Methods, devices, storage media, and computer equipment for screening redundant images from gastrointestinal capsule endoscopy.

Abstract

This application provides a method, device, storage medium, and computer equipment for screening redundant images from a gastrointestinal capsule endoscopy image set. The screening method includes: identifying feature points in each frame of a gastrointestinal capsule endoscopy image set; matching feature points between adjacent frames to obtain matching feature points and a feature matching rate; calculating the relative motion distance between adjacent frames based on the matching feature points; performing feature distribution difference analysis on adjacent frames to obtain a difference degree; calculating the redundancy degree of adjacent frames based on the feature matching rate, relative motion distance, and difference degree; and removing or retaining the next frame in the adjacent frames based on the redundancy degree. This screening method comprehensively considers multiple factors and dynamically adjusts the screening conditions to adapt to different scenarios, improving the accuracy and effectiveness of redundant image screening.

Description

Technical Field

[0001] This application belongs to the field of capsule endoscopy technology, and more specifically, relates to a method, screening device, storage medium, and computer equipment for screening redundant images from a digestive tract capsule endoscope. Background Technology

[0002] With the development of medical and information technologies, wireless capsule endoscopy (WCE), or simply capsule endoscopy, has become an important tool for diagnosing gastrointestinal diseases. This miniature imaging device, due to its patient-centered, non-invasive, convenient, and accessible nature, which is difficult for traditional endoscopes to reach, demonstrates significant advantages in the early diagnosis, lesion monitoring, and screening of gastrointestinal diseases.

[0003] Capsule endoscopy typically acquires tens of thousands or even hundreds of thousands of image sequences during the entire examination. These images not only provide clinicians with rich information about the digestive tract environment but also provide data support for computer-aided diagnosis and intelligent medical systems. At the same time, the screening, processing, and detection of massive amounts of data have brought unprecedented challenges, requiring efficient and accurate data screening and detection while ensuring diagnostic accuracy.

[0004] While capsule endoscopy has opened up new avenues for clinical diagnosis, image redundancy has become increasingly prominent in practical applications. Due to the restricted movement of the capsule within the digestive tract and the strong continuity of the scene, a large amount of highly similar redundant data exists in the acquired image sequences. These redundant images not only increase the workload of clinicians in interpreting images but also increase the computational and storage costs of the system during subsequent image analysis, data storage, and transmission, prolonging the diagnostic cycle and potentially affecting the highlighting of lesions, thus increasing the risk of misdiagnosis and missed diagnosis. To address the limitations of redundant images and achieve efficient and accurate examination, redundant image screening technology has become crucial. Redundant image screening determines whether an image is redundant by evaluating the similarity, motion information, and structural consistency of each frame with adjacent frames. However, due to the irregular shape, sparse texture, and high repetitiveness of the digestive tract, redundant image screening faces many challenges. Therefore, there is an urgent need to develop a redundant image screening method for capsule endoscopy to achieve efficient and intelligent screening and lesion protection for wireless capsule endoscopy data.

[0005] Existing research has proposed a series of screening methods, including redundant image removal methods based on low-level feature comparison using inter-frame correlation and color histograms, local block matching algorithms based on PatchMatch, self-supervised motion estimation methods, and classification models combining deep neural networks and support vector machines. These methods have achieved certain redundant image removal effects in specific scenarios. However, most of these methods rely on static rules and cannot dynamically adapt to the movement state and lesion feature distribution of capsules in different digestive tract regions, generally suffering from poor adaptability to scene changes. Summary of the Invention

[0006] The technical problem addressed in this application is: how to provide a method for screening redundant images from a digestive tract capsule endoscopy that is highly adaptable to scene changes and has dynamically adjustable thresholds.

[0007] This application provides a method for screening redundant images from a digestive tract capsule endoscopy, the screening method comprising:

[0008] Feature point identification was performed on each frame of the digestive tract capsule endoscopy image set;

[0009] Feature points of adjacent frames in a set of capsule endoscopy images of the digestive tract are matched to obtain the matched feature points and feature matching rate.

[0010] The relative motion distance between two adjacent image frames is calculated based on the matched feature points.

[0011] The feature distribution difference analysis is performed on the two adjacent frames to obtain the difference degree;

[0012] The redundancy of two adjacent frames is calculated based on the feature matching rate, the relative motion distance, and the difference degree. The latter frame of the two adjacent frames is then discarded or retained based on the redundancy degree.

