Image processing method and device, program product and storage medium

Abstract

This invention provides an image processing method and apparatus. Addressing the characteristics of accelerator neutron source intensity fluctuations and the transient and random nature of neutron radiograph white spot noise, a detection-location-replacement mechanism is employed to process multi-frame image sequences, achieving image denoising. First, for each pixel location, a noise reference value obtained based on temporal median filtering is used to determine whether each frame at that pixel location contains noise. The average gray value of the noisy frame at that pixel location is replaced with the average gray value of the noise-free frame. The replaced frame sequence is then fused and iterated through all pixel locations to obtain a first-stage denoised fused image. Next, a spatial filtering method is used to determine whether each fused pixel contains residual noise. For pixels containing residual noise, another noise reference value (not based on temporal filtering) is used to replace the gray value of the noisy frame again, and this process is repeated across all pixel locations to obtain a second-stage denoised fused image.

Description

Technical Field

[0001] This invention relates to an image processing method and apparatus capable of removing white spot noise caused by radiation particles or waves, and more particularly to a low-noise image processing method and apparatus suitable for neutron radiography, as well as corresponding program products and storage media, and further to a neutron radiography system. Background Technology

[0002] Neutron radiography is a high-performance non-destructive testing technique that uses the difference in the number of neutrons projected after a neutron interacts with a sample to create an image, revealing information such as the internal structure and defects of the irradiated object. Neutron radiography is classified into different types according to the neutron energy used, including cold neutron radiography, thermal neutron radiography, and fast neutron radiography. Among these, thermal neutron radiography is widely used in aerospace, defense, and precision manufacturing industries due to its excellent image quality, low cost, and high sensitivity to specific nuclides.

[0003] In a commonly used neutron radiography system, neutrons from a neutron source penetrate the sample and interact with a conversion screen, converting the irradiated neutrons into, for example, visible light. A neutron image is then captured using a camera or similar device.

[0004] However, neutron images are always subject to various noises that degrade the image. Among these noises, white spot noise is the most prominent. It is mainly generated by the interaction of gamma rays or charged particles with the imaging device of the camera, such as the CCD chip. Although the camera itself is not directly irradiated by the neutron beam, many radiating particles (such as scattered neutrons and secondary gamma) may still reach the CCD chip and deposit energy, producing prominent random white spot noise in the image. This leads to a decrease in image resolution and loss of image details, and may affect quantitative analysis and 3D reconstruction. Therefore, image denoising methods are commonly used in the industry to eliminate image noise.

[0005] However, white spot noise in neutron images is often large in size, high in magnitude, and unevenly distributed, which may make commonly used denoising algorithms (image domain and frequency domain denoising methods) unsuitable for real neutron images containing a large amount of white spot noise.

[0006] Therefore, how to ensure imaging resolution while avoiding the impact of radiation-induced white spot noise on the imaging quality of neutron images has become an urgent problem to be solved. Summary of the Invention

[0007] This application aims to solve the aforementioned technical problem, namely, the degradation of neutron image imaging quality caused by radiation-induced white spot noise.

[0008] To address the problems existing in the prior art, this invention provides an image processing method and apparatus that can effectively remove white spot noise caused by radiation during neutron radiography. Thus, white spot noise in neutron images is eliminated while maintaining imaging resolution, thereby improving the resolution of the imaging system.

[0009] Furthermore, the present invention is not limited to neutron photography, but can be equally applied to removing white spot noise caused by radiation particles or waves in other situations, such as noise originating from cosmic rays in remote sensing images.

[0010] Specifically, this invention provides an image processing method applied to a neutron radiography system, characterized by comprising: an image acquisition step, wherein, for a single radiography task, multiple frames of probe images with the same imaging duration are acquired through multiple exposures to obtain a probe image frame sequence; and a denoising step, wherein the probe image frame sequence is subjected to temporal denoising processing and fusion processing at each pixel position, wherein, in the temporal denoising processing, the mean or median of the pixel grayscale at the processed pixel position of each frame in the probe image frame sequence is used as the grayscale reference value of the pixel frame sequence, and the pixel grayscale at the processed pixel position of each frame is successively subtracted from the grayscale of the pixel frame sequence. A reference value is used. If the result is not greater than a first threshold, the gray level of the pixel in that frame is retained. If the result is greater than the first threshold, the original gray level value is replaced by the average of all retained pixel gray levels at the processed pixel position in each frame. This yields a temporal denoised pixel gray level sequence at the processed pixel position as a first-stage denoised pixel gray level sequence. In the fusion process, the cumulative value of all pixel gray levels in the first-stage denoised pixel gray level sequence at the processed pixel position is calculated as the gray level value of that pixel position in the first-stage denoised fused image. Thus, a first-stage denoised fused image is obtained by traversing all pixel positions; and a second denoising step, wherein the first-stage denoised fused image is... The image is processed for residual noise at each pixel location, followed by noise reduction and restoration. In the residual noise detection process, the group of pixels closest to the processed pixel location is filtered using mean or Gaussian filtering to obtain a spatial pixel grayscale reference value. The absolute value of the difference between the pixel grayscale of the processed pixel location in the first-stage denoising fusion image and the spatial pixel grayscale reference value is calculated. The presence of residual noise at the processed pixel location is determined by comparing this absolute value with a second threshold. In the noise reduction and restoration process, if the processed pixel location is determined to have no residual noise, the first-stage denoising fusion image is retained. The pixel grayscale of the pixel position in the denoised fused image is used as the pixel grayscale of the corresponding pixel position in the secondary denoised fused image. If the processed pixel position is determined to have residual noise, the ratio of the spatial pixel grayscale reference value to the number of image frames is used as the new pixel frame sequence grayscale reference value. The processed pixel position is then subjected to the temporal denoising process again to obtain a new temporal denoised pixel grayscale sequence as the secondary denoised pixel grayscale sequence. The accumulated value of the obtained pixel grayscale sequence is used as the corresponding pixel grayscale to replace the grayscale value of the corresponding pixel position in the primary denoised fused image. Thus, the secondary denoised fused image is obtained by traversing all pixel positions.

[0011] Furthermore, this invention provides an image processing apparatus applied to a neutron radiography system, characterized by comprising: an image acquisition module, which, for a single radiography task, acquires multiple frames of detection images with the same imaging duration through multiple exposures to obtain a detection image frame sequence; and a primary denoising module, wherein the detection image frame sequence is subjected to temporal denoising and fusion processing at each pixel position, wherein in the temporal denoising processing, the mean or median of the pixel grayscale at the processed pixel position of each frame in the detection image frame sequence is used as the grayscale reference value of the pixel frame sequence, and the pixel grayscale at the processed pixel position of each frame is successively subtracted from the grayscale reference value of the pixel frame sequence. If the result is not greater than the first threshold, the grayscale value of the pixel in that frame is retained. If the result is greater than the first threshold, the original grayscale value is replaced by the average of the grayscale values ​​of all retained pixels at the processed pixel location in each frame. This yields a temporal denoised pixel grayscale sequence at the processed pixel location as a first-stage denoised pixel grayscale sequence. In the fusion process, the cumulative value of the grayscale values ​​of all pixels in the first-stage denoised pixel grayscale sequence at the processed pixel location is calculated as the grayscale value of that pixel location in the first-stage denoised fused image. Thus, a first-stage denoised fused image is obtained by traversing all pixel locations; and a second-stage denoising module, wherein the first-stage denoised fused image... The process involves residual noise detection and denoising repair at each pixel location. In the residual noise detection process, the spatial pixel grayscale reference value is calculated by means filtering or Gaussian filtering for the pixel group closest to the pixel location being denoised. The absolute value of the difference between the pixel grayscale value at the denoised pixel location and the spatial pixel grayscale reference value is calculated. The presence of residual noise at the denoised pixel location is determined by comparing this absolute value with a second threshold. In the denoising repair process, if the denoised pixel location is determined to have no residual noise, the original denoising repair is retained. The pixel grayscale of the pixel position in the noise-fused image is used as the pixel grayscale of the corresponding pixel position in the secondary noise-fused image. If the processed pixel position is determined to have residual noise, the ratio of the spatial pixel grayscale reference value to the number of image frames is used as the new pixel frame sequence grayscale reference value. The processed pixel position is then subjected to the temporal denoising process again to obtain a new temporal denoised pixel grayscale sequence as the secondary noise-fused pixel grayscale sequence. The accumulated value of the obtained pixel grayscale sequence is used as the corresponding pixel grayscale to replace the grayscale value of the corresponding pixel position in the primary noise-fused image. Thus, the secondary noise-fused image is obtained by traversing all pixel positions.

