A Smart Method for Detecting Book Cover Damage Based on Image Recognition

Abstract

This invention provides an intelligent method for detecting book cover damage based on image recognition. It comprehensively utilizes various image processing technologies and intelligent methods to match and compare the acquired image of the book cover to be detected with an undamaged reference book cover image. This detects whether the book cover is damaged and identifies the location, size, and shape of the damage. This provides an efficient technical reference solution for damage identification in automated borrowing and returning systems. Therefore, this invention employs a non-contact image detection and analysis method to detect the presence, location, size, and shape of damage on book covers, thereby regulating borrower behavior and indirectly strengthening the protection of paper books.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, and in particular relates to an intelligent method for detecting book cover damage based on image recognition. Background Technology

[0002] Book damage is a common problem in libraries today, mainly manifesting as stains, folded corners, cracks, and graffiti, which seriously affects the lifespan of books. Currently, automated library systems cannot effectively identify book damage, thus necessitating an efficient method for identifying book damage upon return to ensure the integrity of the books.

[0003] To address the problem of book damage, manual inspection is time-consuming and labor-intensive, leading to the development of various automated detection methods. Existing book damage detection methods include comparing weight changes before and after book borrowing and returning, using internal mechanical structures to inspect each page, and employing non-contact sensors to acquire edge lines and image information from the book's sides to detect foaming, deformation, and side defects. However, these methods suffer from low accuracy, the potential for secondary damage to the book, and a limited range of identifiable damage types.

[0004] Image recognition-based book damage detection methods offer advantages such as non-contact operation, high efficiency, and high accuracy, and are easily integrated with existing automated library systems. However, in the field of machine vision, a complete methodological framework for book damage detection has not yet been established. Related surface defect detection methods are mainly concentrated in the industrial sector, focusing on materials such as metals, wood, fabrics, and PCB circuit boards, and utilizing traditional image recognition methods and deep learning methods for the classification, localization, and segmentation of surface defects. Summary of the Invention

[0005] To address the problem that automated book return systems cannot identify book damage, this invention provides an intelligent book cover damage detection method based on image recognition. This method can automatically detect whether a book is damaged when it is returned, and can detect the location, size, and shape of the damage. It has extremely high detection efficiency for common damage forms such as stains, folds, and cracks, without the need for manual supervision and judgment, saving manpower and improving the efficiency of book borrowing and returning management.

[0006] A method for intelligent detection of book cover damage based on image recognition includes the following steps:

[0007] S1: Using the SIFT feature matching method, image matching is performed between the cover image A of the book to be returned and the undamaged reference cover image B, and four corner points corresponding to the four reference corner points of image B are obtained on image A.

[0008] S2: Use gradient descent to correct the positions of the four corner points, and obtain the perspective transformation matrix between the corrected positions of the four corner points and the four reference corner points;

[0009] S3: Perform a perspective transformation on image A according to the perspective transformation matrix to obtain an image A* that better matches image B;

[0010] S4: Obtain the difference image between image A* and image B;

[0011] S5: Use the trained network model Mask R-CNN to extract the difference regions in the difference images, and realize the automatic detection of the location, size and shape of the dirt.

[0012] Furthermore, obtaining the four corner points on image A corresponding to the four reference corner points of image B specifically involves:

[0013] S11: Apply the classic SIFT feature detection method to perform scale space construction, local extremum detection, key point orientation calculation and key point descriptor calculation on image B and image A respectively, to obtain 128-dimensional key point descriptors corresponding to each pixel of the two images;

[0014] S12: Apply KNN feature matching to perform similarity matching between the key point descriptors of image B and the key point descriptors of image A in step S11, and obtain n pairs of feature points with similar features;

[0015] S13: Using the RANSAC algorithm, homography matching is performed on the n pairs of feature points that were successfully matched in step S12 to obtain the homography matrix between image B and image A. Based on the homography matrix, homography transformation is performed on the four selected reference corner points on image B to obtain four corner points on image A corresponding to the four reference corner points.

[0016] Furthermore, the positions of the four corner points obtained are corrected using the gradient descent method as follows:

[0017] S21: The row and column information contained in the four corner points obtained on image A will be represented as an eight-dimensional vector x;

[0018] S22: Calculate the initial perspective transformation matrix between the four corner points of image A represented by the eight-dimensional vector x and the four reference corner points of image B, and perform perspective transformation on image A according to the initial transformation matrix. Use the mean square error between the perspective transformed image A and image B as the loss function, and the loss function is a function of the eight-dimensional vector x, denoted as L=f(x).

