Intelligent blank label image recognition and composite coding system

Abstract

The invention discloses an intelligent blank label image recognition and composite coding system, and relates to the technical field of image processing. The system comprises a label image feature processing part, an embedded code part and a composite code generation part, the label image feature processing part is used for obtaining multi-dimensional features of a target label image; curvature flow feature analysis is carried out, and local curvature distribution is calculated; calculating a geometric invariant moment of the target label image based on the local curvature distribution; the embedded coding part is used for extracting continuous homology-based topological characteristics of a target label image; according to the topological features, feature popular codes are constructed by using the Gaussian-Blog theorem; and the composite code generation part is used for carrying out unique coding based on differential topology by adopting a Hessian matrix according to a feature popular code and a fusion result. According to the method, the geometric and topological characteristics of the label image can be accurately identified, and a code with uniqueness and robustness is generated.

Description

Technical Field

[0001] The present disclosure relates to, but is not limited to, the field of image processing technology, and in particular to an intelligent blank label image recognition and composite coding system. Background Art

[0002] Label image recognition and coding technology has a wide range of applications in industrial production, logistics management, product tracking and other fields. As a product identification carrier, the efficient recognition and information coding of labels directly affects the operating efficiency and management accuracy of the entire supply chain. However, the existing technology still has many problems in label image recognition and coding, such as insufficient robustness in complex scenes, low uniqueness and discrimination of coding methods, and limited adaptability to geometric deformation and topological changes.

[0003] At present, label image recognition mainly relies on the following technologies: Template matching is a more traditional label image recognition technology. Its basic principle is to compare the pre-stored label template with the target image pixel by pixel to determine whether it matches. The advantage of this method is that it is simple to implement and is suitable for identifying label images with strong regularity. However, when the shape of the label image is complex or there is rotation, scaling, and noise interference, the performance of template matching is significantly reduced. In addition, template matching has high requirements for computing resources, especially in large-scale label image processing scenarios. It is difficult to meet the real-time requirements. In recent years, feature point extraction algorithms such as SIFT (Scale Invariant Feature Transform), SURF (Speeded Up Robust Features) and ORB (Orbital Robust Features) have been widely used in label image recognition. These methods complete feature matching by extracting key points and descriptors of label images. Compared with template matching, feature point extraction methods have better robustness to image rotation, scaling and illumination changes. However, these methods mainly focus on the local features of the image, and it is difficult to fully describe the global geometric characteristics and topological structure of the label image. Especially when processing label images with complex morphology, the matching accuracy is low. In recent years, deep learning technology has made great progress in the field of image recognition, and convolutional neural networks (CNNs) have been widely used in label image classification and recognition tasks. By training a large amount of label data, neural networks can extract high-level image features and achieve efficient recognition of label images. However, the performance of deep learning methods is heavily dependent on the quality and scale of training data, and has high requirements for computing resources. In addition, existing deep learning methods focus more on image classification tasks, and have limited ability to extract the geometric and topological characteristics of label images, making it difficult to meet the uniqueness and robustness requirements of label image encoding. Summary of the invention

[0004] The present invention provides an intelligent blank label image recognition and composite coding system, which realizes multi-dimensional feature extraction, topological feature calculation, feature popular coding generation and unique coding based on differential topology of label images. The present invention can accurately identify the geometric and topological characteristics of label images, generate unique and robust coding, adapt to morphological changes, light interference and noise influence in complex scenes, and has efficient processing capabilities and extremely high discrimination.

[0005] In order to solve the above problems, the technical solution of the present invention is achieved as follows:

[0006] An intelligent blank label image recognition and composite coding system, the system comprises: a label image feature processing part, an embedded coding part and a composite coding generation part; the label image feature processing part is used to extract multi-dimensional features of a target label image to obtain multi-dimensional features of the target label image; based on the multi-dimensional features, a curvature flow feature analysis is performed to calculate the local curvature distribution for the boundary curve of the target label image; based on the local curvature distribution, the geometric invariant moment of the target label image is calculated; the embedded coding part is used to extract the topological features of the target label image based on continuous homology based on the geometric invariant moment; according to the topological features, the feature manifold coding is constructed using the Gauss-Bonnet theorem; the composite coding generation part is used to perform multi-scale fusion of scale space theory using the Hessian matrix according to the feature manifold coding to obtain a fusion result; according to the fusion result, a unique coding based on differential topology is performed.

[0007] Furthermore, before the label image feature processing part performs multi-dimensional feature extraction on the target label image to obtain the multi-dimensional features of the target label image, it will also determine whether the target label image is a blank image. If so, it will perform multi-dimensional feature extraction on the target label image to obtain the multi-dimensional features of the target label image. If the target label image is not a blank image, a prompt will be issued to terminate the processing process; the label image feature processing part uses a method based on pixel value checking to check whether the target label image is a blank image, specifically including: converting the target label image into a grayscale image, going through all pixel values ​​of the grayscale image, checking whether all pixel values ​​are the same, and if so, judging that the target label image is a blank image.

[0008] Furthermore, the multidimensional feature Φ(x, y) is expressed using the following formula:

[0009]

[0010] Wherein, x represents the X-axis coordinate; y represents the Y-axis coordinate; I represents the target label image; Represents the gradient of the target label image I; represents the curl operator, Ω represents the entire image domain of the target labeled image I.

[0011] Furthermore, the local curvature distribution K is calculated by the following formula:

[0012]

[0013] Where, κ is the local curvature value; C is the boundary curve of the target label image I; is the local curvature formula of a two-dimensional curve, (x′, y′) represents the tangent vector of the boundary curve, (x″, y″) represents the rate of change of the curvature of the boundary curve; x′ is the first derivative of x; y′ is the first derivative of y; x″ is the second derivative of x; y″ is the second derivative of y; ds is the micro-arc element of the curve, indicating the small step size in the integration process; σ is the standard deviation of Gaussian smoothing.

[0014] Furthermore, the geometric invariant moment M of the target label image is calculated by the following formula: pqr :

[0015]

[0016] Wherein, V is a closed volume obtained by expanding the boundary curve of the target label image, which is used as the characteristic volume of the target label image; is the X-axis coordinate of the centroid of the feature volume; is the Y-axis coordinate of the centroid of the feature volume; is the Z-axis coordinate of the centroid of the characteristic volume; p is the order of the moment of the X-axis; q is the order of the moment of the Y-axis; r is the order of the moment of the Z-axis; is the divergence operator of the local curvature distribution K; H is the average curvature of the closed volume; K G is the Gaussian curvature; is the gradient of the local curvature distribution K; dV is the infinitesimal element of the volume integral.