[0012] In addition, the present invention provides a program product comprising a program that, when executed by a processor, implements the steps of the above-described image processing method.

[0013] In addition, the present invention provides a storage medium on which a program is stored, which, when executed by a processor, implements the steps of the above-described image processing method.

[0014] In addition, the present invention provides a neutron radiography system, comprising: a neutron source that provides a collimated neutron beam; a conversion screen disposed in front of the neutron emission direction of the neutron source, with respect to a sample, for converting the irradiated neutrons into visible light; an imaging detector that detects the visible light converted by the conversion screen in a multiple short-exposure mode; and the aforementioned image processing device.

[0015] Using the image processing method and apparatus of the present invention as described above, temporal denoising and fusion processing are performed on the detection image frame sequence, targeting the pixel grayscale sequence at each pixel position (i.e., the sequence of grayscale values ​​at the corresponding pixel positions in each frame). A first-stage denoised fused image (equivalent to first-stage denoising) is obtained by traversing all pixel positions. Then, for the first-stage denoised fused image, pixels without residual noise retain their pixel grayscale values, while pixels still containing residual noise have their pixel grayscale sequence at their pixel positions modified with reference values, and temporal denoising and fusion processing is performed again. A second-stage denoised fused image (equivalent to second-stage denoising) is obtained by traversing all pixel positions. Therefore, white spot noise caused by radiation during neutron radiography can be effectively removed, thereby eliminating white spot noise in neutron images while maintaining imaging resolution and improving the resolution of the imaging system.

[0016] Furthermore, the secondary denoising process can be repeatedly performed iteratively until no residual noise remains or the number of iterations reaches a preset limit. This further eliminates noise. Attached Figure Description

[0017] Figure 1 This is a schematic diagram showing the structure of the neutron radiography system 10 of the present invention.

[0018] Figure 2 This is a flowchart illustrating the image processing method of this embodiment.

[0019] Figure 3 Is with Figure 2 The illustrated processing procedure corresponding to the image processing method shown is explained.

[0020] Figure 4 This describes the specific process of the normalization step S202.

[0021] Figure 5 This describes the specific process of one noise reduction step S203.

[0022] Figure 6 This describes the specific process of the secondary denoising step S204.

[0023] Figure 7 This describes the specific process of time-domain repair step S204b.

[0024] Figure 8 This is a diagram illustrating the effect of the image processing method of this embodiment.

[0025] Figure 9 This is a functional block diagram illustrating the image processing apparatus 900 of this embodiment.

[0026] Figure 10 This refers to the hardware structure of an image processing device that utilizes a processor to execute computer programs. Detailed Implementation

[0027] The specific embodiments of the present invention will now be described with reference to the accompanying drawings.

[0028] It should be understood that all the accompanying drawings are merely representative schematic diagrams, used to aid in understanding the technical ideas of the invention, and do not limit the specific shapes, sizes, or positional relationships.

[0029] In the following embodiments, when referring to numbers (including number, value, quantity, range, etc.) of elements, they are not limited to that specific number, except in cases where they are specifically stated or where they are clearly limited to a specific number in principle, and may be above or below that specific number.

[0030] Furthermore, in the following embodiments, the structural elements (including step elements, etc.) are not necessarily required except where specifically stated or where they are obviously understood to be necessary in principle, and may also include elements not explicitly mentioned in the specification.

[0031] The embodiments described in this specification are merely examples of a complete description and do not limit the scope of protection of this invention. All other embodiments that can be obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0032] The image processing method of this invention can be implemented by a computer executing a program stored in a storage medium, or the various steps of the method can be implemented by dedicated hardware. The storage medium can be any readable medium capable of storing data and programs, such as a hard disk, optical disk, flash drive, or chip. The program can be any form of computer program, such as binary code, source code, or instruction set. This program can be executed by a processor on devices such as computers, servers, and cloud computing platforms to implement the image processing method of this invention.

[0033] [Implementation Method]

[0034] The following embodiments are specifically described for neutron radiography, but the present invention is not limited to neutron radiography and can be applied to other situations involving the removal of white spot noise caused by radiating particles or waves. However, considering the characteristics of the source intensity fluctuations of accelerator neutron sources, conventional time-domain median filtering cannot effectively remove noise. Therefore, the present invention is particularly suitable for neutron radiography using accelerator neutron sources.

[0035] First, let me briefly explain the structure of the neutron photography system. Figure 1 This is a schematic diagram showing the structure of the neutron radiography system 10 of the present invention.

[0036] like Figure 1 As shown, the neutron imaging system 10 includes: a neutron source 1, a sample 2, a neutron conversion screen 3, a reflector 4, an imaging detector 5, a data acquisition and processing unit 6, and a dark box 7. The neutron conversion screen 3, the reflector 4, and the imaging detector 5 are placed inside the dark box 7, while the neutron source 1, the sample 2, and the data acquisition and processing unit 6 are placed outside the dark box 7.

[0037] Neutron source 1 is, for example, a radioactive isotope neutron source, a reactor neutron source, or an accelerator neutron source. The present invention does not limit this, but an accelerator neutron source is preferred, in which case the noise reduction effect of the present invention is particularly significant.

[0038] Neutron source 1 emits a collimated neutron beam 8. Sample 2 and neutron conversion screen 3 are positioned in front of the emission direction of neutron beam 8 relative to neutron source 1. Sample 2 and neutron conversion screen 3 are configured perpendicular to the direction of neutron beam 8. Neutron conversion screen 3 contains neutron-converting material, fluorescent material, etc. The neutron-converting material interacts with the neutrons penetrating sample 2, producing α, β, γ rays or recoil material, etc. These charged particles or secondary rays cause the fluorescent material to emit light.

[0039] The neutron conversion screen 3 is installed inside the dark box 7, which has an inverted "L"-shaped cross-section, at the front. The light emitted by the fluorescent material travels along the emission direction of the neutron beam 8. The reflector 4 is placed at a certain angle at the corner of the dark box 7 to ensure that the light emitted by the neutron conversion screen 3 can be received by the imaging detector 5. The imaging detector 5 is a device that can convert the received light into image data for output, including but not limited to CCD cameras, CMOS cameras, etc.

