followed by verification with new M-frame images. This evaluation-adjustment-verification cycle continues until the accuracy target is met, thus ensuring the robustness and reliability of the calibration.

[0211] S4.2 Adaptive compensation for bad pixels:

[0212] This step translates the detection results into actual image quality enhancement actions. Its core idea is to adopt the most targeted compensation strategy based on the physical characteristics and spatial context of the bad pixels. The specific process is as follows:

[0213] S4.2.1 Initialization of the compensation strategy library:

[0214] During startup or initialization, a compensation strategy lookup table is built in the differential compensation strategy library (implemented via memory), which serves as the basis for subsequent compensation execution.

[0215] Policy Index: The policy library is indexed using a composite primary key of bad pixel type and image region type.

[0216] Strategy content:

[0217] 1. Standard Neighborhood Interpolation Strategy: The association key is (STATIC_DEAD, non-edge region). This strategy is defined as a basic operation.

[0218] 2. Weighted Neighborhood Interpolation Strategy: The association key is (STATIC_DEAD, edge region). This strategy includes additional weight allocation rules.

[0219] 3. Temperature Prediction-Based Pre-Compensation Strategy: The association key is (DYNAMIC, *) (* indicates no region distinction). This strategy includes temperature query, trigger judgment, and parameter calling logic.

[0220] S4.2.2 Static defect compensation execution:

[0221] For each pixel that is determined to be STATIC_DEAD:

[0222] 1. Region Determination: The Canny edge detection algorithm is used to extract edges from the current scene image A, generating a binary edge map. Centered on the coordinates of the defective pixel, a 3×3 neighboring window is checked to see if any edge pixels exist. If they do, the defective pixel is determined to be located in an edge region; otherwise, it is determined to be a non-edge region.

[0223] 2. Non-edge region compensation (conventional neighborhood interpolation):

[0224] Operation: Centered on the defective pixel p, take its preset 5×5 compensation neighborhood window (window size is configurable).

[0225] Calculation: Iterate through all pixels within the window except p. If a pixel is not marked as a bad pixel of any type, it is considered a "normal pixel". Collect the grayscale values ​​of all normal pixels and calculate their arithmetic mean.

[0226] Replacement: Assign this arithmetic mean directly to pixel p to complete the compensation.

[0227] 2. Edge region compensation (weighted neighborhood interpolation):

[0228] Operation: Similarly, take a 5×5 compensation neighborhood window centered on the defective pixel p.

[0229] Edge direction analysis: Calculate the image gradient direction at pixel p (using the Sobel operator). This image gradient direction is the normal direction of the edge. The tangent direction perpendicular to the normal direction is the edge direction that needs to be protected (i.e., the edge direction).

[0230] Pixel filtering: Traverse all pixels within the compensation neighborhood window. Only pixels that are not marked as any type of bad pixel (i.e., normal pixels) are included in the subsequent weight allocation and weighted average calculation. The bad pixel to be compensated, as well as other bad pixels in its neighborhood, are excluded.

[0231] Weighting:

[0232] For two adjacent pixels within the compensation neighborhood window that are located along the edge tangent direction (i.e., along the edge direction), a higher weight is assigned. .

[0233] For pixels in other directions within the compensation neighborhood window, assign standard weights. .

[0234] The defective pixel itself and other pixels within the window marked as defective pixels have a weight of 0 and are not included in the calculation.

[0235] Calculation and Replacement: Calculate the weighted average of the grayscale values ​​of all normal pixels within the compensation neighborhood window, and replace the original value of p with this weighted average as the compensation value. This method effectively utilizes edge continuity information, avoiding blurred or broken edges after compensation. The calculation formula is as follows: , where i represents all the filtered normal pixels.

[0236] S4.2.3 Dynamic bad pixel compensation execution:

[0237] For each pixel q that is determined to be DYNAMIC, the compensation is preventative and data-driven.

[0238] 1. Temperature-Defect Pixel Mapping: The system continuously learns in the background. Whenever a dynamic defect pixel q is identified, its activation temperature is recorded. After accumulating data over multiple work cycles (several hours or days), a set of activation temperature samples for q can be obtained. By performing kernel density estimation or cluster analysis (such as DBSCAN) on these samples, the predominantly active temperature ranges (temperature thresholds) for q can be determined. This temperature range represents the temperature-defect mapping relationship of q, which is the temperature threshold at which dynamic defects are likely to occur.