[0019] S23: Add a perturbation value d = 0.1 to each of the eight values ​​in the eight-dimensional vector x. Then, sequentially obtain the gradient components of the loss function corresponding to the new eight-dimensional vectors with the added perturbation values. Finally, integrate the eight obtained gradient components of the loss function into an eight-dimensional loss function gradient ΔL = [ΔL1 ΔL2 … ΔL8]. The method for obtaining the gradient components of the loss function corresponding to each new eight-dimensional vector is as follows:

[0020] Let the new eight-dimensional vector x be formed by adding a perturbation value d to the i-th value of the eight-dimensional vector x. Obtain the eight-dimensional vector according to step S22. Corresponding loss function components Then, the gradient components of the loss function are obtained using the first-order difference method.

[0021] S24: Construct the SGD-M optimizer based on the gradient ΔL of the loss function as follows:

[0022] x t+1 =x t -α·m t

[0023] m t =β·m t-1 +(1-β)·ΔL

[0024] Where, x t+1 Let x be an eight-dimensional vector in the (t+1)th iteration. t Let x be an eight-dimensional vector in the t-th iteration, and x t The initial value is the eight-dimensional vector x in step S21, α is the learning rate, and m t Let m be the auxiliary variable for the t-th iteration, β be the momentum coefficient, and m be the auxiliary variable for the t-th iteration. t-1 Auxiliary variables for the (t-1)th iteration;

[0025] Each iteration of the SGD-M optimizer determines ||x t+1 -x t If the result of either judgment is yes, the final eight-dimensional vector is obtained, and the correction of the four corner points obtained on image A is completed; if the result of either judgment is no, the next iteration in the SGD-M optimizer is entered until the final eight-dimensional vector is obtained.

[0026] Furthermore, obtaining the difference image between image A* and image B specifically involves:

[0027] S41: Obtain the local contrast similarity c(x, y) between image A* and image B in each color channel using the following method:

[0028] S41a: Calculate the local mean μ of image A*. x The local mean μ of image B y ;

[0029] S41b: Based on the local mean μ x Calculate the local weighted variance σ of image A*. x And based on the local mean μ y Calculate the local weighted variance σ of image B. y as follows:

[0030]

[0031]

[0032] Among them, w i (·) 2 Let x be a circularly symmetric Gaussian function with standard deviation σ = 1.5, i = 1, 2, ..., N, where N is the number of pixels in image A* or image B, and x i Let y be the pixel value of the i-th pixel in image A*. i Let be the pixel value of the i-th pixel in image B;

[0033] S41c: Based on the local weighted variance σ x and local weighted variance σ y Calculate the local contrast similarity c(x, y) between image A* and image B:

[0034]

[0035] Where C is a set constant;

[0036] S42: Linearly map the local contrast similarity c(x, y) of image A* and image B in each color channel to the color code range [0, 255] to obtain the difference image between image A* and image B.

[0037] Furthermore, the training method for the Mask R-CNN network model is as follows:

[0038] S51: Prepare multiple book cover images to be tested, and obtain the difference images corresponding to each book cover image to be tested. At the same time, use image masks to mark the smudged areas of each book cover image to be tested. Among them, the difference image corresponding to a book cover image to be tested and the smudged areas marked by the mask constitute a training sample. All training samples constitute a smudged dataset.

[0039] S52: Using the difference images corresponding to the book cover images of the test books in each training sample of the soiled dataset as the input of the Mask R-CNN network model, and the soiled areas of each book cover image marked by the mask as the theoretical output of the Mask R-CNN network model, a loss function is constructed using the theoretical output of the Mask R-CNN network model and the actual output of the Mask R-CNN network model under each training sample. It is then determined whether the loss function meets the requirements. If the result is yes, the final network model Mask R-CNN is obtained; if the result is no, backpropagation is performed using the loss function to adjust the model parameters of the Mask R-CNN network model. Each training sample in the soiled dataset is then used as the input of the Mask R-CNN network model after the model parameters have been adjusted, until the training reaches the set number of iterations or the result is yes.