[0017] Furthermore, the topological features are calculated by the following formula:

[0018]

[0019] in, is a topological feature, an n-dimensional homology group, describing the topological space The n-th dimensional topological property of ; the value of n is 0, 1 or 2; is the number of independent connected parts in the target label image; is the number of holes in the target label image; is the number of cavities on the surface of the target labeled image; ker is the kernel space operator; im is the image space operator; Lap is the Laplace operator; is a one-dimensional boundary operator; is a two-dimensional boundary operator; is the integral differential element of the topological space; eig(Δ) is the eigenvalue sequence of the Laplace operator;

[0020] Furthermore, the following formula is used to construct the feature popular coding E based on the topological features using the Gauss-Bonet theorem:

[0021]

[0022] Furthermore, the following formula is used to adopt the Hessian matrix to perform multi-scale fusion of scale space theory and obtain the fusion result:

[0023]

[0024] Among them, F is the fusion result.

[0025] Furthermore, the following formula is used to generate a unique code based on differential topology:

[0026] C=hash(∮ C K∧ds·F·exp(∫ x Ric(H1(x)dx)));

[0027] Among them, hash(·) is the hash mapping function; Ric(·) means calculating the Ricci curvature; ∧ is the exterior integral operator.

[0028] The intelligent blank label image recognition and composite coding system of the present invention has the following beneficial effects: the present invention realizes accurate recognition of complex label images through a unique design combining multidimensional feature extraction, topological characteristic analysis, and Gauss-Bonnet theorem with differential topology. Complex label images may have porous structures, irregular shapes, or deform due to environmental changes (such as illumination, rotation, and scaling). Traditional methods are difficult to accurately extract the key characteristics of such label images, while the present invention adopts a method of multidimensional feature extraction and curvature distribution calculation, which can capture the key characteristics of local details and global morphology in label images, thereby showing high robustness in different scenarios. For example, geometric characteristic analysis based on local curvature distribution can not only accurately depict subtle changes in label boundaries, but also effectively eliminate the influence of noise by combining Gaussian smoothing and curvature gradient filtering. For deformations such as rotation or scaling, the present invention introduces geometric invariant moments to ensure the stability and consistency of extracted features, significantly enhancing the adaptability of the system to complex label images. In the process of label image feature extraction, the present invention deeply integrates geometry and topology for the first time. Geometric characteristics focus on the shape, boundary, texture and other features of the label image, while topological characteristics focus on the connectivity, hole distribution and spatial structure of the label image. By introducing the Gauss-Bonnet theorem and homology group theory, the present invention can simultaneously extract the geometric and topological characteristics of the label image, making the system's description of the label image more comprehensive and accurate. The present invention achieves efficient processing capabilities while improving recognition accuracy and coding quality. First, the feature extraction part adopts a multi-dimensional feature combination method to avoid redundant calculations, and at the same time reduces the complexity of subsequent processing through the compressed expression of feature popular coding. Secondly, the hash mapping function used in the system can quickly map high-dimensional features into fixed-length codes, greatly improving the storage and matching efficiency of label images. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A schematic diagram of the system structure of an intelligent blank label image recognition and composite encoding system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0030] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present disclosure clearer and more understandable, the present disclosure is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present disclosure and are not used to limit the present disclosure.

[0031] Example 1, reference Figure 1:An intelligent blank label image recognition and composite coding system, the system includes: a label image feature processing part, an embedded coding part and a composite coding generation part; the label image feature processing part is used to extract multi-dimensional features of the target label image to obtain the multi-dimensional features of the target label image; based on the multi-dimensional features, a curvature flow feature analysis is performed, and the local curvature distribution is calculated for the boundary curve of the target label image; based on the local curvature distribution, the geometric invariant moment of the target label image is calculated; the embedded coding part is used to extract the topological features based on the continuous homology of the target label image based on the geometric invariant moment; according to the topological features, the Gauss-Bonnet theorem is used to construct the feature popular coding; the composite coding generation part is used to use the Hessian matrix to perform multi-scale fusion of scale space theory according to the feature popular coding to obtain the fusion result; according to the fusion result, a unique coding based on differential topology is performed.

[0032] Specifically, the target label image is usually a two-dimensional graphic with various shapes and complex details, and its features are highly important in the label application scenario, such as the curve shape of the label edge, the local curvature and the overall structural layout. The present invention obtains information from the texture, shape and spatial distribution characteristics of the image through a specific multidimensional feature extraction method to ensure that the extracted features are both stable and adaptable to a variety of complex scenarios. In the specific implementation, the system accurately captures the boundary features of the label image by introducing curvature flow feature analysis. Curvature flow analysis is a method that uses the local geometric information of the image boundary points to characterize the shape of the curve. Its core lies in characterizing the curvature of the boundary points by calculating the curvature. The size of the boundary curvature reflects the local morphology of the curve, such as sharp corners, smooth transitions and other features, which provides a basis for accurately extracting the image shape. Furthermore, in order to overcome the sensitivity of traditional methods to transformations such as image scaling, rotation and flipping, the present invention realizes the global expression of features through the calculation of geometric invariant moments. Geometric invariant moments are a mathematical method that can capture the global morphological features of an image. Its core lies in using the distribution information of image pixels to generate feature description quantities that are invariant to different morphological changes. In the present invention, the introduction of geometric invariant moments greatly enhances the system's ability to recognize label images, especially in a variety of scenes and environmental changes, and can ensure the robustness and consistency of extracted features. Combined with curvature flow analysis, geometric invariant moments further improve the expression of the morphological features of label images, making the function of this part far beyond the existing traditional image feature extraction technology. The present invention captures the morphological details of the label image from a local perspective through the curvature distribution analysis of the boundary curve; then, these local information is integrated into global features through geometric invariant moments to describe the shape characteristics of the label image from an overall perspective. Such a hierarchical processing method can effectively avoid the problem of feature loss caused by the imbalance of global and local information in the traditional feature extraction process. It is worth noting that curvature flow and geometric invariant moments complement each other in the processing process. The local accuracy provided by curvature flow is combined with the global stability of geometric invariant moments, so that the system can more comprehensively capture the diverse features of label images. Traditional methods usually rely on simple edge detection or shape matching algorithms, which have limited processing capabilities for complex images, especially when there are morphological variations or background interference in the label image, the problem of feature extraction failure often occurs. The present invention introduces curvature flow feature analysis and geometric invariant moments to accurately capture local details of the label image while ensuring the integrity of global features. This dual feature extraction mechanism ensures that the system can maintain a high level of recognition capability and stability when facing complex images and dynamically changing scenes.