[0239] 2. Pre-compensation parameter learning: within the above temperature range Within the system, when q represents a bad pixel, the system attempts to apply different compensation parameters (such as different window sizes and different weighting schemes) for compensation, and evaluates the compensation effect in the next frame of the background image (such as the difference between the compensated pixel value and the background). The set of compensation parameters with the best evaluation effect (such as the smallest difference) is saved as the historical best compensation parameters for q within that temperature range. These parameters may be a special 7×7 window shape, a set of asymmetric weight distributions, or even a directly learned fixed grayscale compensation value (i.e., the direct best compensation value).

[0240] 3. Pre-compensation triggering and execution:

[0241] Real-time monitoring: The system continuously reads the real-time temperature of the detector. .

[0242] Trigger judgment: Set a warning temperature difference .when satisfy or When the temperature is close to the active range of the dynamic bad point q, it is considered that the temperature has approached the active range of the bad point q.

[0243] Perform pre-compensation: Once the triggering condition is met, the system immediately retrieves the value of q from storage. The optimal compensation parameters are pre-stored within the interval and directly applied to pixel q in the current frame image to replace its grayscale value.

[0244] The generation and application of optimal compensation parameters specifically include:

[0245] A. Parameter learning phase:

[0246] In the temperature range Within the context, when a dynamic defective pixel q actually appears in the background image, the compensation learning module performs the following steps:

[0247] 1. Multi-strategy parallel experiment: Multiple candidate compensation strategies are used simultaneously to repair q, including:

[0248] (a) Direct assignment strategy: Try to replace it with multiple candidate fixed gray values ​​{V1,V2,...}.

[0249] (b) Fixed weight interpolation strategy: Try to use multiple sets of preset neighbor pixel weight coefficients {W1,W2,...} for weighted averaging.

[0250] (c) Model-based interpolation strategy: Try calling different interpolation algorithms (such as Gaussian weighted interpolation, bicubic interpolation, and anisotropic diffusion repair) and adjust their parameters.

[0251] 2. Quantitative evaluation of the effect: In the uniform background image acquired later, where the temperature is still within the range, calculate the absolute difference or squared difference between the pixel value at position q after compensation by each strategy and the ideal background value.

[0252] 3. Optimal Parameter Selection: Select the compensation strategy that yields the best evaluation results (minimum difference) and its corresponding specific parameters (i.e., fixed parameters). or weight set Algorithm identifier and parameter set This parameter is bound and stored as the optimal compensation parameter for bad point q within that temperature range.

[0253] B. Parameter Application Phase (Pre-compensation Execution):

[0254] When the real-time temperature meets the pre-compensation trigger condition, the system executes:

[0255] 1. Parameter call: Reads the optimal compensation parameters pre-stored for the bad point q in this temperature range.

[0256] 2. Parameter parsing and execution:

[0257] 2.1 If the parameter is the direct optimal compensation value Then the grayscale value of pixel q in the current frame will be directly replaced with .

[0258] 2.2 If the parameter is a weighted coefficient distribution Then, taking q as the center, the neighborhood is calculated according to... The real-time grayscale values ​​of normal pixels in the neighborhood are weighted and averaged, and the result is assigned to q.

[0259] 2.3 If the parameters are algorithm identifier and parameter set Then the corresponding interpolation algorithm is called, and the input is... Using the neighborhood information of q in the current image, calculate the compensation value and assign it to q.

[0260] Specifically, if the parameters are algorithm identifier and parameter set Then the corresponding interpolation algorithm is called, and the input is... The process of calculating the compensation value and assigning it to q based on the neighborhood information of q in the current image is as follows:

[0261] 2.3.1 A compensation algorithm registry was established, and the specific process is as follows:

[0262] 2.3.1.1 Algorithm Registration: During the system initialization phase, all supported compensation algorithms (such as Gaussian weighted interpolation, bicubic interpolation, anisotropic diffusion repair, etc.) are registered with the central registry. Each registration entry includes:

[0263] Algorithm identifier: A unique string or number code that identifies the algorithm (e.g., "GAUSSIAN_INTERP").

[0264] Algorithm function pointer: Points to the memory address of the specific computation function of the algorithm.

[0265] Parameter template: Defines the type and structure of the parameters required by the algorithm (e.g., the Gaussian algorithm requires "window size" and "standard deviation σ").

[0266] 2.3.1.2 Unified Interface: All algorithms follow the same calling interface specification, namely: input parameters and image data, output a compensated pixel value. This allows the system to schedule different algorithms in a unified manner.