[0013] Optionally, methods for feature point recognition of each frame of a digestive tract capsule endoscopy image set include:

[0014] Each frame of the image is preprocessed, including adaptive histogram equalization, Laplacian filtering, and image normalization.

[0015] Feature point recognition is performed on each frame of the preprocessed image using a feature point extraction model based on deep learning.

[0016] Optionally, methods for matching feature points between two adjacent frames in a digestive tract capsule endoscopy image set include:

[0017] A deep learning-based matching algorithm is used to match feature points between two adjacent frames of images to obtain an initial set of matching feature points.

[0018] Based on the density threshold condition and the geometric consistency condition, some feature points in the initial set of matching feature points are removed to obtain the final set of matching feature points.

[0019] Optionally, the method for calculating the relative motion distance between two adjacent frames based on matching feature points includes:

[0020] Two circles are obtained by fitting matching feature points on two adjacent image frames respectively;

[0021] The difference in radii between the two circles is used as the relative motion distance between two adjacent frames.

[0022] Optionally, the method for performing feature distribution difference analysis on the two adjacent frames to obtain the difference degree includes:

[0023] Feature maps of two adjacent frames are extracted respectively;

[0024] Dimensionality reduction is performed on the two feature maps respectively to obtain the feature distribution vectors of two adjacent frames;

[0025] The difference between two adjacent frames is calculated based on two feature distribution vectors.

[0026] Optionally, the method for calculating the redundancy of two adjacent frames based on the feature matching rate, the relative motion distance, and the difference degree includes:

[0027] Determine the weights for the feature matching rate, the relative motion distance, and the difference.

[0028] The sum of the products of the weight of the feature matching rate and the feature matching rate, the weight of the relative motion distance and the relative motion distance, and the weight of the difference and the difference is taken as the redundancy degree.

[0029] Optionally, the screening method further includes:

[0030] Before removing the next frame image from two adjacent frames based on redundancy, determine whether the next frame image is a lesion frame image.

[0031] If yes, then retain the next frame image; otherwise, discard the next frame image.

[0032] This application also discloses a device for screening redundant images from a digestive tract capsule endoscopy, the device comprising:

[0033] The feature point recognition unit is used to perform feature point recognition on each frame of the digestive tract capsule endoscopy image set.

[0034] The feature point matching unit is used to match feature points of two adjacent frames in the digestive tract capsule endoscopy image set to obtain matched feature points and feature matching rate.

[0035] A distance calculation unit is used to calculate the relative motion distance between two adjacent image frames based on matching feature points.

[0036] A feature distribution difference analysis unit is used to perform feature distribution difference analysis on two adjacent frames of images to obtain the difference degree.

[0037] An image filtering unit is used to calculate the redundancy of two adjacent frames based on the feature matching rate, the relative motion distance, and the difference degree, and to remove or retain the next frame image in the two adjacent frames based on the redundancy degree.

[0038] This application also discloses a computer-readable storage medium storing a screening program for redundant images from a digestive capsule endoscopy, wherein the screening program for redundant images from a digestive capsule endoscopy is executed by a processor to implement the above-described stitching method.

[0039] This application also discloses a computer device, which includes a computer-readable storage medium, a processor, and a screening program for redundant images of a digestive tract capsule endoscopy stored in the computer-readable storage medium. When the processor executes the screening program for redundant images of a digestive tract capsule endoscopy, it implements the above-described screening method for redundant images of a digestive tract capsule endoscopy.

[0040] This application provides a method for screening redundant images from a digestive tract capsule endoscopy, a feature point matching device, a computer-readable storage medium, and a computer device, which have the following technical advantages:

[0041] First, the feature matching rate, relative motion distance, and difference between two adjacent frames are calculated. Then, the redundancy of the two adjacent frames is calculated based on these three factors, which determines whether to remove or retain the images. This screening method takes into account multiple factors and dynamically adjusts the screening conditions to adapt to different scenarios and improve the accuracy and effectiveness of redundant image removal. Attached Figure Description

[0042] Figure 1 A flowchart of a method for screening redundant images from a digestive tract capsule endoscopy according to one or more embodiments;