[0033] The embedded coding part is based on the geometric characteristics of the label image, deeply mines the topological structure information, and transforms the image's shape, spatial distribution and other features into a high-level topological representation in a mathematical form through the processing of geometric invariant moments, and generates feature popular codes. Specifically, the design of the embedded coding part integrates two highly abstract mathematical tools, continuous homology and Gauss-Bonnet theorem, aiming to extract universal and robust topological characteristics from image features, and fix these characteristics in the form of popular codes, providing an indispensable input for subsequent composite code generation. Geometric invariant moments are the basis of the embedded coding part. The system first extracts stable features related to the morphology of the label image from geometric invariant moments. In label image recognition, shape features often deviate due to shooting angles, changes in illumination or image noise, and geometric invariant moments have a natural resistance to these influences. Therefore, it can express the shape characteristics of the label image in a way that does not depend on specific transformations. However, geometric invariant moments themselves are only a static representation of geometric features, lacking a comprehensive capture of spatial distribution and global structure. The present invention further mines the high-order information contained in geometric invariant moments through continuous homology, thereby realizing dynamic capture of image structural characteristics. Persistent homology is a powerful topological tool used to analyze the topological structure of images at different scales. For labeled images, their boundary curves, holes, and the correlation between shapes constitute the unique topological characteristics of the image. In the embedded coding part, the system uses persistent homology as a tool to observe the topological changes of labeled images at different scales and capture its features that are not fully expressed in geometric invariant moments. For example, a labeled image may contain some small curves or sharp corners, which are significant at the local scale but may be ignored at the global scale. Persistent homology records the duration of these topological features during the change process by constructing persistent barcodes at different scales, thereby expressing the topological information of the labeled image in a more comprehensive and dynamic way. This topological feature can not only describe the shape of the labeled image, but also reflect its structural properties that remain unchanged in different environments, and has extremely high recognition value.

[0034] The Gauss-Bonet theorem plays a key role in the embedded coding part, which converts the topological characteristics extracted by continuous homology into a mathematical quantitative expression, thereby forming a feature popular coding that can be used for further coding. The Gauss-Bonet theorem essentially establishes a connection between the local geometric characteristics of the surface (such as curvature) and the global topological characteristics (such as the Euler characteristic). In the application of label images, this means that we can integrate this information into a global topological expression by analyzing the local curvature of the boundary curve. Specifically, the embedded coding part calculates the local curvature distribution of the label image and combines the Gauss-Bonet theorem to generate a set of feature popular coding that expresses the overall morphological characteristics of the image. The coding not only contains the geometric characteristics of the label image, but also fully integrates its topological characteristics, so it has a high degree of distinction and uniqueness. The design of the embedded coding part cleverly combines the advantages of geometry and topology, unifies the static characteristics and dynamic characteristics of the label image, and thus generates a highly recognizable feature popular coding. This coding form plays an irreplaceable and important role in the intelligent blank label image recognition and composite coding system of the present invention. On the one hand, feature popular coding expresses the key characteristics of label images in mathematical form, laying a solid foundation for the subsequent composite coding generation; on the other hand, this coding is highly compressible and can retain the core information of the image while greatly reducing the computational complexity, thereby improving the overall efficiency of the system. Compared with traditional image coding methods, the embedded coding part realizes cross-domain information integration from geometry to topology through a deep combination of continuous homology and Gauss-Bonnet theorem. Traditional methods usually rely on simple feature point detection or shape matching technology, which is difficult to effectively capture the global characteristics and local details of the image. The present invention introduces a high-level topological method to abstract the geometric characteristics of the label image into a higher-dimensional topological expression, which fundamentally improves the system's recognition ability and robustness for complex label images. Especially when the label image is complex in morphology or the background interference is severe, the feature popular coding generated by the embedded coding part can effectively resist external interference and ensure the stability and accuracy of the system.

[0035] The composite coding generation part first performs multi-scale processing on the feature popular coding output by the embedded coding part, and this process is centered on the scale space theory. The introduction of scale space theory is to ensure that the feature performance of the label image at different observation scales is consistent, because in actual scenes, the shooting distance, angle, and lighting conditions of the label image may be different, which leads to changes in the distribution of features at different scales. By placing the feature popular coding in the scale space, the system can simultaneously analyze its characteristics at different scales and synthesize these characteristics to generate a global, multi-scale feature expression. This method essentially extracts the global characteristics of the coding at a large scale and the local detail characteristics at a small scale by combining blurring and sharpening, so that the generated composite coding retains the global morphological characteristics of the label image and captures the important local details. In order to achieve this goal, the composite coding generation part uses the Hessian matrix as a mathematical tool to capture the changing characteristics of the feature popular coding in the scale space. The Hessian matrix is ​​a second-order partial derivative matrix, and its function is to reveal the local curvature and changing trend of the target function. In the present invention, the Hessian matrix is ​​used to analyze the gradient change of the feature popular coding, and the fuzzy kernel of the scale space is combined to perform multi-scale fusion of the features. Through this process, the system is able to generate a set of representative feature descriptions at different scales, ensuring that the generated codes are robust and unique regardless of the actual shooting conditions of the labeled images. Compared with the traditional single-scale method, multi-scale fusion not only enhances the expressive power of the codes, but also significantly improves the system's recognition accuracy for complex labeled images.