[0267] 2.3.2 Parameter Set Composition and storage:

[0268] During the learning phase, when the system determines that the optimal compensation method for bad pixels q within a certain temperature range is a specific algorithm, the stored... It contains two parts of information:

[0269] Algorithm identifier: Explicitly specifies which registered algorithm to use (e.g., identifier "GAUSSIAN_7x7").

[0270] Algorithm-specific parameters: These are the specific configuration values ​​required for the algorithm's computation. These parameters are stored in structured key-value pair format. For example:

[0271] For the "GAUSSIAN_7x7" algorithm, the parameters might be: {"window_size": 7, "sigma": 1.5};

[0272] For the "BICUBIC" (bicubic interpolation) algorithm, the parameters might be: {"b":0.0,"c":0.75} (parameters of the bicubic function);

[0273] For the "ANISO_DIFFUSION" (anisotropic diffusion repair) algorithm, the parameters might be: {"iterations":5,"kappa":30};

[0274] After this information is serialized, it is associated with the coordinates of the bad point q and its corresponding temperature range, and stored in the system's non-volatile memory (such as flash memory).

[0275] 2.3.3 Real-time pre-compensation execution process:

[0276] When the detector's real-time temperature meets the pre-compensation trigger condition, the system performs the following steps for the bad pixel q:

[0277] Step 1: Parameter loading and parsing:

[0278] The system retrieves the corresponding value from storage based on the coordinates of q and the current temperature range. The data packet is then parsed to separate the algorithm identifier and the specific parameter key-value pairs.

[0279] Step 2: Algorithm search and instantiation:

[0280] The system matches the parsed algorithm identifier against the compensation algorithm registry. Once found, the system knows which specific function to call for calculation, and according to the parameter template defined by the algorithm, performs the calculation. The key-value pairs in the memory are converted into a memory data structure (such as a struct) that the function can use directly.

[0281] Step 3: Neighborhood Data Preparation:

[0282] The system uses the coordinates of the bad pixel q in the current real-time image as the center, and follows... The system extracts a 7x7 pixel block from the image cache using the specified window size (e.g., window_size: 7). Simultaneously, the system examines every pixel within this block: if a pixel is also marked as a bad pixel (whether static or dynamic), it is marked as "invalid"; only pixels marked as "normal" will have their grayscale values ​​used in subsequent calculations.

[0283] Step 4: Algorithm Scheduling and Computation:

[0284] The system will pass the prepared neighboring pixel data (including valid / invalid markers) and the instantiated parameter structure to the algorithm calculation function found in step 2.

[0285] This function performs calculations based on its internal logic and the passed-in parameters. For example:

[0286] A. If the algorithm is Gaussian weighted interpolation, it will generate a 7×7 Gaussian weight matrix based on sigma=1.5, and then only for pixels marked as "normal" in the neighborhood, it will multiply their gray values ​​by the corresponding weights and sum them up, and finally divide by the total weights to obtain the compensation value.

[0287] The core of this implementation lies in constructing a dynamic, effective Gaussian weighting system based solely on normal pixels.

[0288] Input: Algorithm parameters The ImagePatch structure of the pixel q to be compensated (containing 7×7 neighborhood grayscale data and pixel status markers).

[0289] Output: Compensated pixel grayscale values.

[0290] Specific steps:

[0291] 1. Generate the basic Gaussian kernel: Using the center of the 7×7 window as the origin (0,0), calculate the Gaussian weight at each integer coordinate (u,v) (u,v∈[-3,3]):

[0292] ,in Then Normalize them so that the sum is 1.

[0293] 2. Construct an effective weight mask: Generate a binary mask matrix M of the same size based on ImagePatch.pixel_status. If pixel_status[i][j] is marked as "normal", then M(i,j)=1; if it is marked as "bad pixel / invalid" (including the center point q itself), then M(i,j)=0.

[0294] 3. Calculate the effective Gaussian weight matrix: Multiply the base Gaussian kernel element-wise with the mask to obtain the effective weight matrix. This operation forces the weight of the bad pixel location to zero, preventing it from participating in the calculation.

[0295] 4. Calculate the compensation value:

[0296] Molecular: Calculates the grayscale value of all normal pixels in the neighborhood and its corresponding value in the local area. The sum of the products of the corresponding weights: , where I(i,j) is the grayscale value.

[0297] Denominator: Calculation The sum of all weights in the middle: .

[0298] Compensation value: .

[0299] denominator It is dynamically changing, ensuring that even if there are multiple bad pixels in the neighborhood causing the total effective weight to be less than 1, the final compensation value can still maintain brightness consistency through normalization, avoiding being too dark or too bright.

[0300] B. If the algorithm is a bicubic interpolation algorithm.

