A method for detecting changes in fundus image sequences

Abstract

This invention relates to a method for detecting changes in a fundus image sequence, comprising the following steps: S1: acquiring a sequence containing multiple retinal fundus images, selecting a background image from all the sequence images, expanding the background image, and registering the other images with the corresponding background image to obtain a registered background sequence image; S2: expanding the registered background sequence image, and performing adaptive contrast enhancement on both the expanded background sequence image and other images to generate an expanded sequence image set; S3: inputting the expanded sequence image set into a pre-constructed detection model for change detection to obtain the change results of the fundus image sequence. Compared with the prior art, this invention can construct a more robust background model, and the detected change areas are more obvious and accurate.

Description

Technical Field

[0001] This invention relates to the field of image sequence change detection technology, and in particular to a method for detecting changes in fundus image sequences. Background Technology

[0002] In computer-aided diagnostic (CAD) systems, change detection in retinal fundus image sequences is an important but challenging task. Medical images exhibit significant brightness variations, and illumination is complex and variable. Detecting changes in images of the same scene at different times is a very common problem in video surveillance, remote sensing, and medicine. However, change detection in longitudinal medical sequences differs from change detection in video sequences. Medical image sequences typically contain only two or a few images, with only one frame considered as background, which presents numerous challenges for change detection.

[0003] Current technologies rarely address the issue of change detection in longitudinal sequences of medical images because these sequences are typically captured when patients present with complex clinical symptoms, and these symptoms change over time. Change detection in medical image sequences could provide physicians with more visual information, helping them to quantitatively analyze areas of change and providing quantitative analysis for disease treatment and drug use.

[0004] Change detection in video sequences and longitudinal fundus image sequences differs significantly. On one hand, video sequences offer a large number of usable frames, many without foreground objects; these background frames are used to construct the background model. For longitudinal medical image sequences, typically only the image closest to normal is selected as the background model, while all other frames are considered foreground frames. On the other hand, the sampling span of consecutive frames in video surveillance is less than one second, resulting in continuous and smooth illumination changes within the sequence. However, longitudinal fundus sequences typically involve sampling over weeks or even months, and the longer time spans between sampling frames and varying imaging conditions lead to discontinuous and abrupt illumination changes in medical image sequences.

[0005] Change detection methods in video surveillance are primarily based on background modeling. Robust principal component analysis (RPCA) uses matrix decomposition to separate the changing parts from the background when the background is static and the foreground is dynamic with minimal changes. Masked-RPCA was proposed to address the problem of moving foregrounds overlapping on the background, thereby recovering sparse objects and obtaining low-rank components. These matrix decomposition-based RPCA methods transform each image into a column vector, destroying the spatial correlation of the image itself and making it difficult to obtain accurate and clear change regions. The TRPCA method decomposes video frames into a spatiotemporally correlated background and a spatiotemporally continuous foreground using total variation and Tucker decomposition. ETRCA preserves the salient content of the image by processing tensor singular values ​​differently. Although background modeling and background subtraction are well-established research areas, existing solutions are still susceptible to the frequent and complex lighting changes that occur in real-world scenarios, and these issues remain difficult to address. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a method for detecting changes in fundus image sequences. This invention can construct a more robust background model, and the detected change areas are more obvious and accurate.

[0007] The objective of this invention can be achieved through the following technical solutions: This invention provides a method for detecting changes in a fundus image sequence, comprising the following steps: S1: Obtain a sequence containing multiple retinal fundus images, select a background image from all sequence images, perform illumination correction on all sequence images, normalize the brightness of the background image and other images in the sequence image one by one, and then expand the background image. Register other images in the sequence image with the background image corresponding to the brightness of the sequence image to obtain the registered background sequence. S2: Perform secondary interpolation on the registered background sequence image to expand the background sequence image. Perform adaptive contrast enhancement on the secondary interpolated background sequence image and other images in the sequence image to generate an expanded sequence image set. S3: Input the expanded sequence image set into the pre-built detection model to perform change detection and obtain the change results of the fundus image sequence.