[0036] After completing the multi-scale fusion, the composite coding generation part further introduces the differential topology method to ensure the uniqueness of the final coding. Differential topology is a mathematical tool for studying the continuous change characteristics of space. Its core lies in revealing the global topological structure by analyzing the local differential properties of space. In the present invention, differential topology is used to convert the multi-scale fused features into a unique coding expression. The key to this process is that the system generates a set of mathematical descriptions with consistency and discrimination by analyzing the changing trend of feature popular coding at different differential scales. These descriptions are further topologically integrated and finally fixed as the unique coding of the labeled image. This coding can not only accurately characterize the characteristics of the labeled image, but also has a strong anti-interference ability, and can maintain its uniqueness and stability even in complex scenes. Traditional image coding methods usually focus on the simple extraction of geometric characteristics, lack of in-depth analysis of the global structure of the image, and thus often show limitations when processing complex labeled images. The present invention introduces the differential topology method in the composite coding generation part to unify the geometric characteristics and topological characteristics into a mathematical framework, which not only ensures the global consistency of the coding, but also retains the accurate expression of local details. Especially when the label image has complex morphology and diverse features, this method can fully capture the essential characteristics of the image, thereby providing important support for the efficient recognition of the system. In addition, the unique code generated by the composite coding generation part has extremely high compression efficiency and expression ability. This code can not only accurately express the core characteristics of the label image in a limited storage space, but also realize rapid retrieval and matching in a large-scale label image database. This feature makes the intelligent blank label image recognition and composite coding system of the present invention have great advantages in practical applications, especially in scenarios where massive label image data needs to be processed, its efficiency and accuracy far exceed the existing technology. Through the design of the composite coding generation part, the present invention successfully integrates the geometric, topological and multi-scale characteristics of the label image into a unique and robust coding expression, which significantly improves the overall performance of the system.

[0037] Embodiment 2: before the label image feature processing part performs multi-dimensional feature extraction on the target label image and obtains the multi-dimensional features of the target label image, it will also determine whether the target label image is a blank image. If so, it performs multi-dimensional feature extraction on the target label image to obtain the multi-dimensional features of the target label image. If the target label image is not a blank image, a prompt is issued to terminate the processing process; the label image feature processing part uses a method based on pixel value checking to check whether the target label image is a blank image, specifically including: converting the target label image into a grayscale image, going through all pixel values ​​of the grayscale image, checking whether all pixel values ​​are the same, and if so, judging that the target label image is a blank image.

[0038] Specifically, grayscale image conversion simplifies each pixel value of a color image from multiple channels (such as RGB three channels) to a single-channel grayscale value, thereby greatly reducing data complexity. This not only speeds up the pixel value check, but also effectively reduces the computational burden of the system. Secondly, pixel value traversal, as the core step of the detection method, directly checks each pixel of the grayscale image to confirm whether it is consistent with other pixel values. Since all pixel values ​​of a blank image are theoretically exactly the same, this method can complete the detection quickly and accurately, avoiding the additional overhead that may be caused by complex image analysis algorithms.

[0039] Embodiment 3: The multidimensional feature Φ(x, y) is expressed using the following formula:

[0040]

[0041] Wherein, x represents the X-axis coordinate; y represents the Y-axis coordinate; I represents the target label image; Represents the gradient of the target label image I; represents the curl operator, Ω represents the entire image domain of the target labeled image I.

[0042] Specifically, in this formula, the expression of the multidimensional feature Φ(x, y) depends on the high-order variation pattern of the target image I(x, y). First, the basic composition of Φ(x, y) is the fourth-order partial derivative of the target image: This part captures the high-frequency variation patterns of the image in the spatial dimension. Compared with the first-order or second-order derivatives, the fourth-order partial derivatives are significantly more sensitive to boundaries, textures, and complex local details. This high-order variation directly reflects the detailed information in the labeled image, such as subtle edge features, complex curved shapes, and cross structures. This is crucial for feature extraction of labeled images, especially in complex scenes, and can effectively avoid the problem of insufficient feature expression caused by detail loss in low-order methods. At the same time, the Hessian matrix determinant det(H) introduced in the formula provides local curvature information of the labeled image. The Hessian matrix consists of the second-order partial derivatives of the image, and its determinant represents the curvature distribution characteristics of the image at a certain point. In the labeled image, the curvature of different regions reflects the local changes in the image structure, such as the distribution of convex, concave, or flat areas. By multiplying the Hessian matrix determinant with the fourth-order partial derivative, the formula not only strengthens the expression of local characteristics of the image, but also combines this information in a unique mathematical form, thereby more comprehensively describing the geometric characteristics of the labeled image. This combination overcomes the separation problem of local features and global features in traditional methods and provides a unified framework for label image feature extraction. Furthermore, the curl operator in this formula The spatial structural information of the labeled image is supplemented from another dimension. The curl is an operator that reflects the rotational properties of the image gradient field. For a two-dimensional image, its value directly represents the degree of asymmetry of the gradient in the image. Specifically, when the curl is zero, the gradient field is irrotational, indicating that the intensity change of the image is linear; when the curl is non-zero, it indicates that the image has a complex nonlinear change pattern, such as a cross boundary or a subtle texture structure. By introducing the curl operator into the formula, the feature extraction method of the present invention can capture the complex structural characteristics of the labeled image that are difficult to find by traditional methods. This design is particularly suitable for processing complex labeled images, such as multi-boundary images or labels with overlapping structures. In addition, the overall feature expression of the formula is completed over the entire image domain Ω through an integral operation. This integral process is of great significance. It not only eliminates the interference of local noise, but also unifies local features into a global feature expression. In practical applications, labeled images are often affected by noise, illumination changes or other interference factors, and the integral operation can effectively smooth these interferences, thereby ensuring the robustness and stability of feature extraction. At the same time, the integral further enhances the system's ability to capture global characteristics, so that the generated feature expression can accurately reflect the overall information structure of the label image. This feature extraction method from a global perspective plays an important role in the efficient recognition of label images, especially in the processing scenario of large-scale label image databases, which significantly improves the accuracy and speed of feature matching. Compared with traditional low-order feature methods, this formula has significant advantages in capturing the complex characteristics of label images. On the one hand, through the joint operation of high-order partial derivatives, Hessian matrix determinants and curl, the formula can extract the global and local characteristics of the label image, avoiding the limitations of the single feature method. On the other hand, the integral operation in the formula effectively integrates these characteristics and generates a unified feature expression. This multi-dimensional and multi-level design ensures that the system fully expresses the characteristics of the label image and provides high-quality input for subsequent embedded coding and composite coding generation. From the perspective of practical application, this formula is particularly suitable for the processing of complex label images. In scenarios such as industrial production and logistics management, the types and qualities of label images may vary due to environmental changes, and the design of this formula can stably extract the core features of label images under various conditions. Whether processing images with complex lighting conditions or needing to identify highly detailed labeled images, this formula can provide reliable feature expression, thus ensuring the efficiency and stability of the system.