[0040] Imaging detector 5 transmits image data to data acquisition and processing unit 6 via a data cable, where the image data is processed. Here, imaging detector 5 can adjust the exposure time of each output image (hereinafter referred to as "detection image"). For example, the exposure mode of imaging detector 5 can be set through data acquisition and processing unit 6, and imaging can be performed using multiple short-time exposure modes (the definition of "short-time" will be described later).

[0041] Thus, compared with the neutron radiography systems commonly used in the prior art, the difference of the neutron radiography system of the present invention can be considered to be only in the data acquisition and processing unit 6. This difference can be implemented in hardware or software. For example, it can be considered that the difference only involves the hardware structure of the data acquisition and processing unit 6, or it can be considered that the difference only involves the program stored and running on the computer in the data acquisition and processing unit 6.

[0042] Therefore, during the operation of the neutron imaging system 10, the neutron source 1 irradiates the sample 2 under test with a neutron beam 8, and the imaging detector 5 (sometimes referred to as a CCD camera below) detects the light generated from the neutron conversion screen 3. Under the control of the data acquisition and processing unit 6, the CCD camera outputs a sequence of detected images at fixed time intervals.

[0043] As mentioned above, the detection images output by CCD cameras contain white spot noise caused by radiated particles, especially when using accelerator neutron sources, and existing technologies cannot effectively remove this white spot noise.

[0044] In response, the image processing method of the present invention can effectively remove the white spot noise and improve the visual effect of the neutron image. In this image processing method, denoising is applied at least twice in stages to a sequence of multiple detection images obtained from the imaging detector.

[0045] Specifically, considering the fluctuations in the source strength of the accelerator neutron source and the transient and random nature of neutron radiograph white spot noise, a detection-localization-replacement mechanism is adopted. For the sequence of frames of the denoising target image obtained from the probe image sequence, temporal denoising and fusion processing are performed on the pixel grayscale sequence at each pixel location to obtain denoised pixels at each pixel location. Specifically, for the pixel grayscale sequence at a given pixel location, a defined noise reference value for that pixel location is used to determine whether each frame in the frame sequence contains noise. The average grayscale value at that pixel location of all frames determined not to contain noise is used to replace the grayscale value at that pixel location of frames determined to contain noise. All frames in the replaced frame sequence at that pixel location are then fused into a single fused pixel. Specifically, firstly, temporal denoising and fusion processing are performed on the frame sequence at each pixel location using a first noise reference value based on adaptive temporal median filtering, thereby obtaining a denoised fused image (corresponding to one denoising step) that eliminates large-scale white spot noise. Based on this, each pixel of the first denoised fused image is identified as a pixel containing residual noise and to be repaired based on spatial domain filtering. The frame sequence at the pixel position of the pixel to be repaired is then subjected to temporal denoising and fusion processing again using the second noise reference value to perform temporal repair (corresponding to the second denoising step).

[0046] The two time-domain denoising processes target different types of noise (large-scale white spot noise and residual noise) and use different criteria (noise reference values) to determine whether noise is present, thereby effectively removing white spot noise caused by radiation during neutron radiography.

[0047] Furthermore, the secondary denoising step can be repeated iteratively until there is no residual noise in the resulting denoised fused image or the number of iterations reaches a preset upper limit, thereby further removing noise.

[0048] The details of the image processing method of this embodiment will now be described with reference to the accompanying drawings.

[0049] Figure 2 This is a flowchart illustrating the image processing method of this embodiment. Figure 3 This is a description of the corresponding exemplary processing procedure. Figures 4-7 express Figure 2 The specific flow of the corresponding steps in the image processing method shown.

[0050] like Figure 2 As shown, the image processing method of this embodiment generally includes an image acquisition step S201, a normalization step S202, a primary denoising step S203, and a secondary denoising step S204. The primary denoising step S203 and the secondary denoising step S204 mainly involve the aforementioned temporal denoising and fusion processes, only the target pixels may differ, and different noise reference values ​​may be used. Therefore, they can also be collectively referred to as "denoising step S205". As detailed later, the purpose of the normalization step S202 is to eliminate image grayscale value fluctuations caused by changes in neutron source intensity, which is particularly suitable for situations using accelerator neutron sources. However, when the neutron source intensity is stable or the source intensity fluctuation is negligible, this normalization step S202 can be omitted.

[0051] (Step S201)

[0052] First, in the image acquisition step S201, multiple consecutive detection images with the same imaging duration are acquired from the imaging detector 5 to obtain a detection image sequence.

[0053] Specifically, unlike traditional single-exposure imaging, this embodiment sets the exposure time of the imaging detector 5 and uses short-time multiple exposures to acquire multiple equally spaced detection images. Each detection image is, for example, an M×N pixel grayscale image. For example, raw data can be acquired through a CCD camera, and the CCD camera can be controlled to output process images at fixed time intervals to obtain an image sequence of grayscale changes within the same time interval. This allows for the acquisition of multiple equally spaced neutron images, thereby enabling more precise control and processing of neutron image data.

[0054] For example, compared to the traditional exposure time of approximately 3 to 5 minutes per image, in one embodiment, the CCD camera outputs a detection image every 10 seconds (exposure time of 10 seconds) while maintaining the same exposure time and aperture size, thereby obtaining a sequence of detection images with the same time interval. This image sequence is output to the data acquisition and processing unit 6 through the CCD camera's data transmission interface for subsequent processing.

[0055] It should be noted that the "short time" mentioned in this application means that the thermal noise of the neutron radiograph itself is negligible compared with the imaging signal of the detection image. For example, if the gray value of the imaging signal is 500, and the gray value fluctuation of the image caused by other noise (such as thermal noise, dark current noise, readout noise, etc.) is 5, and the fluctuation signal accounts for less than 1% of the imaging signal, then it can be considered to meet the conditions for short time imaging.

[0056] The reason for employing this imaging mode in this invention is that, under long exposure conditions, the amount of white spot noise increases with the imaging time. Assuming that the noise percentage in a one-minute image is 10%, and the white spot noise percentage in a five-minute image is 30%, and considering five imaging attempts, the probability that the median value of the pixel sequence in the one-minute image is noise is 0.9%, and the probability that the median value of the pixel sequence in the five-minute image is noise is 16.3%. Such a long exposure method is not particularly suitable for the image processing described later in this invention.

[0057] The image acquisition step S201 described above corresponds to Figure 3 In the process “①”, Figure 3 As an example, five detection images were acquired, each with roughly the same exposure time. As shown in the figure, each detection image contains a large amount of white spot noise.

[0058] (Step S202)

[0059] In the normalization step S202, a weighting factor based on the source strength of the neutron source is calculated for each detection image obtained in the image acquisition step S201, and the weighting factor is used to normalize each detection image.

[0060] Here, the weighting factor can be calculated according to the specific requirements and methods of the application. For example, it can be calculated based on factors such as the grayscale distribution and contrast of each detected image, or it can be calculated by directly measuring the source strength or based on the accelerator current, as long as the influence of source strength fluctuations between each detected image can be eliminated. This invention does not limit this.

[0061] Furthermore, since this invention can also be applied to fields other than neutron photography, the "source strength of the neutron source" here can also be appropriately expressed as "source strength of the detection signal source", etc.

[0062] Figure 4 The specific process of the normalization step S202 is described, which includes, for example, the following steps: weight factor calculation step S202a and normalization processing step S202b.

[0063] (Step S202a)

[0064] In the weight factor calculation step S202a, each probe image obtained in the image acquisition step S201 is first pre-smoothed, and then the average gray value of each pre-smoothed probe image is calculated as the source strength weight factor.