[0043] Figure 2 A schematic block diagram of a device for screening redundant images from a digestive capsule endoscope according to one or more embodiments;

[0044] Figure 3 This is a schematic block diagram of a computer device according to one or more embodiments. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0046] Before describing the various embodiments of this application in detail, the technical concept of this application is first briefly described: Currently, when removing redundant images from gastrointestinal capsule endoscopy, there are problems such as the inability to dynamically adapt to different regions, different motion states, and different lesion feature distributions. Therefore, this application provides a method for screening redundant images from gastrointestinal capsule endoscopy. Its key improvements include calculating the feature matching rate, relative motion distance, and difference between two adjacent image frames, and further calculating the redundancy degree of the two adjacent image frames based on these three factors, thereby determining whether to remove or retain the image. This screening method comprehensively considers multiple factors and dynamically adjusts the screening conditions to adapt to different scenarios, improving the accuracy and effectiveness of redundant image removal. The specific principles of the gastrointestinal capsule endoscopy redundant image screening method of this application are described below with reference to more embodiments.

[0047] Specifically, such as Figure 1 As shown, the method for screening redundant images from a digestive tract capsule endoscopy in this embodiment includes the following steps:

[0048] S10. Feature point identification is performed on each frame of the digestive tract capsule endoscopy image set.

[0049] S20. Match the feature points of two adjacent frames in the digestive tract capsule endoscopy image set to obtain the matched feature points and feature matching rate.

[0050] S30. Calculate the relative motion distance between two adjacent frames based on the matching feature points;

[0051] S40. Perform feature distribution difference analysis on the two adjacent frames to obtain the difference degree;

[0052] S50. Calculate the redundancy of two adjacent frames based on the feature matching rate, the relative motion distance, and the difference degree, and discard or retain the next frame image in the two adjacent frames based on the redundancy degree.

[0053] In one or more embodiments, image acquisition and image preprocessing are performed before step S10. Specifically, images of the digestive tract are acquired using a wireless capsule endoscopy device. The wireless capsule endoscopy device has a built-in miniature camera that acquires images of the inside of the digestive tract via wireless transmission technology and transmits the image data to an external receiving device. The capsule endoscopy scans the entire digestive system, capturing images of the entire digestive tract region. After acquisition, the images are transmitted to the external receiving device via an antenna. A single capsule endoscopy examination process generates 70,000 to 80,000 images, providing sufficient data support for subsequent redundant screening. Furthermore, to improve the quantity and quality of feature point extraction, especially in weak texture areas or digestive tract scenes with uneven lighting, image preprocessing is required. First, the acquired images are initially screened to remove images irrelevant to the inside of the digestive tract, such as those taken after swallowing the capsule endoscopy or after expelling the contents of the digestive tract. Image preprocessing mainly includes adaptive histogram equalization (CLAHE), Laplacian filtering, and image normalization.

[0054] Adaptive Histogram Equalization (CLAHE) is a method to enhance local image contrast by dividing the image into small regions, performing histogram equalization on each region, and limiting contrast to avoid over-enhancement. Its core mapping formula is as follows:

[0055] ,

[0056] In the formula, S k The mapped grayscale value. L For gray levels, h ( r j () is the original grayscale histogram. C L The contrast limit threshold is T×T, where T×T is the sub-block size.

[0057] Laplacian filtering is an image enhancement method based on the second derivative, used to highlight edges and rapidly changing regions in an image, thereby enhancing feature point detection performance. A commonly used discrete form is:

[0058] ,

[0059] in, x , y Represents two-dimensional coordinates in an image. I Describing the degree value function, These represent the second derivatives of the image grayscale value in the horizontal and vertical directions, respectively, and are used to detect edge regions in the image where grayscale changes drastically.

[0060] Simultaneously, to meet the input requirements of deep learning models and improve feature extraction efficiency, standardization is performed. Standardization aims to eliminate differences between different images, ensuring consistency in the model's input. The mean standardization formula is as follows:

[0061] ,

[0062] In the formula, x This is the value before standardization. It is the standardized value. Represents the mean of the dataset. This represents the standard deviation of the dataset.