[0040] Furthermore, an intelligent detection method for book cover damage based on image recognition also includes the following steps:

[0041] To capture the book covers, a 500*500 resolution camera and a fill light were added to the existing automatic borrowing and returning system.

[0042] Collect images of various book covers without damage and add them to the search database;

[0043] When a borrower returns a book, the automatic borrowing and returning system uses a camera to capture an image of the book cover and uses radio frequency identification technology to obtain basic information about the book. It then retrieves an image of the book's cover when it is undamaged from the database.

[0044] Beneficial effects:

[0045] 1. This invention provides an intelligent method for detecting book cover damage based on image recognition. It comprehensively utilizes various image processing technologies and intelligent methods to match and compare the acquired image of the book cover to be detected with an undamaged reference book cover image. This detects whether the book cover is damaged and identifies the location, size, and shape of the damage. This provides an efficient technical reference solution for damage identification in automated borrowing and returning systems. Therefore, this invention employs a non-contact image detection and analysis method to detect the presence, location, size, and shape of damage on book covers, thereby regulating borrower behavior and indirectly strengthening the protection of paper books.

[0046] 2. This invention provides an intelligent detection method for book cover damage based on image recognition. It uses gradient descent to correct the positions of the four corner points and obtains the perspective transformation matrix between the corrected positions of the four corner points and the four reference corner points, so as to minimize the mean square error between the transformed book cover image to be returned and the undamaged reference image, thereby improving the accuracy of intelligent detection of book cover damage.

[0047] 3. This invention provides an intelligent method for detecting book cover damage based on image recognition. It uses a machine learning-optimized feature matching algorithm for high-precision image matching and constructs a classification convolutional neural network by analyzing local contrast differences in the image for damage identification. After the automatic book lending system acquires the cover image of the book to be returned and retrieves its corresponding undamaged reference image, the algorithm can automatically detect whether the book cover is damaged and detect the location, size, and shape of the damage. It has extremely high detection efficiency for common damage forms such as stains, folds, and cracks. No manual supervision is required, saving manpower and improving the efficiency of book lending and returning management. It overcomes the shortcomings of traditional methods, such as low detection accuracy, easy damage to books, or incomplete detection areas, providing a feasible technical reference solution for damage identification in automatic book lending systems.

[0048] 4. This invention provides an intelligent detection method for book cover damage based on image recognition. It performs image matching based on feature matching and machine learning optimization, and also performs target detection based on local contrast difference features and convolutional neural networks. It can detect the location, size and shape of damage, and has extremely high detection efficiency for common damage forms such as stains, folds and cracks. No manual supervision is required, saving manpower and improving the efficiency of book borrowing and returning management. Attached Figure Description

[0049] Figure 1 A flowchart illustrating the overall process of soiling identification;

[0050] Figure 2 Flowchart for optimizing gradient descent matching accuracy. Detailed Implementation

[0051] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0052] It should be noted that, based on the existing automatic borrowing and returning system, a 500*500 resolution camera and a supplementary light are added to capture the book covers. In addition, the library needs to collect image information of various book covers when they are undamaged and add it to the retrieval database. When borrowers return books, the automatic borrowing and returning system uses the camera to capture the book cover image, obtains the basic information of the book through radio frequency identification technology, and retrieves the undamaged cover image of the book from the retrieval database.

[0053] After obtaining the cover image of the book at the time of return and an undamaged reference image, damage detection is performed based on an image recognition-based intelligent book cover damage detection method, such as... Figure 1 As shown, it includes the following steps:

[0054] S1: Using the SIFT feature matching method, perform image matching between the book cover image A to be returned and the undamaged reference cover image B, and obtain the four corner points on image A that correspond to the four reference corner points of image B. The specific steps include:

[0055] S11: Apply the classic SIFT feature detection method to perform scale space construction, local extremum detection, key point orientation calculation and key point descriptor calculation on image B and image A respectively, to obtain 128-dimensional key point descriptors corresponding to each pixel of the two images;

[0056] S12: Apply KNN feature matching to perform similarity matching between the key point descriptors of image B and image A in step S11, and obtain n pairs of feature point pairs with similar features; wherein, the present invention can use a KDTree search algorithm with a depth of 2 to improve the matching search efficiency;

[0057] S13: Using the RANSAC algorithm, homography matching is performed on the n pairs of feature points that were successfully matched in step S12. The maximum allowable reprojection error threshold is set to 5.0 to obtain the homography matrix between image B and image A. Based on the homography matrix, homography transformation is performed on the four selected reference corner points on image B to obtain four corner points on image A corresponding to the four reference corner points.