[0008] Preferably, step S1 specifically includes the following steps: S101: Select the image with the most normal disease state from all sequence images as the background image; S102: Process all retinal fundus images using the IRHSF model. By combining homomorphic filtering, curve fitting, and fundus image information, estimate the reflectance component of the image to reduce the illumination difference within the image, and then perform illumination correction on the sequence of images. S103: Based on S102, the brightness of the background image and other images in the sequence are normalized one by one to reduce the brightness difference between images; S104: Using the PIIFD method, register other images in the sequence image and the background image corresponding to the brightness of the sequence image respectively to obtain the registered background sequence image.

[0009] Preferably, in step S102, the process of processing all retinal fundus images using the IRHSF model specifically includes: Image Modeling as illumination components and reflection component The product of: in, Indicates the wavelength of the color channel; By using a logarithmic transformation, each specific channel image is mapped to a new scale space, and a fourth-order polynomial with 15 parameters is constructed to simulate the illumination pattern. express The logarithmic image under a specific channel image, then The parametric surface fitting equation formula for IRHSF is as follows: in, It is a diagonal weight matrix of size squared to the input image resolution. Non-zero elements are pixel brightness values ​​in the region of interest. These are the illumination parameters; the IRHSF illumination component and reflection component can be obtained using the following formula: .

[0010] Preferably, in step S103, the illumination normalization process is as follows: Assuming the sequence contains Zhang Image , , It is the background model. After brightness correction, the sequence can be represented as Select the background The brightness is normalized to the brightness of other images in the sequence, using the following formula: It is the normalized background, and the first in the sequence. Frames with the same brightness level and background sequences with different brightness levels can be represented as follows: .

[0011] Preferably, in step S2, the process of performing secondary interpolation on the registered background sequence specifically involves: Assume that in the background subsequence after brightness normalization there are Zhang Image It is the first in the background subsequence Zhang background image, in, Indicates the first Pixels in a background image The brightness value at that location, This represents a time period corresponding to an image sequence, consisting of three consecutive background images. and To obtain a quadratic sampling model, due to , and They are the first The, the The and the first Pixels of an image The corresponding brightness value at that point, therefore, through three consecutive points , and The relationship is used to construct a background model by setting different time periods. This obtains a new brightness value different from the sampling point. By traversing all pixels of the image, a new brightness value is obtained. The background model of the entire image, consisting of individual pixels.

[0012] Preferably, in step S3, the detection model is an improved H-TRPCA model. The construction process and model of the improved H-TRPCA model are as follows: In the formula, , and These represent the image sequence matrix, transformation matrix, and perturbation, respectively. It is a vectorized representation of the ideal background model. Represents the changing components spatiotemporal continuity, yes Norm, yes Norm, It's a hyperparameter. Indicates the tensor along the first Expand each dimension into a matrix, and then perform matrix multiplication.

[0013] Preferably, the construction process of the improved H-TRPCA model is as follows: The entire image sequence is modeled as a tensor using TRPCA. The size of the images in the sequence is There are a total of Frame image, tensor Decomposed into robust background components and change components The sum, based on robust background components and change components Background modeling and change region modeling are performed separately to construct an improved H-TRPCA model.

[0014] Preferably, based on robust background components The process of background modeling is as follows: Stable background weight It can be decomposed into low-rank components. and disturbance components The sum of, therefore: Expand the background component along the time direction ,get : in, It is a low-rank matrix, rank , It is a disturbance. It is a matrix. It is the number of frames in the image sequence. Unfold the background component along the height and width directions to obtain and : in and It is a high-rank matrix, and its rank satisfies the following conditions: and , and These are the height and width of the images in the sequence, respectively. and It is a matrix. Through Tucker decomposition It can be represented as follows: in It is the kernel tensor. It is a disturbance component. Indicates the tensor along the first Expand each dimension into a matrix, then perform matrix multiplication, and let... ,but It is an ideal background model.

[0015] Preferably, based on the changing components The process of modeling the change region is as follows: For the changing components The total variational method is used to constrain the spatiotemporal continuity of the changing region, as shown below: in, express medium voxel Spatiotemporal continuity at that point For all voxels, we have: Expand along the height, width, and time directions respectively. Then we can get Continuity along three different directions is represented as follows: Will , and Vectorization yields: The expression is as follows: in, It represents the spatiotemporal continuity of the changing components.