[0063] After image preprocessing, feature point extraction models based on deep learning (such as SuperPoint, ALIKED, RoMa, XFeat, etc.) are used to extract feature points from the image sequence. This automatically learns and extracts stable and discriminative keypoints and feature descriptors from the image, ensuring sufficient feature points in the image. This addresses the problem of traditional extraction methods (such as SIFT, SURF, etc.) having few or no features. The deep learning-based feature point extraction method can extract more stable and robust keypoints in image regions with low texture and complex backgrounds, improving image matching accuracy and solving the problem of strong texture dependence in traditional methods. This is the foundation for subsequent motion estimation and redundancy determination. This step provides high-quality and accurate feature point support, thereby improving the accuracy and robustness of subsequent screening.

[0064] After feature point recognition, a deep learning matching algorithm is used to match feature points between adjacent frames. Based on the extracted image feature points and descriptors obtained from the previous steps, a deep learning-based matching algorithm (such as SuperGlue, LoFTR, LightGlue, or other end-to-end feature matching networks) is also used to achieve highly robust feature point pair generation under spatial consistency constraints. Since deep learning-based matching algorithms lack rotation invariance, each image pair is rotated 0-360° for enhancement, using the rotation method with the highest matching degree. The output after matching includes the corresponding point coordinates and confidence score, providing stable input for subsequent geometric fitting and displacement calculation.

[0065] Furthermore, based on density threshold and geometric consistency conditions, some feature points in the initial matching feature point set are removed to obtain the final matching feature points. Specifically, to further eliminate mismatched point pairs and improve the accuracy of displacement estimation, a dual-constraint screening mechanism of density clustering and geometric consistency is introduced. For example, spatial clustering based on DBSCAN and geometric constraint removal based on MAGSAC are mainly used. DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is a density-based clustering method that specifies a neighborhood radius... ε The algorithm uses MinPts, a minimum number of points, to cluster feature points with high density in the space and remove outlier noise points. Its core expression is as follows:

[0066] ,

[0067] In the formula, p Indicates the current feature point, q Represents other feature points in the dataset. D Represents the set of all feature points. ε Used to define the proximity relationship between points. dist ( p , q )express p , q The Euclidean distance between them, if ,but p The core point.

[0068] For example, MAGSAC replaces hard threshold judgment with a score for each sampled model on a marginal basis and evaluates the support of each sample for the model based on a probabilistic model, achieving robust estimation without a threshold, as shown in the following formula:

[0069] ,

[0070] In the formula, e i Indicates sample i The error, Represents the loss function. Indicates the standard deviation The prior probability, This indicates the maximum boundary of the standard deviation.

[0071] After performing a false match removal operation on the matching feature points, the feature matching rate is calculated. m r Its definition is based on the sum of the number of valid feature point pairs matched in two frames of images and the total number of all feature points detected in the two frames of images. The formula is as follows:

[0072] ,

[0073] In the formula, N m This indicates the number of valid matching pairs retained after rejection. This represents the total number of feature points detected in the first frame of the image. This represents the total number of feature points detected in the second frame image.

[0074] In one or more embodiments, a method for calculating the relative motion distance between two adjacent frames based on matching feature points includes: fitting two circles on the two adjacent frames based on the matching feature points; and using the difference in radii of the two circles as the relative motion distance between the two adjacent frames.

[0075] For example, based on the filtered set of matching points, a circle model is fitted using the least squares method, and the motion amplitude is estimated by combining the inter-frame variations. Matching feature points between adjacent image sequences are fitted with two circles respectively, and the displacement is estimated using the fitting parameters between the two circles. The least squares method fits the coordinates of matching points in two adjacent frames on the image plane, with the fitting objective being to calculate the center coordinates and radius of the circle, minimizing the sum of squared residuals of the objective function, as shown in the following formula:

[0076] ,

[0077] In the formula, x i , y i Let α represent the coordinates of the matched feature points in the image. b The coordinates of the center of the circle. r This represents the radius of the circle.