[0043] Embodiment 4: Calculate the local curvature distribution K by the following formula:

[0044]

[0045] Where, κ is the local curvature value; C is the boundary curve of the target label image I; is the local curvature formula of a two-dimensional curve, (x′, y′) represents the tangent vector of the boundary curve, (x″, y″) represents the rate of change of the curvature of the boundary curve; x′ is the first derivative of x; y′ is the first derivative of y; x″ is the second derivative of x; y″ is the second derivative of y; ds is the micro-arc element of the curve, indicating the small step size in the integration process; σ is the standard deviation of Gaussian smoothing.

[0046] Specifically, the core part of the formula is the calculation of local curvature, which is in the form of This part directly reflects the local geometric characteristics of the boundary curve of the target label image and describes the degree of curvature of the curve at a specific point. The calculation of local curvature depends on the tangent vector (x′, y′) and curvature change rate (x″, y″) of the curve, where the tangent vector describes the direction of the boundary curve and the curvature change rate reflects the curvature change of the curve at that point. By integrating this information into the curvature formula, the system can accurately capture the shape characteristics of the boundary of the label image, such as the sharpness of the edge, the size of the smooth transition area, and the change pattern of the corner. This information is of great significance for feature extraction of label images, because the boundary shape of the label is often an important basis for its identification and classification. Based on the curvature calculation, the formula further introduces the characteristic function Φ(x, y) and the Gaussian smoothing factor The feature function Φ(x, y) is a multidimensional feature expression, which comes from the output of the label image feature extraction part and contains the global and local information of the target label image. Introducing Φ(x, y) in the curvature calculation can combine the geometric characteristics of the boundary with the overall characteristics of the image, so that the local curvature distribution not only reflects the shape characteristics of the boundary, but also combines the global information such as the texture and spatial distribution of the image to generate a more comprehensive feature description. Gaussian smoothing factor It plays a key weighting role. The change in curvature may be affected by noise or other interference in some cases, resulting in errors in feature extraction. In order to reduce this interference, the formula adjusts the curvature value through Gaussian weighted smoothing, where σ represents the standard deviation of the Gaussian distribution. Specifically, larger curvature values ​​(such as sharp corners or steep boundaries) are given smaller weights under Gaussian weighting, which effectively reduces the impact of noise on the calculation results. Smaller curvature values ​​(such as smooth transition areas) receive higher weights, thereby enhancing the contribution of smooth areas in feature description. In this way, the formula achieves smoothing of local features and improves the robustness and stability of the system. The path integral ∮ contained in the formula CThe whole calculation process covers the complete boundary of the target label image. This integral process accumulates each tiny point on the boundary through the curve micro-arc element ds, thereby generating a global curvature distribution expression. The introduction of integral can not only integrate the local information of the label image boundary, but also eliminate the influence of local noise, and integrate the complex boundary characteristics into a global feature. This globality makes the curvature distribution K highly general and representative for the overall shape description of the label image, providing an extremely important input for the subsequent embedded coding and composite coding generation. Traditional boundary feature extraction methods usually rely only on simple low-order information (such as edge detection or local gradient), lack the description of high-order geometric characteristics, and are easily affected by noise and deformation. However, this formula can not only extract the detailed characteristics of the boundary, but also ensure the stability and robustness of the feature description through high-order curvature calculation, combination of feature functions and Gaussian weighted smoothing. In addition, the formula elevates local features to the level of global features through integral operations, so that the generated curvature distribution can fully reflect the boundary morphology and overall structure of the label image. In practical applications, this formula is particularly suitable for scenes with complex label image morphology and significant boundary features. For example, in industrial production, label images may have a variety of shapes and boundary complexities, and this formula can accurately extract these features, thereby providing reliable data support for label classification, detection, and tracking. At the same time, because the design of the formula fully considers robustness, the curvature distribution it generates has good resistance to environmental factors such as lighting changes and noise interference, thus ensuring the stability and efficiency of the system in complex scenarios.

[0047] Example 5: Calculate the geometric invariant moment M of the target label image using the following formula: pqr :

[0048]

[0049] Wherein, V is a closed volume obtained by expanding the boundary curve of the target label image, which is used as the characteristic volume of the target label image; is the X-axis coordinate of the centroid of the feature volume; is the Y-axis coordinate of the centroid of the feature volume; is the Z-axis coordinate of the centroid of the characteristic volume; p is the order of the moment of the X-axis; q is the order of the moment of the Y-axis; r is the order of the moment of the Z-axis; is the divergence operator of the local curvature distribution K; H is the average curvature of the closed volume; K G is the Gaussian curvature; is the gradient of the local curvature distribution K; dV is the infinitesimal element of the volume integral.

[0050] Specifically, the geometric invariant moment M in the formula pqris the integral result over the closed volume V. Volume V is a three-dimensional feature volume generated by expanding the boundary curve of the target label image, which reflects the global morphological characteristics of the target label image. This three-dimensional extension elevates the label image from a two-dimensional boundary curve to a three-dimensional space expression, so that the geometric invariant moment can capture more complex structural information of the image in space. In this way, the formula not only describes the planar geometric characteristics of the label image, but also effectively combines its spatial morphological characteristics, making the feature extraction have a higher dimension and stronger description ability. One of the core parts of the formula is about The weighted form of this term is obtained by calculating the centroid coordinates of the image feature volume to measure the offset of a specific point relative to the centroid, and adjust its influence by the order of p, q, and r. The centroid of the feature volume is the geometric center of the target image, reflecting its global symmetry and mass distribution. When the integral is expanded to this weighted form, the formula can capture the offset characteristics of the target image of different orders in space, and then fully describe its three-dimensional geometric characteristics. This weighted integral makes the geometric invariant moment invariant to the rotation, translation, and scaling of the labeled image, thereby greatly improving the robustness of the system. On this basis, the formula introduces the local curvature distribution K and its related operators, including the gradient and divergence The local curvature distribution K is derived from the aforementioned boundary curvature calculation, reflecting the curvature characteristics of the target label image boundary. By combining K in the volume integral, the formula effectively transfers the geometric information of the boundary to the global feature expression. In addition, represents the gradient of the curvature distribution, revealing the changing trend of K in three-dimensional space, while As its divergence operator, it further describes the diffusion characteristics of the curvature distribution in the feature volume. The combination of these operators enables the geometric invariant moment to reflect the directionality and diffusion of the local geometric changes of the target label image, thereby enhancing the delicateness of the feature description.