[0301] The key to this implementation lies in simulating the fitting process of a bicubic surface within the local neighborhood where invalid pixels exist, and using valid pixel sample points to estimate the center missing value.

[0302] Input: Algorithm parameters (Standard bicubic parameters); ImagePatch of the pixel q to be compensated.

[0303] Output: Compensated pixel grayscale values.

[0304] Specific steps:

[0305] 1. Establish the interpolation region and sampling points: Treat the 7×7 ImagePatch as a continuous local coordinate plane, with the center point q located at (0, 0). Define the 4×4 core region required for interpolation (usually the 16 nearest neighbors). Check the pixel_status corresponding to the positions of these 16 preset sampling points.

[0306] 2. Handling invalid sampling points:

[0307] If a preset sampling point is marked as "invalid", then the grayscale value of that point will not be used directly.

[0308] Search for the nearest pixel in a smaller neighborhood (e.g., 3×3) around the invalid point that is in a "normal" state and use its grayscale value as the replacement value for the sampling point.

[0309] If no alternative pixel can be found, the weight of that sampling point will be set to zero in subsequent calculations.

[0310] 3. Perform bicubic interpolation calculation:

[0311] Based on the bicubic interpolation formula, the value of the interpolation point (0,0) is obtained by weighted summation of the gray values ​​of the surrounding 4×4 sampling points. The weights are determined by the horizontal and vertical distances (dx,dy) between the sampling point and (0,0), and are adjusted by parameters b and c.

[0312] The weighting function R(x) is usually: , for , for ;0,otherwise}, where a is usually -0.75 (corresponding to b=0, c=0.75).

[0313] Final compensation value , where (i,j) iterates through 4×4 sampling points after step 2, and I(i,j) is its gray value.

[0314] This paper transforms global bicubic interpolation into a locally adaptive sample point inpainting technique that bypasses invalid pixels. By dynamically finding alternative sampling points, it can still construct a smooth fitting surface to estimate the center value even when local samples are incomplete (with bad pixels), preserving more detail than simple neighborhood interpolation.

[0315] C. If the algorithm is anisotropic diffusion, it will start with the neighborhood of the bad point q as the initial state, and perform several diffusion iterations (iterations=5) based on parameters such as kappa=30, and finally use the value of the center point after the iteration is stable as the compensation value.

[0316] The core of this implementation is to constrain anisotropic diffusion within a very small local window and use the central bad pixel as the repair target, with its diffusion flux coming only from reliable pixels.

[0317] Input: Algorithm parameters ImagePatch for the pixel q to be compensated.

[0318] Output: Compensated pixel grayscale values.

[0319] Specific steps:

[0320] 1. Initialization: Copy the grayscale data from ImagePatch to the working buffer. (t=0 represents the 0th iteration). Initialize the value of the central bad point q. Set to the median gray value of all normal pixels in its neighborhood (not the mean, to improve robustness to outliers).

[0321] 2. Perform customized diffusion iterations (from t=0 to iterations-1):

[0322] a. Gradient Calculation and Correction: For each normal pixel p within the window, excluding the center point q, calculate its gradient with its four direct neighbors (top, bottom, left, and right). Key Correction: If a neighboring pixel in a certain direction is marked as "invalid," ignore the gradient contribution in that direction (treat the gradient as zero).

[0323] b. Diffusion coefficient calculation: For each effective gradient direction d of pixel p, calculate the Perona-Malik diffusion coefficient: Kappa=30 controls edge sensitivity. This represents the gradient of the image grayscale value I in a specific direction d.

[0324] c. Pixel value update (non-center point): Update the value of the non-center normal pixel p according to the diffusion equation: ,in Let the step size be , and then sum them up. Only traverse the valid gradient directions. This represents the new grayscale value of pixel p after the (t+1)th iteration. This represents the new grayscale value of pixel p after the t-th iteration. This represents the diffusion coefficient in the direction d.

[0325] d. Special update for the center bad pixel q: The value update of the center pixel q is driven only by its surrounding normal neighboring pixels N(q): .in, The gradient is the distance from the normal pixel s to the center q, and the diffusion coefficient c is calculated based on this gradient. This ensures that repair information only "flows" into the bad pixel from its reliable neighbors. Represents a set The summation operator. This represents the diffusion coefficient from the normal pixel s towards the center q. Let represent the gradient from the normal pixel s to the center q during the t-th iteration.

[0326] 3. Output compensation value: After completing the preset 5 iterations, take the final center point value. As compensation value.