[0078] For example, the fitting problem is transformed into a linear least squares problem, and a constraint matrix equation is constructed to be solved. The constraint matrix A is decomposed using the Singular Value Decomposition (SVD) algorithm. SVD has advantages such as strong numerical stability and high robustness to outliers, making it suitable for real-world scenarios where image matching points have slight noise interference. The specific formula is as follows:

[0079] ,

[0080] In the formula, A Represents the constraint matrix. U Let ∑ denote the left singular vector matrix, and let ∑ denote the singular value diagonal matrix. V T The matrix representing the transpose of the right singular vector. A + Indicates the pseudo-inverse of a matrix. This represents the least squares solution. Further solutions are needed to find the coordinates of the circle's center α. b With radius r :

[0081] ,

[0082] The displacement is estimated based on the difference in radii of the two fitted circles obtained from the corresponding matching points, as shown in the following formula:

[0083] ,

[0084] In the formula, D This represents the relative motion distance between two adjacent frames. r t , r t+1 This represents the radius of the fitted circle between two adjacent frames.

[0085] In one or more embodiments, the method for performing feature distribution difference analysis on the two adjacent frames of images to obtain the difference degree includes: extracting feature maps of the two adjacent frames of images respectively; performing dimensionality reduction processing on the two feature maps respectively to obtain feature distribution vectors of the two adjacent frames of images; and calculating the difference degree of the two adjacent frames of images based on the two feature distribution vectors.

[0086] For example, pre-trained or specially designed convolutional neural networks (CNNs), such as ResNet, MobileNet, or lightweight UNet, are used to extract features from the image. The final output layer is discarded, and a pre-trained backbone model is used to select high-level feature maps, such as the fourth or fifth layer, to obtain high-level semantic representations of the image, including texture, edges, and lesion structures. Assume the input feature image sequence is... I t The extracted feature map is represented as follows:

[0087] ,

[0088] In the formula, C For the number of channels, H , W Indicates the dimension of the feature map.

[0089] After feature map extraction, the original feature maps have high dimensionality and spatial complexity, which is not conducive to direct inter-image difference analysis. The feature maps of each frame are converted into one-dimensional feature distribution vectors to express the overall semantic feature distribution of the image in a more compact and easily comparable form. This process is accomplished by performing a global aggregation operation (such as global average pooling) on ​​the feature maps, which effectively preserves the global response information of each channel, improves the comparability between features, and significantly reduces computational complexity, laying the foundation for subsequent inter-frame similarity measurement. The three-dimensional feature maps are transformed into one-dimensional vectors that can be used for inter-image sequence difference measurement. Global Average Pooling (GAP) is used to reduce the dimensionality of the feature maps.

[0090] ,

[0091] In the formula, v t Let represent the feature distribution vector of the t-th frame image. This vector comprehensively reflects the semantic distribution characteristics of the entire image.

[0092] After generating the feature distribution vector, a difference analysis is performed on the feature distribution vectors of adjacent image frames to measure and filter semantic changes between image frames. The magnitude difference of the overall feature response is measured using Euclidean distance. To enhance adaptability and numerical stability, the final feature difference value can be uniformly normalized and used as a key parameter to measure the degree of change in image content. For any two adjacent image frames (e.g., ... I t , I t+1 Calculate the corresponding feature distribution vectors for each. v t and v t+1 Difference between :

[0093] ,

[0094] In the formula, Denotes the Euclidean norm. The denominator is used for normalization, so that the result value is stable within [0,1].

[0095] In one or more embodiments, a method for calculating the redundancy of two adjacent frames based on feature matching rate, relative motion distance, and difference degree includes: determining the weights of the feature matching rate, the relative motion distance, and the difference degree; and using the sum of the products of the weights of the feature matching rate and the feature matching rate, the weights of the relative motion distance and the relative motion distance, and the weights of the difference degree and the difference degree as the redundancy degree. The specific calculation formula is as follows:

[0096] ,

[0097] In the formula, α, b , c The weighting coefficients, which are adaptively adjusted based on the current image sequence, can be dynamically generated using heuristic rules. m r Indicates the feature matching rate. D Indicates the relative distance traveled. Indicates the degree of difference. According to... score Determine whether the next frame in two adjacent frames is a redundant image and remove it, then continue to determine the subsequent images.

[0098] Weighting coefficient α b , c These three metrics, used to measure the importance of feature matching rate, relative motion distance, and difference in redundancy determination, respectively, satisfy normalization constraints:

[0099] ,

[0100] Default weight settings To achieve a balance among various indicators, the weighting coefficients are dynamically set based on heuristic rules, which are dynamically formulated according to the characteristics of the digestive tract region (stomach, small intestine, colon) where the capsule endoscopy image is located.