[0065] The pre-smoothing process can employ any smoothing method commonly used in the art, or it can be omitted without compromising the technical effect of the present invention. In one example, this smoothing process can be performed using a 5×5 spatial median filter, but the size of the sliding window is not limited and will not be described in detail here.

[0066] Let the detection image sequence consist of m images, where the weight factor of the nth image is... It can be calculated using the following formula (1).

[0067]

[0068] Where M and N are the image dimensions, i.e., the horizontal and vertical pixel counts of each detected image. This represents the nth detected image. This represents the grayscale value of pixel (i,j) in the nth probe image after the above pre-smoothing process (spatial median filtering).

[0069] (Step S202b)

[0070] Next, in the normalization process step S202b, the weight factors obtained in step S202a are used to normalize each detection image (the original image obtained in step S201).

[0071] The normalization process is calculated using the following formula (2).

[0072]

[0073] in, It is the actual gray value of the pixel at pixel position (i,j) in the nth probe image. It is the normalized gray value of the pixel, 1≤i≤M, 1≤j≤N.

[0074] In one embodiment, the CCD camera captures an image every 10 seconds, obtaining, for example, a sequence of 10 equally spaced neutron images with an exposure time of 10 seconds. Each neutron image is pre-smoothed using a 5×5 spatial median filter, and the average gray value of each filtered image is calculated. For example, assuming the average gray value of the first image after median filtering is 50 according to equation (1), the weighting factor based on the source strength is... The value is 50. Then, each probe image is normalized, and the normalized gray value of each pixel in each image is calculated using equation (2). For example, for pixel (i,j) in the first image, if its actual gray value is 100, then its normalized gray value is 100 / 50=2.

[0075] Therefore, through the normalization step S202, the multiple detection images obtained in the image acquisition step S201 are each normalized to obtain a sequence (normalized frame sequence) composed of multiple normalized images.

[0076] The normalization step S202 described above corresponds to Figure 3 The process “②” in the diagram illustrates the process of calculating weight factors after the probe image has undergone median filtering (pre-smoothing) and then undergoing further normalization.

[0077] (Step S203)

[0078] In the first denoising step S203, a detection-localization-replacement mechanism is adopted to perform temporal denoising and fusion processing on the normalized frame sequence output by the normalization step S202 at each pixel position to eliminate white spot noise, and obtain a first denoising fusion image described later by traversing all pixel positions.

[0079] It should be noted that this embodiment is based on the normalization step S202. Therefore, in the first denoising step S203, the normalized image is used as the denoising object, and normalization restoration processing is required in the fusion process. However, if the normalization step S202 is omitted, the multiple probe images obtained in the image acquisition step S201 can be directly used as the denoising object for the first denoising, in which case the restoration processing is also omitted. That is, in this invention, the processing object of the first denoising step S203 can be either the probe image itself or the image obtained after the probe image has undergone the above-mentioned normalization step and / or other image processing commonly used in the art.

[0080] Additionally, before the denoising step S203, pixel alignment can be performed on each image as needed. This alignment process can be performed on each normalized image after the normalization step S202, or on each acquired image during the image acquisition step S201. The aligned image is then used as the aforementioned probe image. Of course, the alignment step can also be omitted depending on the situation.

[0081] Figure 5 The specific process of a denoising step S203 is described, for example, including the following steps: first noise reference value calculation step S203a, first pixel noisy frame judgment step S203b, first pixel noisy frame replacement step S203c, and first fusion step S203d. In a denoising step S203, the above-mentioned denoising fusion process is performed using the first noise reference value at each pixel position of the M×N pixels of the normalized image.

[0082] (Step S203a)

[0083] In the first noise reference value calculation step S203a, for each pixel position of the M×N pixels of the normalized image, the median or mean of the gray value at that position of each frame in the normalized frame sequence is calculated, and the median or mean is used as the first noise reference value at that pixel position, that is, the gray value of the pixel frame sequence.

[0084] For example, the first noise reference value at pixel position (i,j) can be expressed as: .

[0085] (Step S203b)

[0086] In the first pixel noisy frame determination step S203b, for each pixel position, based on the relationship between the normalized grayscale value of each frame in the normalized frame sequence at that pixel position and the first noise reference value calculated in the first noise reference value calculation step S203a, it is determined whether the frame contains white spot noise at that pixel position. If there is no noise, its pixel grayscale is retained; if it contains noise, its grayscale value is replaced in the following steps. For example, it can be determined based on whether the difference between the normalized grayscale value and the first noise reference value at the corresponding pixel position is greater than a predetermined threshold (first threshold).

[0087] For example, the following formula (3) can be used to make a judgment.

[0088]

[0089] in, It is the first threshold, the size of which is related to the degree of signal fluctuation of the background noise of the CCD camera, and can be appropriately set by those skilled in the art according to the actual situation. This is a noisy Boolean value. A value of 1 indicates that the frame contains noise at pixel position (i,j), and a value of 0 indicates that there is no noise.

[0090] (Step S203c)

[0091] In the first pixel noisy frame replacement step S203c, according to each pixel position, the average value of the gray values ​​of all retained pixels at that pixel position is used to replace the gray values ​​of all noisy pixels at that pixel position.

[0092] In one embodiment, for example, a sequence of 10 equally spaced neutron images, each 320×240 pixels in size, is acquired and normalized, with each image having an 8-bit grayscale value. For each pixel location, the median of the corresponding normalized grayscale value in each image (each frame) is calculated as the corresponding first noise reference value. Then, for each pixel location, it is compared whether the corresponding normalized grayscale value in each frame is higher than the first noise reference value (for simplicity, it is assumed...). If the value is higher, then z(i,j) of the corresponding frame for that pixel position is set to 1; otherwise, it is set to 0. Then, temporal adaptive median filtering is used to eliminate white spot noise. Specifically, for each pixel position (i,j), the value of z(i,j) in each frame is detected. If z(i,j) = 0, it is considered that there is no white spot noise and no replacement is needed; if z(i,j) = 1, it is considered that there is white spot noise and replacement is needed. In the case of noise, the average of the gray values ​​of all retained pixels at that pixel position is used to replace the pixel gray value of the noisy frame. For example, if frames 2, 4, and 7 at (i,j) contain white spot noise, then the average of the normalized gray values ​​of frames 1, 3, 5, 6, 8, 9, and 10 at (i,j) can be used to replace the normalized gray values ​​of frames 2, 4, and 7.

[0093] (Step S203d)

[0094] In the first fusion step S204, the corresponding pixel grayscale sequences after replacement in the first pixel noisy frame replacement step S203c are fused (accumulated value is calculated) for each pixel position. For example, weighted fusion is performed using the source strength-based weighting factor calculated in the normalization step S202 (equivalent to accumulation after restoration processing). Specifically, for each pixel position, the grayscale value of the nth frame after the first pixel noisy frame replacement step S203c is multiplied by the corresponding weighting factor. Then, the images are superimposed at n=1 to m, and a denoised fused image is obtained by traversing all pixel positions. Of course, if the probe image is not normalized, no restoration processing is needed during fusion.

[0095] As described in the normalization step S202, the weighting factor can be calculated according to the specific requirements and methods of the application. For example, it can be calculated based on factors such as the grayscale distribution and contrast of each detected image, or it can be calculated by directly measuring the source strength or based on external detectors, accelerator currents, etc., as long as the influence caused by the source strength fluctuation between each detected image can be eliminated. This invention does not limit this.