[0327] This invention transforms a partial differential equation method for overall image smoothing and edge preservation into a precise, directionally protected point-based repair tool for isolated pixel defects. By strictly controlling the gradient source (only from normal pixels) and the center point update rule, it effectively fills in defective pixels near complex textures or edges while strictly avoiding damage to the original image structure.

[0328] Step 5: Pixel value replacement

[0329] The algorithm function returns the calculated compensation value. The system immediately replaces the original grayscale value at the location of the bad pixel q in the current frame's output image with this value.

[0330] 2.3.4 Exception Handling and Degradation Strategies

[0331] To ensure system robustness, a backup mechanism is designed:

[0332] Algorithm search failed: If If the algorithm identifier is not found in the registry (e.g., the algorithm changes after a software upgrade), the system will automatically downgrade and call a preset, reliable default algorithm (such as a simple 3×3 mean interpolation) to complete the compensation and record the log.

[0333] Calculation failure: If an error occurs during the algorithm calculation (such as invalid parameters or all neighboring pixels being invalid), the system will adopt a conservative strategy, such as replacing the value of the nearest normal pixel around the bad pixel in the current frame.

[0334] The system learns in advance The algorithm retrieves the specified algorithm and parameters from the algorithm registry, then performs real-time and precise calculations on the neighborhood of the defective pixel, ultimately completing pixel restoration. This mechanism upgrades the compensation strategy from simple fixed-value replacement to complex, data-driven adaptive image restoration, greatly improving the technical ceiling of the compensation effect.

[0335] In summary, the system preemptively corrects dynamic dead pixels (q) before they actually appear as abnormal bright or dark spots in the image due to temperature reaching a critical point. As a result, the transient defects caused by dynamic dead pixels are completely eliminated in the image observed by the end user, achieving seamless image output.

[0336] Through the collaborative work of modules S4.1 and S4.2, this invention can not only detect defective pixels with high precision, but also enable the detection system to continuously optimize itself and repair image defects in the most intelligent way, ultimately ensuring that the infrared detector outputs stable and high-quality images throughout its entire life cycle.

[0337] To more clearly illustrate the specific embodiments and technical effects of this invention, the following detailed explanation is provided using a specific detector model (a certain type of 640×512 uncooled infrared focal plane detector). This case study will fully demonstrate the entire process from image acquisition and bad pixel detection to adaptive compensation.

[0338] I. System Deployment and Initialization

[0339] 1. Hardware platform: The software system of the method of the present invention is integrated into the embedded image processing board (such as a platform based on TI TDA4VM or Xilinx Zynq series FPGA+ARM architecture) of the detector.

[0340] 2. Parameter initialization:

[0341] Image parameters: Preset frame rate FPS=20, current image buffer capacity is 1 frame, and image normalization resolution is set to 640×512.

[0342] Background acquisition parameters: Temperature acquisition accuracy Temperature change step size The background image database has a capacity of N=10.

[0343] Detection algorithm parameters: wavelet basis initialized to bior3.5; decomposition scale J=4; singularity detection coefficient k=3.0; judgment difference threshold. .

[0344] Compensation parameters: The neighborhood window for static defect compensation is 5×5; the pixel weights in the edge direction during weighted interpolation. ,other Dynamic defect warning temperature difference .

[0345] II. Work Process

[0346] Assume the detector starts operating at an ambient temperature of 20.0℃, and the target scenario is outdoor park monitoring.

[0347] Step A: Synchronous Image Acquisition

[0348] The system continuously outputs non-uniformity corrected images of the park scene (current scene image A) at 20 frames per second.

[0349] Meanwhile, a temperature sensor monitors the chip temperature. When the temperature rises to... (Changes) (The shutter closes, capturing the first frame of a uniform background image.) (related temperature) And store it in the database. Subsequent temperature changes... That is, triggering data collection ( ), ( ...The database always scrolls and retains the 10 most recent frames.

[0350] Step B: Singularity Detection (using current scene image A and background) ( (For example)

[0351] 1. Wavelet Transform: For A and Perform a 4-level two-dimensional wavelet transform to obtain their respective high-frequency detail components LH, HL, and HH.

[0352] 2. Adaptive threshold filtering: Calculation Standard deviation of non-marginal regions in LH component Therefore, the threshold T = 3.0 * 4.2 = 12.6. Iterate through the LH components and find the pixels with a local maximum grayscale value > 12.6. Perform the same operation on the HL and HH components, and take the intersection of the three. 150 candidate singularities have been preliminarily identified.