[0016] Preferably, in step S3, the detection model is an improved PG-TRPCA model, which is as follows: In the formula, , and These represent the image sequence matrix, transformation matrix, and perturbation, respectively. It is a vectorized representation of the background tensor. Represents the changing components spatiotemporal continuity, yes Norm, yes Norm, It's a hyperparameter. Indicates the tensor along the first Expand each dimension into a matrix, and then perform matrix multiplication.

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention provides a method for detecting changes in fundus image sequences by combining quadratic interpolation with tensor decomposition. The quadratic interpolation is used to expand the background sequence, making the constructed background model more robust. (2) The present invention provides a method for detecting changes in fundus image sequences. By constructing an improved PG-TRPCA model and an improved H-TRPCA model, and using three-dimensional total variation to constrain the spatiotemporal correlation between the background and the change region, the change detection results obtained are clearer and more accurate. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a method for detecting changes in a fundus image sequence provided in this embodiment.

[0019] Figure 2 This is a foreground sequence of 30 retinal fundus images.

[0020] Figure 3 (a) is the background model. Figure 3 (b) is the background model after IRHSF correction. Figure 3 (c) is the IRHSF illumination model.

[0021] Figure 4 This is the result of registering the first image in the sequence with the background model using the PIIFD method.

[0022] Figure 5 These are 30 background images obtained through double interpolation.

[0023] Figure 6 These are 91 images from the entire sequence after ACE processing.

[0024] Figure 7 , Figure 8 and Figure 9The changes detected in the 3rd, 12th, and 29th images of the sequence are compared. (a), (c), (e), and (g) are the background models reconstructed by H-TRPCA, PG-TRPCA, Improved H-TRPCA, and Improved PG-TRPCA, respectively, and (b), (d), (f), and (h) are their absolute difference images, respectively.

[0025] Figure 10 , Figure 11 and Figure 12 The PR and ROC curves for the 3rd, 12th, and 29th images in the sequence are the curves obtained by different methods.

[0026] Figure 13 This section compares the detection results of Masked-RPCA, ETRPCA, Improved H-TRPCA, and Improved PG-TRPCA. The first row shows the original image; 1a and 1b show the background and change detection results constructed using Masked-RPCA, respectively; 2a and 2b show the background and change detection results constructed using ETRPCA, respectively; 3a and 3b show the background and change detection results constructed using Improved H-TRPCA, respectively; and 4a and 4b show the background and change detection results constructed using Improved PG-TRPCA, respectively.

[0027] Figure 14 The comparison results for the third fundus image in the sequence with and without ACE are shown. (a) is the absolute difference image obtained by combining PG-TRPCA with quadratic interpolation and ACE (i.e., improved PG-TRPCA), (b) is the absolute difference image obtained by combining PG-TRPCA with quadratic interpolation, (c) is the absolute difference image obtained by combining H-TRPCA with quadratic interpolation and ACE (i.e., improved H-TRPCA), and (d) is the absolute difference image obtained by combining H-TRPCA with quadratic interpolation.

[0028] Figure 15 Comparison of the detection results of matrix RPCA and Improved PG-TRPCA at frame 9. (a) is frame 9 in the sequence, (b) is the absolute difference image based on matrix RPCA, (c) is the background reconstructed based on Improved PG-TRPCA, and (d) is the absolute difference image based on Improved PG-TRPCA.

[0029] Figure 16Comparison of detection results based on matrix RPCA and improved PG-TRPCA for frame 30. (a) is frame 9 in the sequence, (b) is the absolute difference image based on matrix RPCA, (c) is the background reconstructed based on improved PG-TRPCA, and (d) is the absolute difference image based on improved PG-TRPCA.

[0030] Figure 17 The results of comparing different change detection methods, namely matrix RPCA, improved PG-TRPCA, PG-TRPCA, improved H-TRPCA, and H-TRPCA, on the third fundus image are as follows: Figure 17 (d) to Figure 17 As shown in (h). Detailed Implementation

[0031] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0032] refer to Figure 1 As shown, this embodiment provides a method for detecting changes in a fundus image sequence, including the following steps: S1: Obtain a sequence containing multiple retinal fundus images. Select the image with the most normal condition from all sequence images as the background image. Perform illumination correction on all sequence images. Normalize the brightness of the background image and other images in the sequence image one by one. Then expand the background image. Register the other images in the sequence image with the background image corresponding to the brightness of the sequence image to obtain the registered background sequence.