[0051] The formula also incorporates higher-order curvature information, including mean curvature H and Gaussian curvature K G The mean curvature H is an important feature to measure the smoothness of a surface, which represents the local shape characteristics of the surface at a certain point. By introducing H, the geometric invariant moment can capture the smoothness of the boundary of the target label image during the three-dimensional expansion process and its contribution to the overall shape. At the same time, the Gaussian curvature K G It reflects the overall curvature of the surface at a certain point and is a global description of curvature. Introducing H in volume integral 2 ·K G, making the geometric invariant moment more sensitive to the global morphological features of the labeled image. The combination of this high-order curvature information enables the formula to describe the geometric characteristics of the target labeled image on a multi-dimensional scale, providing a more comprehensive feature expression for subsequent embedded coding generation. Another key design of the formula is the weighted smoothing of the local curvature gradient, using the Gaussian weighting function This part reduces the feature extraction error caused by noise or complexity of boundary details by adjusting the weight of the gradient value. The large values ​​(corresponding to the sharp changes in the curvature gradient) will be suppressed by Gaussian weighting, while the smaller gradient values ​​retain more information. This smoothing process effectively improves the robustness of the geometric invariant moment, so that it can still generate stable and reliable feature expressions when processing complex or disturbed label images. Through volume integration, the formula effectively integrates the above geometric information, curvature characteristics and weighted smoothing, and finally generates the geometric invariant moment M. pqr This result is not only a comprehensive feature description of the target label image in the spatial dimension, but also invariant to rotation, translation and scale changes, ensuring feature consistency in different scenarios. Compared with the traditional two-dimensional image moment, this formula is extended to three-dimensional volume, and by combining high-order geometric information and global characteristics, it significantly enhances the feature expression ability. From the perspective of practical application, the geometric invariant moment M pqr It is particularly suitable for processing label images with complex and diverse morphologies. For example, when it is necessary to distinguish highly similar labels or images with complex boundary structures, geometric invariant moments can provide powerful recognition capabilities. At the same time, because the formula is highly robust to environmental noise and morphological changes, it has significant advantages in label image recognition tasks in scenarios such as industrial production and logistics management.

[0052] Example 6: The topological characteristics are calculated by the following formula:

[0053]

[0054] in, is a topological feature, an n-dimensional homology group, describing the topological space The n-th dimensional topological property of ; the value of n is 0, 1 or 2; is the number of independent connected parts in the target label image; is the number of holes in the target label image; is the number of cavities on the surface of the target labeled image; ker is the kernel space operator; im is the image space operator; Lap is the Laplace operator; is a one-dimensional boundary operator; is a two-dimensional boundary operator; is the integral differential element of the topological space; eig(Δ) is the eigenvalue sequence of the Laplace operator.

[0055] Specifically, the topological feature H in the formula n (X) is the homology group of the target label image in n-dimensional space. Homology group H in algebraic topology n (X) is a mathematical tool used to describe the structural characteristics of topological space, where H0(X) represents the number of independent connected parts, reflecting the distribution of discrete regions in the target label image; H1(X) represents the number of holes in the image, such as ring structures or blank areas with closed boundaries; and H2(X) represents the number of cavities on the image surface, which is particularly important when the label image has a complex three-dimensional structure. This hierarchical characteristic expression enables the formula to comprehensively and accurately characterize the topological structure of the target label image. The first part of the formula is the core expression of the homology group, where ker (kernel space) and im (image space) represent the kernel and image of the boundary operator respectively. Specifically, and are n-dimensional and (n+1)-dimensional boundary operators, which are used to describe the boundary relationship of the target labeled image in different dimensions respectively. By calculating ker and im, the formula can extract the topological invariants of the image in n-dimensional space, such as the number of connected parts, the formation of holes, and the distribution of surface cavities. These topological properties are different from traditional geometric properties. They are independent of scale, rotation and deformation, so they are particularly suitable for robust feature extraction of labeled images. In addition, the second part of the formula combines the geometric information of the image with the topological structure through integral operations, further enhancing the expressive power of the features. The core components in the integral term include: Δ(Φ(x, y)), which represents the Laplace operator of the characteristic function Φ(x, y); eig(Δ), which is the eigenvalue sequence of the Laplace operator; and and K G , respectively represent the Gaussian smoothing factor and Gaussian curvature. The role of the Laplace operator Δ(Φ(x, y)) is to capture the local variation characteristics of the target label image feature function Φ(x, y) through the second-order differential, thereby extracting the global information of the image. Its eigenvalue sequence eig(Δ) is used to describe the characteristic distribution of the target label image at different frequencies. This information can reflect the detailed changes and global morphology of the image.

[0056] Gaussian smoothing factor The introduction of makes the formula more robust to noise or small interference in the image. Regions with large curvature gradients (such as sharp corners or complex boundaries) are given lower weights after smoothing, while smooth regions retain more feature information, thereby effectively improving the stability of topological features. Gaussian curvature K GThe combination of further enhances the geometric description ability of the formula, which reflects the global curvature characteristics of the image boundary, making the topological feature not only a pure connectivity description, but also contains the three-dimensional geometric morphological characteristics of the image. The formula is obtained by integrating ∫ X dX integrates the above information to generate the high-dimensional topological features H of the target label image. n (X). This integration process not only ensures the globality of feature extraction, but also eliminates noise interference in local calculations, making the feature expression more stable and robust. This design of global integration combined with local characteristics makes the formula highly resistant to morphological changes, scale transformations and noise interference of label images. Compared with traditional feature extraction methods, this formula has achieved technological breakthroughs in many aspects. Traditional methods usually rely on geometric or statistical characteristics, while ignoring the topological structure information of the image, and therefore are prone to insufficient recognition accuracy when processing complex label images. This formula, by combining algebraic topology, differential operators and Gaussian smoothing, can not only fully capture the geometric and topological characteristics of the target label image, but also generate a robust high-dimensional feature expression. In addition, the formula's hierarchical description of the topological space enables the system to flexibly adjust the dimension of feature extraction in different application scenarios, thereby meeting the diverse needs of label image processing. In practical applications, the topological feature H n (X) can be used for the classification, matching, and unique code generation of label images. For example, in industrial production, for label images with complex morphology or porous structures, the formula can extract their key topological features, thereby achieving efficient recognition and classification. At the same time, because the formula is highly robust to noise and deformation, it can also provide stable and reliable performance support in scenarios such as logistics management and image analysis.