[0058] S2: Use gradient descent to correct the positions of the four corner points, and obtain the perspective transformation matrix between the corrected positions of the four corner points and the four reference corner points, so as to minimize the mean square error between the transformed book cover image to be returned and the undamaged reference image.

[0059] like Figure 2 As shown, the positions of the four corner points obtained by correcting them using the gradient descent method are as follows:

[0060] S21: The row and column information contained in the four corner points obtained on image A will be represented as an eight-dimensional vector x;

[0061] S22: Calculate the initial perspective transformation matrix between the four corner points of image A represented by the eight-dimensional vector x and the four reference corner points of image B, and perform perspective transformation on image A according to the initial transformation matrix. Use the mean square error between the perspective transformed image A and image B as the loss function, and the loss function is a function of the eight-dimensional vector x, denoted as L=f(x).

[0062] S23: Add a perturbation value d = 0.1 to each of the eight values ​​in the eight-dimensional vector x. Then, sequentially obtain the gradient components of the loss function corresponding to the new eight-dimensional vectors with the added perturbation values. Finally, integrate the eight obtained gradient components of the loss function into an eight-dimensional loss function gradient ΔL = [ΔL1 ΔL2 … ΔL8]. The method for obtaining the gradient components of the loss function corresponding to each new eight-dimensional vector is as follows:

[0063] Let the new eight-dimensional vector x be formed by adding a perturbation value d to the i-th value of the eight-dimensional vector x. Obtain the eight-dimensional vector according to step S22. Corresponding loss function components Then, the gradient components of the loss function are obtained using the first-order difference method.

[0064] S24: Construct the SGD-M optimizer based on the gradient ΔL of the loss function as follows:

[0065] x t+1 =x t -α·m t

[0066] m t =β·m t-1 +(1-β)·ΔL

[0067] Where, x t+1 Let x be an eight-dimensional vector in the (t+1)th iteration. t Let x be an eight-dimensional vector in the t-th iteration, and x t The initial value is the eight-dimensional vector x from step S21, α is the learning rate, α = 0.3, m t Let m be the auxiliary variable for the t-th iteration, β be the momentum coefficient, β = 0.9, and m t-1 Auxiliary variables for the (t-1)th iteration;

[0068] Each iteration of the SGD-M optimizer determines ||x t+1 -x tIf the result of either the condition is less than a set threshold or the number of iterations exceeds a set upper limit of 500, the final eight-dimensional vector is obtained, and the correction of the four corner points obtained on image A is completed. If the result of either condition is negative, the next iteration in the SGD-M optimizer is entered until the final eight-dimensional vector is obtained.

[0069] S3: Perform perspective transformation on image A according to the perspective transformation matrix to obtain image A*, which has a higher matching accuracy with image B and the same size;

[0070] S4: Obtain the difference image between image A* and image B, specifically:

[0071] S41: Obtain the local contrast similarity c(x, y) between image A* and image B in each color channel using the following method:

[0072] S41a: Calculate the local mean μ of image A*. x The local mean μ of image B y The diameter of the local circular window is 12.5 pixels;

[0073] S41b: Based on the local mean μ x Calculate the local weighted variance σ of image A*. x And based on the local mean μ y Calculate the local weighted variance σ of image B. y as follows:

[0074]

[0075]

[0076] Among them, w i (·) 2 Let x be a circularly symmetric Gaussian function with standard deviation σ = 1.5, i = 1, 2, ..., N, where N is the number of pixels in image A* or image B, and x i Let y be the pixel value of the i-th pixel in image A*. i Let be the pixel value of the i-th pixel in image B; it should be noted that this invention uses the local weighted variance σ. x and σ y Considered as local contrast of image A* and image B respectively

[0077] S41c: Based on the local weighted variance σ x and local weighted variance σ y Calculate the local contrast similarity c(x, y) between image A* and image B:

[0078]

[0079] Where C is a set constant, and the constant C = 0.03 is taken. The calculation result c(x, y) reflects the local difference between image A* and image B. The value range of c(x, y) is (0, 1]. The larger the value, the smaller the difference, and vice versa.