[0096] Furthermore, the superposition here can be a simple weighted average or other complex image fusion algorithms, such as fusion algorithms based on wavelet transform, optical flow estimation, etc., and this invention does not limit it in this way.

[0097] The above-mentioned noise reduction step S203 corresponds to Figure 3 Processes "③" and "④" in the text perform temporal denoising and fusion processing for each pixel location using the first noise reference value at that location. For example... Figure 3 As shown, white spot noise exists in each normalized image. If a pixel location in a frame is included in the noise, then that frame contains noise at that pixel location and is marked with "×". Otherwise, it is marked with "·". For each pixel location, the frames passing through the dashed line at that location are compared with the gray median as described above. The sums are categorized into noisy and noiseless. The mean grayscale values ​​of the noiseless pixels (those retained) are calculated, and this mean value is used to replace the grayscale values ​​of each noisy pixel. Then, all frames at the same pixel location after replacement are fused by restoring the grayscale values ​​and overlaying them. By traversing all pixel locations, the first denoising and fused image described above is obtained.

[0098] (Step S204)

[0099] Even after the first denoising step, noise may still remain in the image. Therefore, this implementation method performs a second denoising step on top of this.

[0100] In the secondary denoising step S204, the primary denoising fusion image output from the primary denoising step S203 is subjected to a second denoising step to obtain a secondary denoising fusion image.

[0101] As detailed below, during the second denoising process, each pixel in the first denoised and fused image is identified as a pixel to be repaired that contains residual noise that was not completely removed during the first denoising. The normalized frame sequence at the pixel position of each identified pixel to be repaired is used as the object. The second noise reference value, which is not based on temporal filtering, is used to replace the first noise reference value for temporal denoising and fusion processing again, thereby repairing the pixel to be repaired.

[0102] Figure 6 The specific process of the secondary denoising step S204 is described, for example, including the residual noise identification step S204a and the time domain repair step S204b described below.

[0103] (Step S204a)

[0104] In the residual noise identification step S204a, for the first denoised fused image output from the first fusion step S203d, in order to determine whether it contains residual noise that was not completely removed in the first denoising, a residual noise identification value (i.e., spatial pixel grayscale reference value) is calculated for each pixel location using a spatial adaptive mean filter of a specified size. Based on the relationship between the grayscale value of each pixel and the corresponding residual noise identification value, it is determined whether the pixel contains residual noise, i.e., the pixel to be repaired. As mentioned above, Gaussian filtering can also be used for residual noise identification, but this is omitted here.

[0105] For example, for each pixel location, noise discrimination can be performed using the mean of the spatial filtering of its surrounding pixels (within the filter size) as the residual noise discrimination value, which can be done using the following formula (4).

[0106]

[0107] in, Let be the gray value of the pixel at position (i,j) in the first denoised and fused image. Furthermore, a and b are not both 0, and 2f+1 is the filter size. =1 indicates that the pixel is a pixel that needs to be repaired. =0 indicates that the pixel does not need to be repaired. The second threshold, as mentioned above It's also a pre-set value. The reason for setting this... This is because, although a mean result is obtained through spatial mean filtering, in some cases (possibly due to fluctuations or specific imaging conditions), the actual pixel data obtained is slightly larger than this mean result. Therefore, it is possible to use... Correct this situation. Similarly, it can be appropriately set by those skilled in the art based on the actual situation.

[0108] (Step S204b)

[0109] In the temporal repair step S204b, for the pixels to be repaired identified by the residual noise identification step S204a in the first denoised fused image, the normalized frame sequence at each pixel position is re-processed with temporal denoising and fusion according to their pixel positions to replace the pixels to be repaired. For pixel positions other than the pixels to be repaired, the pixel grayscale of that pixel position in the first denoised fused image is retained, thus obtaining the second denoised fused image.

[0110] In the temporal denoising process described here, the noise reference value used is a second noise reference value (i.e., a new pixel frame sequence grayscale reference value) that is different from the first noise reference value. The second noise reference value at each pixel location is obtained based on the aforementioned residual noise discrimination value at that pixel location. More specifically, the second noise reference value is obtained by assigning the residual noise discrimination value to each normalized image.

[0111] Figure 7 The specific process of the temporal restoration step S204b is described, for example, including the following steps: second noise reference value calculation step S204b1, second pixel noisy frame judgment step S204b2, second pixel noisy frame replacement step S204b3, and second fusion step S204b4. In the temporal restoration step S204b, the second noise reference value is used to perform the above-mentioned temporal denoising and fusion processing on the pixel position of the pixel to be restored in the M×N pixels of the first denoised fused image.

[0112] These steps are respectively related to Figure 5 The first noise reference value calculation step S203a, the first pixel noisy frame judgment step S203b, the first pixel noisy frame replacement step S203c, and the first fusion step S203d shown correspond to each other, so only a brief explanation is given here.

[0113] (Step S204b1)

[0114] In the second noise reference value calculation step S204b1, a second noise reference value is calculated for each pixel position of the pixel to be repaired.

[0115] The second noise reference value differs from the first noise reference value. It is not based on time-domain filtering but on spatial-domain mean filtering or Gaussian filtering. Specifically, it is based on the residual noise discrimination value for that pixel location mentioned above.

[0116] As described above, the residual noise discrimination value is used to determine residual noise in each pixel of the first denoised fused image, while the second noise reference value is used to re-evaluate whether each frame of the normalized image (multiple images before fusion) contains noise at the positions of pixels identified as having residual noise. Therefore, the residual noise discrimination value can be allocated to a single normalized image based on the weight factors of each normalized image to obtain the second noise reference value. For example, it can be obtained by dividing the residual noise discrimination value by the sum of the weight factors and then calculating the ratio to the number of image frames. Of course, if the detection image is not normalized, the ratio of the residual noise discrimination value to the number of image frames can be used directly.

[0117] (Step S204b2)

[0118] In the second pixel noisy frame judgment step S204b2, according to the pixel position of the pixel to be repaired, based on the relationship between the normalized gray value of each frame in the normalized frame sequence at that position and the second noise reference value, it is re-determined whether the frame contains white spot noise at that pixel position. If there is no noise, its pixel gray value is retained; if there is noise, its gray value is replaced in the following steps.

[0119] For example, it can be determined whether the difference between the normalized grayscale value at a certain pixel location in a frame and the second noise reference value at the corresponding pixel location is greater than a specified threshold. The third threshold is used to determine this. Similarly, the threshold can be set appropriately by those skilled in the art based on the actual situation, or the first threshold can be used directly.

[0120] (Step S204b3)

[0121] In the second pixel noisy frame replacement step S204b3, according to the pixel position of the pixel to be repaired, the average value of the gray values ​​of all retained pixels at that pixel position is used to replace the gray values ​​of all noisy pixels at that pixel position.

[0122] (Step S204b4)

[0123] In the second fusion step S204b4, the corresponding pixel grayscale sequences after replacement in the second pixel noisy frame replacement step S204b3 are fused (accumulated value is calculated) according to the pixel positions of the pixels to be repaired. For example, the weighting factor based on the source strength calculated in the normalization step S202 is used for weighted fusion (equivalent to accumulation after restoration). Specifically, the grayscale value of the nth frame is multiplied by the corresponding weighting factor according to the pixel positions of the pixels to be repaired. Then, the images are stacked in increments of n=1 to m. Of course, if the probe images are not normalized, no restoration processing is required during fusion.