[0057] Embodiment 7: Use the following formula to construct the characteristic popular code E based on the topological features using the Gauss-Bonet theorem:

[0058]

[0059] Specifically, the formula is based on the Gauss-Bonnet theorem, which is an important conclusion in differential geometry and establishes the connection between the global topological properties and the local geometric properties of a surface. Specifically, the theorem shows that for a two-dimensional surface with a boundary, its total curvature can be decomposed into the sum of the integral of the Gaussian curvature of the surface and the integral of the mean curvature of the boundary, and this total value is directly related to the topological properties of the surface. In the formula, ∫∫ Ω K G dΩ represents the Gaussian curvature integral of the target label image in region Ω, k g ds indicates boundary The two parts are combined to describe the global curvature distribution and boundary characteristics of the labeled image, which are directly related to the topological structure of the image. The formula generates a geometric quantity that can fully express the global characteristics of the labeled image through the geometric and topological fusion of the Gauss-Bonnet theorem. This geometric quantity is further related to the topological feature H n (X), the generated feature manifold code E not only reflects the geometric properties of the labeled image, but also contains its topological properties. n (X) is the homology group of the target labeled image, which describes the number of connected parts, holes, and cavities respectively. n (X) In the summation operation in the formula, the system can comprehensively consider all the topological characteristics of the target label image in 0, 1 and 2 dimensional space, thereby generating a highly discriminative feature expression. The physical meaning of the formula is reflected in its comprehensive description of the shape and structure of the label image. First, ∫∫ Ω K G As the integral of Gaussian curvature, dΩ reflects the overall curvature characteristics of the target label image. For example, for a regular circular label, its Gaussian curvature distribution is uniform, while for a complex shape label, its Gaussian curvature distribution may vary dramatically with position. By integrating the Gaussian curvature, the system can capture the global geometric characteristics of the target label image. Secondly, k g ds is the integral of the boundary curvature, focusing on the boundary shape characteristics of the target label image, such as edge smoothness, sharpness, and shape change. This combination allows the formula to focus on both the internal and boundary characteristics of the label image, thereby generating a more comprehensive geometric description. n (X) multiplied, the formula completes the deep integration of geometric and topological information. Specifically, the topological feature H n (X) provides the topological structure information of the target label image in high-dimensional space, such as the number of connected areas, the distribution of holes, and the morphology of surface cavities. This topological information is one of the core features of the label image and has stronger stability than geometric properties. For example, no matter how the label image is scaled, rotated, or deformed, its topological properties remain unchanged. This joint expression of geometric and topological properties makes the feature manifold encoding E highly unique and robust. In addition, the formula The coefficient is derived from the proportional relationship between the total curvature and the topological Euler characteristic in the Gauss-Bonnet theorem. This coefficient ensures the mathematical consistency of the feature manifold coding E generated by the formula, and further enhances the sensitivity of the coding result to the topological properties of the label image. Through this design, the system can integrate the geometric and topological information of the label image into a unified coding form in a mathematically rigorous way, thereby providing strong technical support for subsequent coding storage, matching and recognition tasks. The application scenarios of feature manifold coding E are very wide. In the classification and recognition of label images, the code can be used as a unique identifier of the target image to effectively distinguish label images with different shapes, structures and topological characteristics. At the same time, due to the strong robustness to noise and deformation in the formula design, feature manifold coding shows extremely high accuracy and stability in label image recognition in complex environments. For example, in real-time label detection in industrial production, fast label matching in logistics management, and label retrieval tasks in large-scale image databases, feature manifold coding can significantly improve the performance of the system.

[0060] Example 8: Use the following formula and Hessian matrix to perform multi-scale fusion of scale space theory to obtain a fusion result:

[0061]

[0062] Among them, F is the fusion result.

[0063] Specifically, the core idea of ​​the formula is to achieve the fusion of different feature dimensions through scale space theory. Scale space theory is an image analysis framework that aims to capture the key information of the target at different resolutions by introducing feature representations of different scales. Labeled images usually have multi-level and multi-scale characteristics, where details may be prominent at small scales, while global structures are more obvious at large scales. In order to express these characteristics at the same time, the formula uses differential operations in scale space Dynamically adjust the scale of the input features. This operation combines the gradient information and spatial distribution of the features, thus ensuring that the fusion result can balance local details and global characteristics. The first term ΔM in the formula pqr The geometric invariant moment M is introduced pqr The Laplace operator of the label image. The geometric invariant moment describes the multidimensional geometric characteristics of the label image in three-dimensional space, and the Laplace operator further captures the rate of change of these geometric characteristics at different spatial positions. Through this operation, the formula can extract the most significant geometric features in the label image, such as prominent boundaries, sharp corners, and curved changes in space. The introduction of this high-order information enhances the formula's ability to express the geometric structure of complex label images. The second term We focus on the topological characteristic gradient of the target label image. Topological feature H n(X) represents the structural characteristics of the labeled image in different dimensions (such as connectivity, holes, and cavities), and its gradient It reflects the rate of change of these topological characteristics in space. Through the square operation, the formula amplifies the severity of the change of the topological characteristics, thereby more clearly expressing the topological characteristics of the labeled image in different spatial regions. This design is particularly important for processing labeled images with complex morphology and significant topological changes. The feature popular coding E in the formula plays a role in integrating geometric and topological characteristics. In Example 7, feature popular coding has been proven to be able to fully express the geometric and topological characteristics of the target labeled image. By directly embedding E into the multi-scale fusion formula, the system can retain these characteristics during the fusion process, thereby ensuring that the generated fusion result is both robust and recognizable. The introduction of the homology group H n The eigenvalue eig of (X) i (H i (X)), the topological characteristics of the labeled image are deeply decomposed. The product operation of eigenvalues ​​captures the relationship between the topological characteristics of different dimensions of the labeled image, while the sum operation of eigenvalues ​​unifies this information into a global quantity. This processing method effectively extracts high-order correlation information in the topological characteristics of the labeled image, further enhancing the expressiveness of the fusion result. By differentially fusing all the above characteristics in the scale space, the formula generates the final fusion result F. F is a highly comprehensive feature expression that combines the advantages of geometric invariant moments, topological characteristic gradients, feature manifold coding, and high-order eigenvalue decomposition. It can not only fully express the multi-scale and multi-dimensional characteristics of the target labeled image, but also maintain a high degree of robustness in different scenarios. For example, regardless of the morphological complexity of the labeled image or how the shooting conditions (such as lighting and resolution) change, the fusion result generated by the formula can maintain stable expressiveness.