[0101] In the stomach region, the displacement direction between images is unstable. Redundancy is assessed using image feature matching rate and difference, and α is appropriately increased. c The weights;

[0102] In the small intestine, the capsule moves along the intestinal wall. The distance traveled is determined by fitting a circle to characteristic points, and α is appropriately increased. b The weights;

[0103] In the colon region, judgment is made more based on movement distance and image difference, with appropriate enhancement. b and c The weight.

[0104] In one or more embodiments, before removing the next frame from two adjacent frames based on redundancy, it is determined whether the next frame is a lesion frame; if so, the next frame is retained; otherwise, it is removed. Specifically, considering the potential for accidental deletion of potential lesion frames during redundancy screening, a dual protection strategy is introduced. First, suspicious lesion frames are screened by analyzing the degree of local mutation in the depth feature distribution. Second, a lightweight convolutional neural network classification model is used to quickly classify or detect suspicious image sequences, identifying typical lesion morphologies. If a lesion is identified, the lesion image sequence is forcibly retained. This ensures that key image sequences potentially containing lesions can still be accurately retained even under high compression ratio screening.

[0105] like Figure 2 As shown in the figure, this embodiment also discloses a screening device for redundant images of a gastrointestinal capsule endoscopy. The screening device includes a feature point recognition unit 100, a feature point matching unit 200, a distance calculation unit 300, a feature distribution difference analysis unit 400, and an image filtering unit 500. The feature point recognition unit 100 is used to identify feature points in each frame of the gastrointestinal capsule endoscopy image set. The feature point matching unit 200 is used to match feature points in two adjacent frames of the gastrointestinal capsule endoscopy image set to obtain matched feature points and feature matching rate. The distance calculation unit 300 is used to calculate the relative motion distance between two adjacent frames based on the matched feature points. The feature distribution difference analysis unit 400 is used to perform feature distribution difference analysis on two adjacent frames to obtain the difference degree. The image filtering unit 500 is used to calculate the redundancy degree of two adjacent frames based on the feature matching rate, relative motion distance, and difference degree, and to remove or retain the next frame of the two adjacent frames based on the redundancy degree.

[0106] The specific working process of each unit of the screening device can be referred to the description in the previous embodiment, and will not be repeated here.

[0107] In one or more embodiments, a computer-readable storage medium stores a screening procedure for redundant images from a gastrointestinal capsule endoscopy, which, when executed by a processor, implements the screening method for redundant images from a gastrointestinal capsule endoscopy as described in Embodiment 1.

[0108] This embodiment also discloses a computer device, at the hardware level, such as... Figure 3As shown, the computer device includes a processor 12, an internal bus 13, a network interface 14, and a computer-readable storage medium 11. The processor 12 reads the corresponding computer program from the computer-readable storage medium and runs it, forming a request processing device at the logical level. Of course, in addition to the software implementation, one or more embodiments of this specification do not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices. The computer-readable storage medium 11 stores a screening program for redundant images of a digestive tract capsule endoscopy. When the processor executes the screening program for redundant images of a digestive tract capsule endoscopy, it implements the above-described screening method for redundant images of a digestive tract capsule endoscopy.

[0109] Computer-readable storage media include both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer-readable storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage, quantum memory, graphene-based storage media or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0110] The specific embodiments of this application have been described in detail above. Although some embodiments have been shown and described, those skilled in the art should understand that modifications and improvements can be made to these embodiments without departing from the principles and spirit of this application as defined by the claims and their equivalents, and such modifications and improvements should also be within the protection scope of this application.

Claims

1. A method for screening redundant images from a digestive tract capsule endoscopy, characterized in that, The screening method comprises: feature point recognition is performed on each frame image of the capsule endoscopy image set; feature points of adjacent two frame images in the capsule endoscopy image set are matched to obtain matching feature points and a feature matching rate; a relative motion distance between the adjacent two frame images is calculated according to the matching feature points; feature distribution difference analysis is performed on the adjacent two frame images to obtain a difference degree; a redundancy degree of the adjacent two frame images is calculated according to the feature matching rate, the relative motion distance and the difference degree, and a last frame image in the adjacent two frame images is removed or retained according to the redundancy degree; the method for calculating the redundancy degree of the adjacent two frame images according to the feature matching rate, the relative motion distance and the difference degree comprises: determining a weight of the feature matching rate, a weight of the relative motion distance and a weight of the difference degree; and taking a sum of a product of the weight of the feature matching rate and the feature matching rate, a product of the weight of the relative motion distance and the relative motion distance, and a product of the weight of the difference degree and the difference degree as the redundancy degree; Feature matching rate m r The calculation formula is as follows: , In the formula, N m represents the number of valid matching point pairs remaining after elimination, represents the total number of feature points detected in the first frame image, represents the total number of feature points detected in the second frame image.