[0124] Therefore, the pixels to be repaired in the first denoising fusion image are replaced, while the other pixels are retained, and the resulting image is the second denoising fusion image.

[0125] The above-mentioned secondary denoising step S204 corresponds to Figure 3 The process “⑤” in the middle, such as Figure 3 As shown in the first denoised fused image, residual single-point white spot noise exists. Therefore, for each pixel location, a second noise reference value (different from the first noise reference value, which is not based on temporal filtering) is used to perform temporal denoising and fusion processing again, and this processing iterates through the pixel locations of each pixel to be repaired. Thus, a second denoised fused image is obtained through the second denoising process.

[0126] Furthermore, the secondary denoising step can be performed iteratively. For example, for the secondary denoising fused image, the secondary denoising step S204 can be performed again, wherein all pixel positions are retraced, and the new residual noise discrimination value obtained based on the secondary denoising fused image is used to determine whether residual noise still exists at each pixel position. If it is determined that there is no residual noise, the pixel grayscale is retained. If it is determined that there is still residual noise, a new second noise reference value is obtained based on the new residual noise discrimination value, and the repair step of step S204b is performed again.

[0127] Here, the number of iterations can be preset with an upper limit, or no upper limit can be set, and it can be iterated until there is no residual noise in the output denoised and fused image.

[0128] This concludes the image processing method of this embodiment. The secondary denoising and fusion image output after at least two denoising steps is the final neutron imaging image. Of course, other image processing techniques known in the art can be further applied to this neutron imaging image as needed, which will not be detailed here.

[0129] The image processing method of this embodiment has been described above, and it has the following effects:

[0130] (1) The imaging detector adopts an imaging method based on multiple exposures in a short time, which obtains the relationship between imaging data and imaging time while ensuring the accuracy of the detection image of neutron radiography, thus providing a basis for the denoising described later.

[0131] (2) To address the instantaneous, random, and fluctuating source intensity characteristics of white spot noise in neutron radiography, especially in accelerator neutron sources, a time-domain adaptive median denoising method based on source intensity normalized image sequences was adopted. Compared with traditional time-domain methods, this method can effectively solve the gray value fluctuations in each detection image caused by changes in neutron source intensity, and has good performance under imaging conditions with unstable source intensity.

[0132] (3) Based on the imaging characteristics of neutron radiography, a secondary denoising method based on spatial domain discrimination and temporal domain data repair is adopted to address the residual noise after the first denoising. This ensures high image fidelity while removing residual noise.

[0133] It is understood that in this invention, at least two temporal denoising processes (and fusion processes) are selectively performed by performing one and two denoising processes. The difference between them is that the reference value used in the first process is obtained based on temporal median filtering, while the reference value used in the second process is obtained based on spatial mean filtering or Gaussian filtering.

[0134] This distinction is made possible by the imaging characteristics of neutron radiography. One of the core ideas of the first denoising step is to identify intra-pixel noise based on the relatively low proportion of noise. However, since the grayscale value of white spot noise is much larger than that of conventional imaging, the result calculated using mean filtering significantly exceeds the actual grayscale value, making this value unreliable. Therefore, only median filtering can be used here. Detecting residual noise is based on the imaging characteristics of neutron radiography—signal loss occurs after passing through the system. This loss can generally be simulated using Gaussian filtering or mean filtering. Therefore, these two methods can be used to detect residual noise in a pixel by comparing the grayscale values ​​of surrounding pixels.

[0135] Figure 8 This is a diagram illustrating the effect of the image processing method of this embodiment. Figure 8 In the middle (a), the image is obtained by directly synthesizing the detection images output by the imaging detector when using an accelerator neutron source for neutron radiography. Figure 8 In the image (b), the neutron image obtained by using the image processing method of this embodiment is shown.

[0136] according to Figure 8 It can be seen that the image processing method of this embodiment can effectively eliminate white spot noise in neutron images.

[0137] The following reference Figure 9 , Figure 10 The image processing apparatus of this embodiment will be described. The image processing apparatus of this embodiment can be configured as the data acquisition and processing unit 6 in the neutron radiography system described above, and can be implemented in either hardware or software.

[0138] Figure 9 This is a functional block diagram illustrating the image processing apparatus 900 of this embodiment.

[0139] like Figure 9 As shown, the image processing apparatus 900 includes an image acquisition module 901, a normalization module 902, a primary denoising module 903, and a secondary denoising module 904, corresponding to each step of the image processing method. Similarly, the primary denoising module 903 and the secondary denoising module can be collectively referred to as the denoising module 905. Furthermore, the normalization module 902 can be omitted, and the restoration process is omitted accordingly depending on whether the normalization module is omitted or not in the first and second fusion modules.

[0140] The following is a brief description of each module. Of course, the foregoing descriptions of each step can also be directly applied to the following descriptions of the features of each module.

[0141] The image acquisition module 901 is connected to the imaging detector 5 to acquire multiple continuous detection images with the same imaging duration, thus obtaining a detection image sequence.

[0142] The normalization module 902 calculates the weight factor based on the source strength of the neutron source corresponding to each detection image obtained by the image acquisition module 901, and uses the weight factor to normalize each detection image.

[0143] The normalization module 902 specifically includes a weight factor calculation module 902a and a normalization processing module 902b.

[0144] In the weight factor calculation module 902a, each probe image obtained by the image acquisition module 901 is first pre-smoothed, and then the average gray value of each pre-smoothed probe image is calculated as the source strength weight factor.

[0145] In the normalization processing module 902b, the weight factors obtained by the weight factor calculation module 902a are used to normalize each detection image.

[0146] The primary denoising module 903 performs temporal denoising and fusion processing on the normalized frame sequence output by the normalization module 902 at each pixel position to eliminate white spot noise, and obtains a primary denoising fused image by traversing all pixel positions.

[0147] The primary noise reduction module 903 specifically includes a first noise reference value calculation module 903a, a first pixel noisy frame judgment module 903b, a first pixel noisy frame replacement module 903c, and a first fusion module 903d.

[0148] In the first noise reference value calculation module 903a, for each pixel position of the M×N pixels of the normalized image, the median or mean of the gray value at that position of each frame in the normalized frame sequence is calculated, and the median or mean is used as the first noise reference value at that pixel position, that is, the gray value reference value of the pixel frame sequence.

[0149] In the first pixel noisy frame determination module 903b, for each pixel position, based on the relationship between the normalized gray value of each frame in the normalized frame sequence at the pixel position and the first noise reference value calculated by the first noise reference value calculation module 903a at the corresponding pixel position, it is determined whether the frame contains white spot noise at the pixel position. If there is no noise, its pixel gray value is retained; if there is noise, its gray value is replaced in the module described later.

[0150] In the first pixel noisy frame replacement module 903c, for each pixel position, the average value of the grayscale of all retained pixels at that pixel position is used to replace all noisy pixel grayscale values ​​at that pixel position.

[0151] In the first fusion module 903d, the grayscale sequences of the corresponding pixels replaced by the noisy frame replacement module 903c are fused (accumulated value) at each pixel position. For example, weighted fusion is performed using the source strength-based weighting factor calculated in the normalization module 902 (equivalent to accumulating after restoration), and a denoised fused image is obtained by traversing all pixel positions. Of course, if the normalization module is omitted, restoration processing is not required here.