[0033] In this embodiment, multiple retinal fundus images were obtained from a patient with wet age-related macular degeneration, spanning from April 24, 2009 to August 6, 2015. The significant differences in illumination within and between these images greatly interfered with the change detection results. (Reference) Figure 2 and Figure 3 As shown, a total of 31 image sequences were acquired. The original image sequence size was 1612×1536. After cropping, the image size was adjusted to 403×384. S101: Select the fourth image from the 31 sequential images, showing the most normal patient condition, as the background image. (Reference) Figure 3 As shown in (a).

[0034] S102: The 31 retinal fundus images were processed using the IRHSF (Iterative Robust Homomorphic Surface Fitting) model. By combining homomorphic filtering, curve fitting and fundus image information, the reflectance components of the images were estimated to reduce the illumination differences within the images, thereby performing illumination correction on the sequence of images.

[0035] image It can be modeled as illumination components and reflection component The product of: in, Indicates the wavelength of the color channel.

[0036] By using logarithmic transformation, each specific channel image is mapped to a new scale space, and then a fourth-order polynomial with 15 parameters is constructed to simulate the illumination pattern. express The logarithmic image under a specific channel image, then The parametric surface fitting equation for IRHSF is then: in It is a diagonal weight matrix of size squared to the input image resolution. Non-zero elements are pixel brightness values ​​in the region of interest. Since these are illumination parameters, the IRHSF illumination component and reflection component can be obtained using the following formula: The result of the background model after IRHSF correction is as follows: Figure 3 As shown in (b).

[0037] S103: Normalize the brightness of the background image and other images in the image sequence one by one to reduce the brightness difference between images. The illumination normalization process is as follows: Assuming the sequence contains Zhang Image , . It is the background model. After brightness correction, the sequence can be represented as follows: Select the background The brightness is normalized to the brightness of other images in the sequence, using the following formula: It is the normalized background, and the first in the sequence. The frames have the same brightness level. Therefore, a background sequence with different brightness levels can be represented as follows: At this point, the background sequence is expanded to 31 images, and the entire image sequence is expanded to 61 images.

[0038] S104: Using the PIFD method, other images in the image sequence are registered with the background image corresponding to the brightness of the image sequence. Specifically, PIFD uses a Harris detector to extract corner points with partial invariance, and then performs matching calculations. The image registration result of the first image and the background image corresponding to the brightness of the first image is as follows: Figure 4 As shown.

[0039] S2: Perform secondary interpolation on the registered background sequence image to expand the background sequence image. Perform adaptive contrast enhancement on the secondary interpolated background sequence image and other images in the sequence image to generate an expanded sequence image set.

[0040] Specifically, the process of performing secondary interpolation on the registered background sequence is as follows: Assume that in the background subsequence after brightness normalization there are Zhang image. It is the first in the background subsequence Zhang background image, in, Indicates the first Pixels in a background image The brightness value at that location, This represents a time segment corresponding to the image sequence. We take three consecutive background images. and Obtain the secondary sampling model. Because... , and They are the first The, the The and the first Pixels of an image The corresponding brightness value at that point can be obtained through three consecutive points. , and Based on the relationship, a background model is constructed in a targeted manner. This is achieved by setting different time periods. This yields a new brightness value different from the sampled point. By traversing all pixels of the image, a new brightness value can be obtained. It can model the background of the entire image for each pixel and robustly estimate new background images.

[0041] In this embodiment, 31 background sequences are interpolated twice to obtain 61 background sequences. The 30 background sequences obtained through the second interpolation are as follows: Figure 5 As shown. The result of ACE processing on all 91 images is as follows. Figure 6 As shown.

[0042] S3: Input the expanded set of 91 retinal fundus image sequences into the pre-built detection model to perform change detection and obtain the change results of the fundus image sequence.

[0043] The detection models include Improved PG-TRPCA (Improved Patch-group-based TensorRobust Principal Component Analysis) and Improved H-TRPCA (Improved HolisticTensor Robust Principal Component Analysis). Improved H-TRPCA first models the entire sequence as a tensor volume, and then separates the regions of variation from the background model. Improved PG-TRPCA constructs local tensor volumes by segmenting the tensor volume into blocks, focusing more on local features and reducing the interference of random noise.