[0080] S42: Linearly map the local contrast similarity c(x, y) of image A* and image B in each color channel to the color code range [0, 255] to obtain an image with the same size as image A* and image B, which is the difference image between image A* and image B.

[0081] S5: Use the trained network model Mask R-CNN to extract the difference regions in the difference images, and realize the automatic detection of the location, size and shape of the dirt.

[0082] The training method for the Mask R-CNN network model built under the detectron2 framework is as follows:

[0083] S51: Prepare multiple book cover images to be tested, and obtain the difference images corresponding to each book cover image to be tested. At the same time, use image masks to mark the smudged areas of each book cover image to be tested. Among them, the difference image corresponding to a book cover image to be tested and the smudged areas marked by the mask constitute a training sample. All training samples constitute a smudged dataset.

[0084] S52: Using the difference images corresponding to the book cover images of the test books in each training sample of the soiled dataset as the input of the Mask R-CNN network model, and the soiled areas of each book cover image marked by the mask as the theoretical output of the Mask R-CNN network model, a loss function is constructed using the theoretical output of the Mask R-CNN network model and the actual output of the Mask R-CNN network model under each training sample. It is then determined whether the loss function meets the requirements. If the result is yes, the final network model Mask R-CNN is obtained; if the result is no, backpropagation is performed using the loss function to adjust the model parameters of the Mask R-CNN network model. Each training sample in the soiled dataset is then used as the input of the Mask R-CNN network model after the model parameters have been adjusted, until the training reaches the set number of iterations or the result is yes.

[0085] Therefore, this invention provides an intelligent book damage detection method based on image recognition technology, offering a feasible technical reference solution for damage identification in automated book lending and returning systems. Currently, there is no machine vision defect detection method based on book cover damage, and traditional methods either have low detection accuracy and poor performance, easily damage books, or have incomplete detection areas. The method of this invention uses a feature matching algorithm optimized by machine learning for high-precision image matching, and constructs a classification convolutional neural network by analyzing the local contrast difference features of the image for damage discrimination. After the automated lending and returning system obtains the book cover image at the time of return and retrieves the reference image when it is undamaged, the algorithm can automatically detect whether there is damage when the book is returned and detect the location, size, and shape of the damage. It has extremely high detection efficiency for common damage forms such as stains, folds, and cracks, without the need for manual supervision and judgment, saving manpower and improving the efficiency of book lending and returning management.

[0086] Of course, the present invention may have other various embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and modifications according to the present invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A method for intelligent detection of book cover damage based on image recognition, characterized in that, Includes the following steps: S1: Using the SIFT feature matching method, image matching is performed between the cover image A of the book to be returned and the undamaged reference cover image B, and four corner points corresponding to the four reference corner points of image B are obtained on image A. S2: Use gradient descent to correct the positions of the four corner points, and obtain the perspective transformation matrix between the corrected positions of the four corner points and the four reference corner points; The positions of the four corner points obtained by using gradient descent are specifically corrected as follows: S21: Represent the row and column information contained in the four corner points obtained from image A as an eight-dimensional vector. ; S22: Calculate an eight-dimensional vector The initial perspective transformation matrix is ​​given between the four corner points of image A and the four reference corner points of image B. Image A is then subjected to perspective transformation according to this initial transformation matrix. The mean square error between the perspective-transformed images A and B is used as the loss function, and the loss function is defined with respect to an eight-dimensional vector. The function, denoted as ; S23: Transform an eight-dimensional vector The eight values ​​are incremented by one perturbation value in turn. Then, the gradient components of the loss function corresponding to the new eight-dimensional vector with added perturbation values ​​are obtained sequentially. Finally, the eight gradient components of the loss function are integrated into an eight-dimensional loss function gradient. The method for obtaining the gradient components of the loss function corresponding to each new eight-dimensional vector is as follows: Eight-dimensional vector The new eight-dimensional vector formed by adding a perturbation value d to the i-th value is denoted as . Obtain the eight-dimensional vector according to step S22. Corresponding loss function components Then, the gradient components of the loss function are obtained using the first-order difference method. ; S24: Based on the gradient of the loss function The SGD-M optimizer is constructed as follows: in, Let be an eight-dimensional vector in the (t+1)th iteration. Let be an eight-dimensional vector in the t-th iteration, and The initial value is the eight-dimensional vector in step S21. , For learning rate, Let be the auxiliary variable for the t-th iteration. The momentum coefficient, Auxiliary variables for the (t-1)th iteration; Each iteration of the SGD-M optimizer is executed, and the following is determined: If the result of either judgment is yes, the final eight-dimensional vector is obtained, and the correction of the four corner points obtained on image A is completed; if the result of either judgment is no, the next iteration in the SGD-M optimizer is entered until the final eight-dimensional vector is obtained. S3: Perform a perspective transformation on image A according to the perspective transformation matrix to obtain an image A* that better matches image B; S4: Obtain the difference image between image A* and image B; S5: Use the trained network model Mask R-CNN to extract the difference regions in the difference images, and realize the automatic detection of the location, size and shape of the dirt.