[0152] In the secondary denoising module 904, a second denoising process is applied to the primary denoising fused image output by the primary denoising module 903 to obtain a secondary denoising fused image. Specifically, for each pixel in the primary denoising fused image, it is determined whether it is a pixel to be repaired containing residual noise that was not completely removed in the first denoising. Taking the normalized frame sequence at each pixel position of the identified pixels to be repaired as the object, a second noise reference value, which is not based on temporal filtering, is used to replace the first noise reference value for temporal denoising and fusion processing again, thereby repairing the pixels to be repaired.

[0153] The secondary noise reduction module 904 specifically includes a residual noise identification module 904a and a time-domain repair module 904b.

[0154] In the residual noise discrimination module 904a, for the first denoised fused image output by the first fusion module 903d, in order to determine whether it contains residual noise that was not completely removed in the first denoising, a residual noise discrimination value (i.e., spatial pixel grayscale reference value) is calculated for each pixel position through spatial adaptive mean filtering or Gaussian filtering of a specified size. Based on the relationship between the grayscale value of the pixel at each position and the corresponding residual noise discrimination value, it is determined whether the pixel contains residual noise, i.e., the pixel to be repaired.

[0155] In the temporal repair module 904b, for the pixels to be repaired identified by the residual noise identification module 904a in the first denoised fused image, the normalized frame sequence at each pixel position is re-processed with temporal denoising and fusion according to their pixel positions to replace the pixels to be repaired. For pixel positions other than the pixels to be repaired, the pixel grayscale of that pixel position in the first denoised fused image is retained, thus obtaining the second denoised fused image.

[0156] The temporal repair module 904b includes the following: a second noise reference value calculation module 904b1, a second pixel noisy frame judgment module 904b2, a second pixel noisy frame replacement module 904b3, and a second fusion module 904b4.

[0157] In the second noise reference value calculation module 904b1, a second noise reference value is calculated for each pixel position of the pixel to be repaired. Specifically, this can be obtained by dividing the residual noise discrimination value by the sum of the weighting factors and then calculating the ratio to the number of image frames. Of course, if the probe image is not normalized, the ratio of the residual noise discrimination value to the number of image frames can be used directly.

[0158] In the second pixel noisy frame judgment module 904b2, according to the pixel position of the pixel to be repaired, based on the relationship between the normalized gray value of each frame in the normalized frame sequence at that position and the second noise reference value, it is re-determined whether the frame contains white spot noise at that pixel position. If there is no noise, its pixel gray value is retained; if there is noise, its gray value is replaced in the following steps.

[0159] In the second pixel noisy frame replacement module 904b3, according to the pixel position of the pixel to be repaired, the average value of the gray values ​​of all retained pixels at that pixel position is used to replace the gray values ​​of all noisy pixels at that pixel position.

[0160] In the second fusion module 904b4, the corresponding pixel grayscale sequences replaced by the second pixel noisy frame replacement module 904b3 are fused according to the pixel positions of the pixels to be repaired. Of course, if the probe image is not normalized, no restoration processing is required during fusion.

[0161] Therefore, the pixels to be repaired in the first denoising fusion image are replaced, while the other pixels are retained, and the resulting image is the second denoising fusion image.

[0162] The secondary denoising fused image is input into the iterative module, which then inputs the secondary denoising fused image back into the secondary denoising module. Thus, all pixel positions are re-traversed, and the new residual noise discrimination value obtained based on the secondary denoising fused image is used to determine whether residual noise still exists at each pixel position. If no residual noise is determined, the pixel grayscale is retained. If residual noise is determined, a new second noise reference value is obtained based on the new residual noise discrimination value, and the temporal domain repair module 904b performs repair again.

[0163] Here, the number of iterations can be preset with an upper limit, or no upper limit can be set, and it can be iterated until there is no residual noise in the output denoised and fused image.

[0164] Figure 10 This refers to the hardware structure of an image processing device that utilizes a processor to execute computer programs.

[0165] like Figure 10As shown, the hardware structure of the image processing device may include at least one processor 501 and a memory such as RAM 503 and ROM 504 that are communicatively connected to the processor 501 via a bus 502. ROM 504 may include various storage media, such as hard disks, optical disks, etc. ROM 504 stores a program 505 that can be executed by the processor, and the processor executes program 505 to implement the image processing method described above. The final neutron image after image processing, as well as various thresholds, set values, etc., used in the image processing method described above, are stored in ROM 504. ROM 504 also stores the original detection image input from the imaging detector 5. Various data generated during the intermediate process of program 505 execution (i.e., during the execution of the image processing method), such as the various noise reference values, frame sequences, fused images, etc., described above, can be temporarily stored in RAM 503 or appropriately stored in ROM 504.

[0166] Input unit 506 is used to input various parameters required for the execution of program 505, such as set values, thresholds, and set values ​​for imaging detector 5, such as exposure time. Input unit 506 may include various input devices, such as a keyboard, mouse, touch screen, etc. Output unit 507 can be used to output the results of image processing, including various output devices, such as a display, speaker, etc. Communication unit 508 communicates with imaging detector 5, for example, bidirectionally, such as exchanging data via the Internet or other telecommunications networks.

[0167] Therefore, by having the CPU 501 execute the program 505 stored in the ROM 504, the data of the detected image is obtained from the imaging detector 5 via the communication unit 508, and the above-mentioned image processing method is applied to it to obtain a neutron image with white spot noise effectively eliminated.

[0168] This invention can be applied to the field of neutron photography, but is not limited thereto. For example, it can also be applied to fields such as remote sensing photography, and can effectively remove white spot noise caused by radiation particles or waves.

Claims

1. An image processing method applied to a neutron radiography system, characterized in that, include: The image acquisition step involves obtaining multiple frames of detection images with the same imaging duration through multiple exposures for a single photographic task, thus obtaining a sequence of detection image frames. In the first denoising step, the probe image frame sequence is subjected to temporal denoising and fusion processing at each pixel position. In the temporal denoising process, the mean or median of the pixel grayscale values ​​at the processed pixel position of each frame in the probe image frame sequence is used as the grayscale reference value of the pixel frame sequence. The pixel grayscale value at the processed pixel position of each frame is subtracted from the grayscale reference value of the pixel frame sequence in turn. If the result is not greater than a first threshold, the pixel grayscale value of that frame is retained. If the result is greater than the first threshold, the original grayscale value is replaced by the average value of all retained pixel grayscale values ​​at the processed pixel position of each frame. Thus, the temporal denoised pixel grayscale sequence at the processed pixel position is obtained as the first denoised pixel grayscale sequence. In the fusion process, the cumulative value of all pixel grayscale values ​​of the first denoised pixel grayscale sequence at the processed pixel position is calculated as the grayscale value of that pixel position in the first denoised fused image. Thus, the first denoised fused image is obtained by traversing all pixel positions. and The second denoising step involves performing residual noise detection and denoising repair processing on each pixel position of the first denoised fused image. In the residual noise detection processing, the spatial pixel grayscale reference value is calculated by means filtering or Gaussian filtering for the pixel group closest to the processed pixel position. The absolute value of the difference between the pixel grayscale of the processed pixel position in the first denoised fused image and the spatial pixel grayscale reference value is calculated. The presence of residual noise at the processed pixel position is determined by comparing this absolute value with a second threshold. In the denoising repair processing, if the processed pixel position is determined to have no residual noise... The pixel grayscale value of the corresponding pixel position in the first denoised fused image is retained as the pixel grayscale value of the corresponding pixel position in the second denoised fused image. If the processed pixel position is determined to have residual noise, the ratio of the spatial pixel grayscale reference value to the number of image frames is used as the new pixel frame sequence grayscale reference value. The processed pixel position is then subjected to the temporal denoising process again to obtain a new temporal denoised pixel grayscale sequence as the second denoised pixel grayscale sequence. The accumulated value of the obtained pixel grayscale sequence is used as the corresponding pixel grayscale value to replace the grayscale value of the corresponding pixel position in the first denoised fused image. Thus, the second denoised fused image is obtained by traversing all pixel positions. The image processing method further includes: The normalization step involves, after acquiring the probe image frame sequence, calculating the source strength weighting factor based on the probe signal source corresponding to each probe image frame, either based on the probe image frame or the source strength signal magnitude of the external detector within the imaging time of each probe image frame. This source strength weighting factor is then used to normalize the probe image frame sequence to obtain a normalized probe image frame sequence. Furthermore, in the first denoising step, the normalized probe image frame sequence is processed instead of the probe image frame sequence. Also, in the fusion process, when calculating the accumulated pixel grayscale value, the grayscale values ​​of all pixels are restored according to the source strength weighting factor of the corresponding frame image before being accumulated. In the denoising and repair process of the second denoising step, the spatial pixel grayscale reference value of the processed pixel position is divided by the sum of the weight factors, and then the ratio of the weight factor to the number of image frames is taken as the new pixel frame sequence grayscale reference value. Furthermore, when calculating the cumulative value of pixel grayscale, all pixel grayscale values ​​are restored according to the source strength weight factor of the corresponding frame image before being accumulated.