[0044] Specifically, the construction process of the improved H-TRPCA model is as follows: The entire image sequence is modeled as a tensor volume using TRPCA (Tensor Robust Principal Component Analysis). , tensor Decomposed into robust background components and change components sum.

[0045] Among them, the robust background component can be further decomposed into low-rank components. and disturbance components The sum of, then It can be represented as In longitudinal fundus image sequences, the anatomical structures and textures in each image are similar and exhibit strong spatial and temporal correlations. The background component is expanded along the temporal direction. ,get : in, It is a low-rank matrix, rank , It is a disturbance. It is a matrix. It is the number of frames in the image sequence.

[0046] Unfold the background component along the height and width directions to obtain and : in and It is a high-rank matrix, and its rank satisfies the following conditions: and , and These are the height and width of the images in the sequence, respectively. and It is a matrix.

[0047] Through Tucker decomposition It can be represented as follows: in It is the kernel tensor. It is a disturbance component. Indicates the tensor along the first Expand each dimension into a matrix, then perform matrix multiplication. Let ,but It is an ideal background model.

[0048] For the changing components As time progresses, lesions dynamically evolve in both the spatial and temporal domains, with the changing regions typically located within a continuous area. Therefore, the total variational method is used to constrain the spatiotemporal continuity of the changing regions, as shown below: in, express medium voxel The spatiotemporal continuity of the location.

[0049] For all voxels, we have: Expand along the height, width, and time directions respectively. Then we can get Continuity along three different directions is represented as follows: Will , and Vectorization yields: The expression is as follows: in, It represents the spatiotemporal continuity of the changing components.

[0050] Therefore, the improved H-TRPCA model is as follows: In the formula, , and These represent the image sequence matrix, transformation matrix, and perturbation, respectively. It is a vectorized representation of the ideal background model. Represents the changing components spatiotemporal continuity, yes Norm, yes Norm, It's a hyperparameter. Indicates the tensor along the first Expand each dimension into a matrix, and then perform matrix multiplication.

[0051] The construction process of the improved PG-TRPCA model is as follows: Improved PG-TRPCA divides the tensor volume into segments of size [size missing]. Multiple overlapping 3D blocks, all of which form a block set. ,in It is an index set. It is the first A 3D block. Background component. It can be divided into There are clusters, each of which is a 4th-order tensor. yes The Middle Vectorized representation of tensor families, which can be adjusted to 4th order tensors. ,in, , Due to intra-cluster similarity, it can be used To approximate Then, the following optimization problem is solved to reconstruct the problem. : The analytical solution to the above equation is: Due to the 4th order tensor There is a strong correlation along the time direction. If it has low rank, then Tucker decomposition modeling can be used: in, It is the kernel tensor. and This represents an orthogonal factor matrix.

[0052] The improved PG-TRPCA model is as follows: In the formula, , and These represent the image sequence matrix, transformation matrix, and perturbation, respectively. It is a vectorized representation of the background tensor. Represents the changing components spatiotemporal continuity, yes Norm, yes Norm, It's a hyperparameter.

[0053] The method for detecting changes in fundus image sequences provided in this embodiment was applied to 31 image sequences from the same patient with wet age-related macular degeneration from April 24, 2009 to August 6, 2015. The experimental results are shown below: Figures 7-9 The images show the detection results for the 3rd, 12th, and 29th images in the sequence. (a), (c), (e), and (g) are the background models reconstructed by H-TRPCA, PG-TRPCA, Improved H-TRPCA, and Improved PG-TRPCA, respectively, and (b), (d), (f), and (h) are their absolute difference images, respectively. Figure 7 The change detection results for the third image are shown. As indicated by the white box, Improved H-TRPCA and Improved PG-TRPCA detected more regions of change than the original H-TRPCA and PG-TRPCA, respectively. Figure 8 As shown, the original PG-TRPCA and H-TRPCA detection results show significant optic disc interference. Figure 9 The detection results for frame 29 are shown. Compared to the original H-TRPCA, the improved H-TRPCA results show less interference from blood vessels. Furthermore, compared to the original PG-TRPCA, the improved PG-TRPCA detects clearer areas of change.

[0054] Figures 10-12The figures show the PR and ROC curves for different methods on images 3, 12, and 29 of the sequence, respectively. It can be seen that the AUC and MAP values ​​of Improved H-TRPCA and Improved PG-TRPCA are higher than those of the other methods, with Improved PG-TRPCA outperforming Improved H-TRPCA in most cases.