2. The intelligent detection method for book cover damage based on image recognition as described in claim 1, characterized in that, Specifically, to obtain the four corner points on image A corresponding to the four reference corner points of image B, the following steps are taken: S11: Apply the classic SIFT feature detection method to perform scale space construction, local extremum detection, key point orientation calculation and key point descriptor calculation on image B and image A respectively, to obtain 128-dimensional key point descriptors corresponding to each pixel of the two images; S12: Apply KNN feature matching to perform similarity matching between the key point descriptors of image B and the key point descriptors of image A in step S11, and obtain n pairs of feature points with similar features; S13: Using the RANSAC algorithm, homography matching is performed on the n pairs of feature points that were successfully matched in step S12 to obtain the homography matrix between image B and image A. Based on the homography matrix, homography transformation is performed on the four selected reference corner points on image B to obtain four corner points on image A corresponding to the four reference corner points.

3. The intelligent detection method for book cover damage based on image recognition as described in claim 1, characterized in that, The specific steps to obtain the difference image between image A* and image B are as follows: S41: Obtain the local contrast similarity between image A* and image B in each color channel using the following method. : S41a: Calculate the local mean of image A* respectively. Local mean of image B ; S41b: Based on local mean Calculate the local weighted variance of image A* And based on local mean Calculate the local weighted variance of image B as follows: in, Standard deviation A circularly symmetric Gaussian function, i =1,2,…, N , N This represents the number of pixels in image A* or image B. For the first image A* i The pixel value of each pixel. For the first image B i The pixel value of each pixel; S41c: Based on local weighted variance and local weighted variance Calculate the local contrast similarity between image A* and image B. : in, To set a constant; S42: Compare the local contrast similarity of image A* and image B across all color channels. Linear mapping to the color code range [0,255] yields the difference image between image A* and image B.

4. The intelligent detection method for book cover damage based on image recognition as described in claim 1, characterized in that, The training method for the Mask R-CNN network model is as follows: S51: Prepare multiple book cover images to be tested, and obtain the difference images corresponding to each book cover image to be tested. At the same time, use image masks to mark the smudged areas of each book cover image to be tested. Among them, the difference image corresponding to a book cover image to be tested and the smudged areas marked by the mask constitute a training sample. All training samples constitute a smudged dataset. S52: Using the difference images corresponding to the book cover images of the test books in each training sample of the soiled dataset as the input of the Mask R-CNN network model, and the soiled areas of each book cover image marked by the mask as the theoretical output of the Mask R-CNN network model, a loss function is constructed using the theoretical output of the Mask R-CNN network model and the actual output of the Mask R-CNN network model under each training sample. It is then determined whether the loss function meets the requirements. If the result is yes, the final network model Mask R-CNN is obtained; if the result is no, backpropagation is performed using the loss function to adjust the model parameters of the Mask R-CNN network model. Each training sample in the soiled dataset is then used as the input of the Mask R-CNN network model after the model parameters have been adjusted, until the training reaches the set number of iterations or the result is yes.

5. A method for intelligent detection of book cover damage based on image recognition as described in any one of claims 1 to 4, characterized in that, It also includes the following steps: To capture the book covers, a 500*500 resolution camera and a fill light were added to the existing automatic borrowing and returning system. Collect images of various book covers without damage and add them to the search database; When a borrower returns a book, the automated borrowing and returning system uses a camera to capture an image of the book cover and uses radio frequency identification technology to obtain basic information about the book. It then retrieves an image of the book's cover when it is undamaged from the database.

Images