[0353] 3. Multi-scale verification: For the aforementioned 150 points, verification was performed on the original image using 3×3, 5×5, and 7×7 windows. For example, if a point (100, 200) was identified as an outlier in all three windows, its confidence level C = 1.0, and it was retained; while another point (150, 300) was only identified in the 3×3 window, with a confidence level C = 0.33, and it was discarded. Finally, Effective set of singularities There are 138 points remaining.

[0354] 4. Calculate singularity values: For For each point, extract the wavelet coefficient modulus at four scales, and fit the attenuation curve to obtain the singularity value S. Assume the point (100, 200) has... .

[0355] Step C: Defect Pixel Determination:

[0356] For any pixel position (x, y) on the detector:

[0357] Scenario 1: If (100, 200) is also a valid singularity in A ( If the condition is met, then proceed to the second level of judgment. Calculation .because It was determined to be noise.

[0358] Scenario 2: If (150, 250) is a valid singular point in both A and B_10, and , ,but It is also judged as noise.

[0359] Scenario 3: If (300, 400) is normal in A, but... If the point is a valid singularity, then proceed to the third level of judgment. A query of the historical database reveals that this point is... ( ), ( It was also a singular point in the data. Therefore, it was determined to be a dynamic bad point, and its activation temperature range was recorded as approximately [value missing]. .

[0360] Scenario 4: If a point is located at A and B... All of them are valid singularities, and , , If it is, then it is determined to be a static bad pixel.

[0361] Step D: Adaptive compensation execution

[0362] 1. Static Defect Pixel Compensation: For pixels identified as static defects, first use the Canny operator to determine if they are located in a non-edge region. Then, use a standard 5×5 neighborhood interpolation to replace the pixel with the average grayscale value of 125 from the surrounding 24 normal pixels.

[0363] 2. Dynamic bad pixel pre-compensation: For dynamic bad pixels (300, 400), the system has learned that they are in... It is easy to occur within the range, and the optimal compensation parameter is "Gaussian weighted interpolation (...)". When the detector temperature rises to 27.5℃ ( ,satisfy When the conditions are met, the system automatically performs Gaussian weighted interpolation pre-compensation on the point in the real-time video stream, and the user's viewing experience is free of any abnormal flickering.

[0364] III. Adaptive Optimization Case Study

[0365] After the system had been running for a period of time, the statistical accuracy during the evaluation period dropped to 92% (below the 95% threshold). Analysis revealed an increase in false alarms, mainly due to image noise caused by recent foggy weather. Increase.

[0366] Calibration action: The system automatically increases the singularity detection threshold coefficient k from 3.0 to 3.2, making the threshold T larger and the screening more stringent.

[0367] Using the newly acquired 10 frames of verification images after adjustment, the accuracy rate rebounded to 96%, and the parameter adjustment was successful.

[0368] IV. Technical Effect Verification Data

[0369] Continuous testing for 48 hours on this model of detector (ambient temperature) After a loop, the following data was obtained:

[0370] Performance indicators Method of the present invention Traditional single-temperature calibration + median filtering method Improvement effect Static defect detection accuracy 98.5% 85% An increase of 13.5 percentage points Dynamic dead pixel detection accuracy 97.2% 70% (and cannot be reliably identified) An increase of 27.2 percentage points Dead pixel false detection rate 1.8% 15% A decrease of 13.2 percentage points PSNR after edge dead pixel compensation 42.5 dB 40.7 dB Improved by 1.8 dB Dynamic dead pixel visibility time 0 frames 2-3 frames / time Completely eliminate Single-frame full-process processing time ≤ 50ms ≤ 20ms Meets the 20fps real-time requirement

[0371] This embodiment demonstrates that the method of the present invention, for this type of detector, achieves ultra-high precision detection and intelligent compensation for static and dynamic defects without the need for blackbody calibration, significantly improving the output image quality and stability of the infrared imaging system, and possessing significant engineering application value.

[0372] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting bad pixels of an infrared detector, characterized in that, Comprise: S1, acquiring a current scene image A during working of an infrared detector, and a plurality of frames of isothermal uniform surface background images collected by the infrared detector at different working temperatures to form a background image sequence wherein i is an acquisition serial number and i≥1; S2, adopt improved wavelet transform algorithm to detect singular points in each frame image in the current scene image A and the background image sequence respectively, filter out effective singular points and calculate the singular value of each effective singular point; the improved wavelet transform algorithm comprises wavelet base adaptive selection, two-dimensional wavelet transform and high-frequency detail component extraction, adaptive threshold setting and singular point screening, and singular value calculation. S3, based on the current scene image A and the background image sequence The singular point detection result, the singularity value difference and the historical background image data corresponding to the pixel position in the current scene image A and the background image sequence are distinguished by a multi-layer judgment rule to obtain a bad pixel detection result containing the bad pixel type, position and corresponding temperature parameter. S4, periodically evaluate the accuracy of the bad pixel detection result, and dynamically calibrate the detection parameters according to the evaluation result, and feed back the bad pixel detection result to the detector control unit.