[0055] Figure 13 This is a visualization comparison of the results from Masked-RPCA, ETRPCA, Improved H-TRPCA, and Improved PG-TRPCA. It can be seen that in the change detection results of Masked-RPCA, interference caused by vascular disturbance and illumination changes is very significant. The background reconstructed by ETRPCA changes with frame rate, making it difficult to build a robust background model. Improved H-TRPCA and Improved PG-TRPCA outperform these two methods. They not only build robust background models from a limited number of frames but also effectively reduce interference caused by vascular and illumination changes.

[0056] Illumination variation is a significant factor affecting change detection. PG-TRPCA and H-TRPCA consider spatial and temporal consistency but neglect the complexity of illumination. ACE can correct for image brightness and help build robust background models. Figure 14 The results show a comparison of using ACE and not using ACE in the third fundus image of the sequence. Combining TRPCA with quadratic interpolation and ACE, the experimental results demonstrate strong robustness to illumination interference. Figure 14 (a) and Figure 14 (c) shows a comparison Figure 14 (b) and Figure 14 (d) Clearer areas of change with less visual interference. By using ACE, TRPCA can detect areas of change more accurately.

[0057] Figure 15 The comparison of detection results based on matrix RPCA and improved PG-TRPCA is shown in frame 9. When there is some random brightness interference in the fundus image, the results of matrix RPCA are severely affected because matrix RPCA only considers the temporal correlation of the background model and ignores the temporal correlation of the changing regions. Figure 15 (d) shows that the improved PG-TRPCA can handle random luminance interference well.

[0058] Figure 16 The comparison of detection results based on matrix RPCA and improved PG-TRPCA is shown in frame 30. Figure 16 (b) shows that the changes detected by matrix RPCA are significantly affected by vascular interference. In the improved PG-TRPCA results, vascular interference is less significant.

[0059] Figure 17 The comparison of different change detection methods, namely matrix RPCA, improved PG-TRPCA, PG-TRPCA, improved H-TRPCA, and H-TRPCA, on the third fundus image is shown below. Figure 17 (d) to Figure 17 As shown in (h). Figure 17 As shown in the red box in (c), the selected background image has a large, very dark spot. Compared to PG-TRPCA and H-TRPCA, matrix RPCA detects clearer areas of change, but ignores small areas of change, retaining only large areas. Figure 17 (e) and Figure 17 As shown in (g), our method can even detect faint light spots and obtain clearer regions of variation than the original PG-TRPCA and H-TRPCA.

[0060] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for detecting changes in a fundus image sequence, characterized in that, Includes the following steps: S1: Obtain a sequence containing multiple retinal fundus images, select a background image from all sequence images, perform illumination correction on all sequence images, normalize the brightness of the background image and other images in the sequence image one by one, and then expand the background image. Register other images in the sequence image with the background image corresponding to the brightness of the sequence image to obtain the registered background sequence. S2: Perform secondary interpolation on the registered background sequence image to expand the background sequence image. Perform adaptive contrast enhancement on the secondary interpolated background sequence image and other images in the sequence image to generate an expanded sequence image set. S3: Input the expanded sequence image set into the pre-built detection model to perform change detection and obtain the change results of the fundus image sequence; In step S2, the process of performing secondary interpolation on the registered background sequence is as follows: Assume that in the background subsequence after brightness normalization there are Zhang Image It is the first in the background subsequence Zhang background image, in, Indicates the first Pixels in a background image The brightness value at that location, This represents a time period corresponding to an image sequence, consisting of three consecutive background images. and To obtain a quadratic sampling model, due to , and They are the first The, the The and the first pixels of an image The corresponding brightness value at that point, therefore, through three consecutive points , and The relationship is used to construct a background model by setting different time periods. This obtains a new brightness value different from the sampling point. By traversing all pixels of the image, a new brightness value is obtained. The background model of the entire image, consisting of 10 pixels.