2. The method of claim 1, wherein the method further comprises: the method for performing feature point recognition on each frame image of the capsule endoscopy image set comprises: preprocessing is performed on each frame image, and the preprocessing comprises adaptive histogram equalization, Laplace filtering and image standardization; feature point recognition is performed on each frame image after preprocessing by using a feature point extraction model based on deep learning.

3. The method of claim 1, wherein the method further comprises: the method for matching feature points of adjacent two frame images in the capsule endoscopy image set comprises: feature points of the adjacent two frame images are matched by using a matching algorithm based on deep learning to obtain an initial matching feature point set; part of the feature points in the initial matching feature point set are removed according to a density threshold condition and a geometric consistency condition to obtain final matching feature points.

4. The method of claim 3, wherein the method further comprises: the method for calculating the relative motion distance between the adjacent two frame images according to the matching feature points comprises: two circles are fitted on the matching feature points in the adjacent two frame images respectively; a radius difference of the two circles is taken as the relative motion distance between the adjacent two frame images.

5. The method of claim 1, wherein the method further comprises: the method for performing feature distribution difference analysis on the adjacent two frame images to obtain a difference degree comprises: feature maps of the adjacent two frame images are extracted respectively; dimension reduction processing is performed on the two feature maps respectively to obtain feature distribution vectors of the adjacent two frame images; a difference degree of the adjacent two frame images is calculated according to the two feature distribution vectors.

6. The method of claim 1, wherein the method further comprises: The screening method further comprises: before the last frame image in the adjacent two frame images is removed according to the redundancy degree, it is judged whether the last frame image is a lesion frame image; if yes, the last frame image is retained, and if no, the last frame image is removed.

7. A device for culling redundant images of a capsule endoscope of the digestive tract, characterized in that, The screening device comprises: a feature point recognition unit, which is configured to perform feature point recognition on each frame image of the capsule endoscopy image set; a feature point matching unit, which is configured to match feature points of adjacent two frame images in the capsule endoscopy image set to obtain matching feature points and a feature matching rate; a distance calculation unit configured to calculate a relative motion distance between the two adjacent frames according to the matched feature points; a feature distribution difference analysis unit configured to analyze feature distribution difference of the two adjacent frames to obtain a difference degree; an image screening unit configured to calculate a redundancy degree of the two adjacent frames according to the feature matching rate, the relative motion distance and the difference degree, and to eliminate or retain the latter one of the two adjacent frames according to the redundancy degree; the method for calculating the redundancy degree of the two adjacent frames according to the feature matching rate, the relative motion distance and the difference degree comprises: determining a weight of the feature matching rate, a weight of the relative motion distance and a weight of the difference degree; and taking a sum of a product of the weight of the feature matching rate and the feature matching rate, a product of the weight of the relative motion distance and the relative motion distance, and a product of the weight of the difference degree and the difference degree as the redundancy degree; Feature matching rate m r The calculation formula is as follows: ， In the formula, N m denotes the number of valid matching point pairs remaining after the elimination, denotes the total number of feature points detected in the first frame image, denotes the total number of feature points detected in the second frame image.

8. A computer-readable storage medium, characterized in that, the computer readable storage medium stores a redundant image screening program of a digestive tract capsule endoscope, and the redundant image screening program of the digestive tract capsule endoscope, when executed by the processor, implements the redundant image screening method of the digestive tract capsule endoscope according to any one of claims 1 to 6.

9. A computer device, comprising: the computer device comprises a computer readable storage medium, a processor and a redundant image screening program of a digestive tract capsule endoscope stored in the computer readable storage medium, and the redundant image screening program of the digestive tract capsule endoscope, when executed by the processor, implements the redundant image screening method of the digestive tract capsule endoscope according to any one of claims 1 to 6.

Images