2. The method of claim 1, wherein the step of detecting the bad pixel is performed by using a method of detecting a bad pixel of an infrared detector, the method comprising the steps of: The specific process of S1 is: ​ S1.1, according to the preset frame rate, the working scene image of the detector is collected, after pixel gray value calibration and image size normalization preprocessing, the current scene image A is formed and stored; S1.2, triggering shutter closing by linkage control mechanism when the working temperature of the detector changes to reach preset temperature step, collecting a single frame of background image and pre-processing, storing to the background image database after associating temperature parameters to form the background image sequence The background image database adopts a cyclic covering mechanism to retain the latest N frames of background images, and N is a preset positive integer.

3. The method of claim 2, wherein the step of detecting the bad pixel is performed by using a method of detecting a bad pixel of an infrared detector, comprising the steps of: The preset frame rate is 10-30 frames / second; and / or, ​ The preset temperature step is ; and / or, The collection accuracy of the temperature parameter is ; and / or, 。 4. The method for detecting defective pixels in an infrared detector according to claim 1, characterized in that, The improved wavelet transform algorithm specifically includes the following sub-steps: S2.1, adaptive selection of wavelet base: select biorthogonal wavelet as wavelet base, and dynamically determine 3-5 level wavelet decomposition scale according to the gray mean and variance of the image; S2.2, two-dimensional wavelet transform and detail component extraction: two-dimensional wavelet transform is performed on the image, and a low-frequency detail component LL and three high-frequency detail components LH, HL and HH are obtained by decomposition; S2.3, adaptive threshold setting and outlier screening: calculate the gray scale standard deviation of non-edge region in each high-frequency detail component As a noise intensity indicator, an adaptive threshold is set wherein k is an empirical coefficient of 2.5-3.5; using the first neighborhood window to traverse each high-frequency detail component, screening out candidate outliers whose center pixel gray scale value of the first neighborhood window is a local maximum value and exceeds the corresponding threshold T, and taking the intersection of candidate outliers in three high-frequency detail components as effective outliers, thereby obtaining an effective outlier set; S2.4, singularity value calculation: the wavelet coefficient modulus value of the position of the effective singular point is extracted under different wavelet decomposition scales, the decay index is obtained by fitting the decay curve of the modulus value with the scale, and the decay index is taken as the singularity value.

5. The method for detecting defective pixels in an infrared detector according to claim 4, characterized in that, Step S2 after obtaining the effective singular point, before step S3, further comprises: For each of the effective singular points, the effective singular point is taken as a to-be-verified point; At least two second neighborhood windows of different sizes are used to perform secondary singular point judgment at the position of the to-be-verified point; The number of times that the to-be-verified point is judged as a singular point in all second neighborhood windows is counted, and the confidence value C of the to-be-verified point is calculated according to the number of times; The confidence value C is compared with a preset confidence threshold value C0, if the to-be-verified point is determined to pass the verification, and the to-be-verified point is kept as a final valid singular point; if the to-be-verified point is determined to fail the verification, and the to-be-verified point is removed from the valid singular point set.

6. The method of claim 1 or 4 or 5, wherein, The multi-layer judgment rule is a three-layer judgment rule, which is based on the effective singular points obtained after step S2 screening and verification, and executes the following three-layer judgment rule: First layer decision: if a pixel position is a valid outlier in the current scene image A, and is not a valid outlier in the latest single frame background image in the sequence of background images then the pixel position is decided to be a noise point. Second layer judgment: if a pixel position is valid singular point in both the current scene image A and the latest single-frame background image B, then calculate the difference of the singular value of the pixel position in the current scene image A and the corresponding singular value in the latest single-frame background image B respectively ; set the difference threshold value , if , then determine that the pixel position is a static bad point, otherwise determine that the pixel position is a noise point, wherein, is the singular value of the pixel position in the current scene image A, is the singular value of the pixel position in the single-frame background image B. The third layer of judgment: If a pixel position is not a valid singular point in the current scene image A, then in the latest single-frame background image... If the pixel in the middle is a valid singularity, then retrieve the pixel position in the historical background image. to The detection results are as follows: if the pixel location was a valid singular point in any frame of the historical background image, then the pixel location is determined to be a dynamic bad point; otherwise, the pixel location is determined to be noise.