2. The method for detecting changes in a fundus image sequence according to claim 1, characterized in that, S1 specifically includes the following steps: S101: Select the image with the most normal disease state from all sequence images as the background image; S102: Process all retinal fundus images using the IRHSF model. By combining homomorphic filtering, curve fitting, and fundus image information, estimate the reflectance component of the image to reduce the illumination difference within the image, and then perform illumination correction on the sequence of images. S103: Based on S102, the brightness of the background image and other images in the sequence are normalized one by one to reduce the brightness difference between images; S104: Using the PIIFD method, register other images in the sequence image and the background image corresponding to the brightness of the sequence image respectively to obtain the registered background sequence image.

3. The method for detecting changes in a fundus image sequence according to claim 2, characterized in that, In step S102, the process of processing all retinal fundus images using the IRHSF model is as follows: Image Modeling as illumination components and reflection component The product of: in, Indicates the wavelength of the color channel; By using a logarithmic transformation, each specific channel image is mapped to a new scale space, and a fourth-order polynomial with 15 parameters is constructed to simulate the illumination pattern. express The logarithmic image under a specific channel image, then The parametric surface fitting equation formula for IRHSF is as follows: in, It is a diagonal weight matrix of size squared to the input image resolution. Non-zero elements are pixel brightness values ​​in the region of interest. These are the illumination parameters; the IRHSF illumination component and reflection component can be obtained using the following formula: 。 4. The method for detecting changes in a fundus image sequence according to claim 2, characterized in that, In step S103, the illumination normalization process is as follows: Assuming the sequence contains Zhang Image , , It is the background model. After brightness correction, the sequence can be represented as Select the background The brightness is normalized to the brightness of other images in the sequence, using the following formula: It is the normalized background, and the first in the sequence. Frames with the same brightness level and background sequences with different brightness levels can be represented as follows: .

5. The method for detecting changes in a fundus image sequence according to claim 1, characterized in that, In S3, the detection model is an improved H-TRPCA model. The construction process and model of the improved H-TRPCA model are as follows: In the formula, , and These represent the image sequence matrix, transformation matrix, and perturbation, respectively. It is a vectorized representation of the ideal background model. Represents the variable components spatiotemporal continuity, yes Norm, yes Norm, It's a hyperparameter. Indicates the tensor along the first Expand each dimension into a matrix, then perform matrix multiplication. It is the kernel tensor. and It is a high-rank matrix. It is a low-rank matrix.

6. The method for detecting changes in a fundus image sequence according to claim 5, characterized in that, The construction process of the improved H-TRPCA model is as follows: The entire image sequence is modeled as a tensor using TRPCA. The size of the images in the sequence is There are a total of Frame image, tensor Decomposed into robust background components and change components The sum, based on robust background components and change components Background modeling and change region modeling are performed separately to construct an improved H-TRPCA model.

7. The method for detecting changes in a fundus image sequence according to claim 6, characterized in that, Based on robust background components The process of background modeling is as follows: Stable background weight It can be decomposed into low-rank components. and disturbance components The sum of, therefore: Expand the background component along the time direction ,get : in, It is a low-rank matrix, rank , It is a disturbance. It is a matrix. It is the number of frames in the image sequence. Unfold the background component along the height and width directions to obtain and : in and It is a high-rank matrix, and its rank satisfies the following conditions: and , and These are the height and width of the images in the sequence, respectively. and It is a matrix; Through Tucker decomposition It can be represented as follows: in It is the kernel tensor. It is a disturbance component. Indicates the tensor along the first Expand each dimension into a matrix, then perform matrix multiplication, and let... ,but It is an ideal background model.

8. The method for detecting changes in a fundus image sequence according to claim 6, characterized in that, According to the changing components The process of modeling the change region is as follows: For the changing components The total variational method is used to constrain the spatiotemporal continuity of the changing region, as shown below: in, express medium voxel Spatiotemporal continuity at that point For all voxels, we have: Expand along the height, width, and time directions respectively. Then we can get Continuity along three different directions is represented as follows: Will , and Vectorization yields: The expression is as follows: in, It represents the spatiotemporal continuity of the changing components.

9. The method for detecting changes in a fundus image sequence according to claim 1, characterized in that, In S3, the detection model is the Improved PG-TRPCA model, which is as follows: In the formula, , and These represent the image sequence matrix, transformation matrix, and perturbation, respectively. It is a vectorized representation of the background tensor. Represents the variable components spatiotemporal continuity, yes Norm, yes Norm, It's a hyperparameter. Indicates the tensor along the first Expand each dimension into a matrix, and then perform matrix multiplication.

Images