2. The image processing method as described in claim 1, characterized in that, Also includes: The iterative process involves traversing all pixel positions in the secondary denoising and fusion image, calculating a new spatial pixel grayscale reference value using the pixel group closest to the processed pixel position, and re-executing the residual noise detection process. If residual noise is still found during the traversal, the ratio of the new spatial pixel grayscale reference value to the image frame number is used as a new pixel frame sequence grayscale reference value for the residual noise pixel positions in the secondary denoising and fusion image. The temporal denoising process is then performed again to obtain a new secondary denoising pixel grayscale sequence. The accumulated value of the newly obtained pixel grayscale sequence is used as the corresponding pixel grayscale value to replace the pixel grayscale value of the residual noise, thereby obtaining a new secondary denoising and fusion image. Repeat the iterative steps until the obtained secondary denoising fusion image has no residual noise or the number of iterations reaches the preset upper limit, and use the final secondary denoising fusion image as the final denoising fusion image.

3. An image processing apparatus, applied to a neutron radiography system, characterized in that, include: The image acquisition module, for a single photographic task, acquires multiple frames of detection images with the same imaging duration through multiple exposures, thus obtaining a sequence of detection image frames; The first denoising module performs temporal denoising and fusion processing on the probe image frame sequence at each pixel position. In the temporal denoising process, the mean or median of the pixel grayscale values ​​at the processed pixel position of each frame in the probe image frame sequence is used as the grayscale reference value of the pixel frame sequence. The pixel grayscale values ​​at the processed pixel position of each frame are subtracted from the grayscale reference value of the pixel frame sequence in turn. If the result is not greater than a first threshold, the pixel grayscale value of that frame is retained. If the result is greater than the first threshold, the original grayscale value is replaced by the average value of all retained pixel grayscale values ​​at the processed pixel position of each frame. Thus, the temporal denoised pixel grayscale sequence at the processed pixel position is obtained as the first denoised pixel grayscale sequence. In the fusion process, the cumulative value of all pixel grayscale values ​​of the first denoised pixel grayscale sequence at the processed pixel position is calculated as the grayscale value of that pixel position in the first denoised fused image. Thus, the first denoised fused image is obtained by traversing all pixel positions. and The secondary denoising module performs residual noise detection and denoising repair processing on each pixel position of the primary denoised fused image. In the residual noise detection processing, the spatial pixel grayscale reference value is calculated by means filtering or Gaussian filtering for the pixel group closest to the processed pixel position. The absolute value of the difference between the pixel grayscale of the processed pixel position in the primary denoised fused image and the spatial pixel grayscale reference value is calculated. The presence of residual noise at the processed pixel position in the primary denoised fused image is determined by comparing this absolute value with a second threshold. In the denoising repair processing, if the processed pixel position is determined to have no residual noise... The pixel grayscale value of the corresponding pixel position in the first denoised fused image is retained as the pixel grayscale value of the corresponding pixel position in the second denoised fused image. If the processed pixel position is determined to have residual noise, the ratio of the spatial pixel grayscale reference value to the number of image frames is used as the new pixel frame sequence grayscale reference value. The processed pixel position is then subjected to the temporal denoising process again to obtain a new temporal denoised pixel grayscale sequence as the second denoised pixel grayscale sequence. The accumulated value of the obtained pixel grayscale sequence is used as the corresponding pixel grayscale value to replace the grayscale value of the corresponding pixel position in the first denoised fused image. Thus, the second denoised fused image is obtained by traversing all pixel positions. The image processing device further includes: The normalization module, for the detection image frame sequence obtained by the image acquisition module, calculates the source strength weighting factor based on the detection signal source corresponding to each detection image frame, either based on the detection image frame or based on the source strength signal magnitude of the external detector within the imaging time of each detection image frame. This source strength weighting factor is then used to normalize the detection image frame sequence to obtain a normalized detection image frame sequence. Furthermore, in the first denoising module, the normalized probe image frame sequence is processed instead of the probe image frame sequence. Also, in the fusion process, when calculating the accumulated pixel grayscale value, the grayscale values ​​of all pixels are restored according to the source strength weighting factor of the corresponding frame image before being accumulated. In the denoising and repair process of the secondary denoising module, the spatial pixel grayscale reference value of the processed pixel position is divided by the sum of the weight factors, and then the ratio of the weight factor to the number of image frames is taken as the new pixel frame sequence grayscale reference value. Furthermore, when calculating the cumulative value of pixel grayscale, all pixel grayscale values ​​are restored according to the source strength weight factor of the corresponding frame image before being accumulated.

4. The image processing apparatus as described in claim 3, characterized in that, Also includes: The iterative module traverses to obtain all pixel positions in the secondary denoising and fusion image. It calculates a new spatial pixel grayscale reference value using the pixel group closest to the processed pixel position. It then re-executes the residual noise detection process. If residual noise is still found during the traversal, the ratio of the new spatial pixel grayscale reference value to the image frame number is used as a new pixel frame sequence grayscale reference value for the residual noise pixel positions in the secondary denoising and fusion image. The temporal denoising process is then performed again to obtain a new secondary denoising pixel grayscale sequence. The accumulated value of the newly obtained pixel grayscale sequence is used as the corresponding pixel grayscale value to replace the residual noise pixel grayscale value, thereby obtaining a new secondary denoising and fusion image. The newly obtained secondary denoising fusion image is input into the iteration module until the obtained secondary denoising fusion image has no residual noise or the number of iterations reaches a preset upper limit. The final secondary denoising fusion image is then used as the final denoising fusion image.

5. A program product, comprising a program, characterized in that, When the program is executed by the processor, it implements the image processing method according to claim 1 or 2.

6. A storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the image processing method according to claim 1 or 2.

7. A neutron radiography system, characterized in that, include: A neutron source that provides a collimated beam of neutrons; A conversion screen, positioned in front of the neutron emission direction of the neutron source across the sample, converts the irradiated neutrons into visible light; An imaging detector that detects visible light converted by the conversion screen in a multiple short-exposure mode; and The image processing apparatus as described in claim 3 or 4.

Images