7. The method of claim 6, wherein the step of detecting the bad pixel is performed by using a method of detecting a bad pixel of an infrared detector, comprising the steps of: the difference threshold value is in the range of 0.1-0.

3. ​ 8. The method for detecting defective pixels in an infrared detector according to claim 1, characterized in that, The specific process of S4 is: S4.1, parameter adaptive calibration: S4.1.1, periodically count the accuracy of the bad pixel detection result, and compare the accuracy with a preset accuracy threshold; S4.1.2, if the accuracy is lower than the preset accuracy threshold, then according to the deviation direction and degree of the accuracy, the wavelet decomposition scale, the singular point detection threshold coefficient k, the bad point judgment difference threshold , the background image acquisition temperature step at least one parameter is adjusted. S4.1.3, using the newly collected verification image of preset number M frames after parameter adjustment, re-executing steps S2 to S3, calculating a new accuracy to verify the parameter adjustment effect, until the accuracy reaches or exceeds the preset accuracy threshold, M is a positive integer; S4.2, adaptive compensation of bad pixels: According to the type, position information and corresponding temperature parameter in the bad pixel detection result, the corresponding compensation strategy is called to perform pixel compensation.

9. The method for detecting defective pixels in an infrared detector according to claim 8, characterized in that, In S4.2, the corresponding compensation strategy is called to perform pixel compensation, specifically including: S4.2.1, compensation strategy initialization: Prestore a differentiated compensation strategy library associated with bad pixel types and image regions; the strategy library includes: a regular neighborhood interpolation strategy for non-edge area static bad pixels, a weighted neighborhood interpolation strategy for edge area static bad pixels, and a temperature prediction based pre-compensation strategy for dynamic bad pixels; S4.2.2, static bad pixel compensation execution: For the pixel of the type of static bad pixel, the image region where the static bad pixel is located is determined; if located in a non-edge region, the regular neighborhood interpolation strategy is called to take the gray scale values of all normal pixels in the preset compensation neighborhood window of the static bad pixel as the center of the static bad pixel, and interpolation calculation is performed to obtain a compensation value; if located in an edge region, the weighted neighborhood interpolation strategy is called to take the gray scale values of pixels in the preset compensation neighborhood window of the static bad pixel as the center of the static bad pixel, and interpolation calculation is performed in combination with a preset weight coefficient to obtain a compensation value, wherein pixels belonging to an edge direction are given a higher weight; S4.2.3, dynamic bad pixel compensation execution: For the pixel of the type of dynamic bad point, the pre-compensation strategy is called, a temperature-bad point mapping relationship based on historical data is queried according to a temperature parameter associated with the dynamic bad point; when an absolute value of a difference between a real-time working temperature of the detector and a temperature threshold value, at which the dynamic bad point is prone to appear, indicated in the mapping relationship is less than or equal to a preset early warning temperature difference , a compensation parameter preset for the temperature threshold value is automatically called to pre-compensate the pixel of the dynamic bad point.

10. The method of claim 9, wherein the step of detecting the bad pixel is performed by using a method of detecting a bad pixel of an infrared detector, comprising the steps of: The regular neighborhood interpolation strategy is: ​ Taking the static bad pixel to be compensated as the center, the gray scale values of all normal pixels in the preset compensation neighborhood window of the static bad pixel are taken, the arithmetic mean of the gray scale values is calculated, and the arithmetic mean is taken as a compensation value to replace the original gray scale value of the bad pixel; The weighted neighborhood interpolation strategy is: Taking the static bad pixel to be compensated as the center, the pixels in the preset compensation neighborhood window of the static bad pixel are taken; After a weight coefficient is given to the gray scale values of the pixels in the compensation neighborhood window, a weighted average value is calculated as a compensation value; wherein the weight coefficient assignment rule is that the adjacent pixels in the edge tangent direction of the static bad pixel are given a weight coefficient higher than that of the pixels in other directions; The pre-compensation strategy is: According to the temperature threshold value where the dynamic bad pixel is prone to appear obtained by querying, compensation parameters verified effective at the temperature threshold value are called from historical compensation records; the compensation parameters at least include compensation neighborhood window size, weight coefficient distribution or direct optimal compensation value; when the pre-compensation trigger condition is met, the compensation parameters are directly applied to replace the gray scale value of the pixel of the dynamic bad pixel.

Images