[0064] Embodiment 9: Use the following formula to generate a unique code based on differential topology:

[0065]

[0066] Among them, hash(·) is the hash mapping function; Ric(·) means calculating the Ricci curvature; ∧ is the exterior integral operator.

[0067] Specifically, the core part of the formula is the fusion of the geometric and topological information of the labeled image, which is reflected in the curvature integral ∮ C K∧ds, Ricci curvature calculation of topological feature H1(X), and multi-level expression of fusion feature F. First, ∮ CK∧ds represents the curvature integral on the boundary of the target label image, which is a key calculation in geometric analysis and reflects the overall shape characteristics of the label image boundary. The integral term captures the local and global change patterns of the boundary through the combination of curvature K and boundary microelement ds. For example, for a label image with complex boundaries, the integral can integrate all boundary characteristics to generate a unified geometric quantity, and this unity is crucial for the uniqueness of subsequent encoding. On this basis, the formula introduces the fusion feature F, which is the feature expression generated by multi-scale fusion in Example 8, and already contains the geometric and topological characteristics of the label image. By directly embedding F into the curvature integral, the formula further extends the multi-scale fusion characteristics to the description of the boundary curvature, thereby enhancing the sensitivity and description ability of the encoding to the complex structure of the label image. In addition, the topological feature H1(X) in the formula is the 1-dimensional homology group of the target label image, reflecting the hole characteristics in the label image. By calculating the Ricci curvature Ric(H1(X)) of H1(X), the formula further introduces the curvature characteristics of the topological structure. Ricci curvature is an important concept in differential geometry, which is used to describe the curvature distribution of a manifold in different directions. Applying Ricci curvature to topological properties can capture the complex topological changes of labeled images in high-dimensional space, such as the arrangement pattern, relative position, and degree of curvature of holes. This information is difficult to capture in traditional geometric analysis, but by introducing Ricci curvature, the formula can generate a more comprehensive feature expression.

[0068] The overall structure of the formula is expressed by the exponential function exp(∫ XRic(H1(X)dX)) combines the Ricci curvature integral with the curvature integral and fusion features. This exponential form design is mathematically very stable and can reduce the impact of local feature fluctuations while integrating features. In addition, by performing volume integration on the Ricci curvature, the formula can also capture the global topological characteristics of the labeled image, thereby ensuring that the generated unique code C is highly robust. Finally, the formula converts the complex feature expression into a fixed-length unique code C through the hash mapping function hash(·). The introduction of the hash function is to meet the storage and comparison requirements in practical systems. The hash function has the ability to map large-scale feature data to a finite space while maintaining the uniqueness of the input. Therefore, the generated code C is not only mathematically unique, but also meets the requirements of efficient storage and fast matching in engineering. From the perspective of practical applications, the unique code C based on differential topology has a wide range of applicability. In tasks such as classification, tracking, and retrieval of labeled images, the code can be used as a unique identifier of the target image to effectively distinguish labeled images with complex morphology and similar structures. For example, in industrial production, labels may have slight differences due to process errors, but the unique code C generated by the formula can still accurately identify these label images. In addition, due to the strong robustness to noise, scale changes and morphological transformations in the formula design, it also shows significant advantages in scenarios such as logistics management and large-scale image database retrieval. Compared with traditional image coding methods, the unique coding formula based on differential topology has achieved technological breakthroughs in many aspects. Traditional methods often rely on statistical characteristics or low-order geometric characteristics, and it is difficult to capture the deep topological characteristics of label images. This formula combines differential topology, curvature analysis and scale fusion, which not only fully expresses the geometric and topological characteristics of label images, but also provides efficient storage and matching capabilities through hash mapping.

[0069] The preferred embodiments of the present disclosure are described above with reference to the accompanying drawings, but the scope of the present disclosure is not limited thereby. Any modification, equivalent substitution and improvement made by those skilled in the art without departing from the scope and essence of the present disclosure shall be within the scope of the present disclosure.

Claims

1. An intelligent blank label image recognition and composite coding system, characterized in that: The system comprises: a label image feature processing part, an embedded coding part and a composite coding generation part; the label image feature processing part is used to extract multi-dimensional features of the target label image to obtain the multi-dimensional features of the target label image; based on the multi-dimensional features, a curvature flow feature analysis is performed to calculate the local curvature distribution for the boundary curve of the target label image; based on the local curvature distribution, the geometric invariant moment of the target label image is calculated; the embedded coding part is used to extract the topological features of the target label image based on the continuous homology based on the geometric invariant moment; according to the topological features, the feature manifold coding is constructed by using the Gauss-Bonnet theorem; the composite coding generation part is used to perform multi-scale fusion of scale space theory by using the Hessian matrix according to the feature manifold coding to obtain a fusion result; according to the fusion result, a unique coding based on differential topology is performed.

2. The intelligent blank label image recognition and composite coding system according to claim 1, characterized in that: The label image feature processing part performs multi-dimensional feature extraction on the target label image to obtain the multi-dimensional features of the target label image, and also determines whether the target label image is a blank image before obtaining the multi-dimensional features of the target label image. If so, the multi-dimensional feature extraction is performed on the target label image to obtain the multi-dimensional features of the target label image. If the target label image is not a blank image, a prompt is issued to terminate the processing process. The label image feature processing part uses a method based on pixel value checking to check whether the target label image is a blank image, specifically including: converting the target label image into a grayscale image, going through all pixel values ​​of the grayscale image, and checking whether all pixel values ​​are the same. If so, the target label image is judged to be a blank image.

3. The intelligent blank label image recognition and composite coding system according to claim 2, characterized in that: The multidimensional feature Φ(x, y) is expressed using the following formula: Wherein, x represents the X-axis coordinate; y represents the Y-axis coordinate; I represents the target label image; Represents the gradient of the target label image I; represents the curl operator, Ω represents the entire image domain of the target labeled image I.

4. The intelligent blank label image recognition and composite coding system according to claim 3, characterized in that: The local curvature distribution K is calculated using the following formula: Where, κ is the local curvature value; C is the boundary curve of the target label image I; is the local curvature formula of a two-dimensional curve, (x′, y′) represents the tangent vector of the boundary curve, (x″, y″) represents the rate of change of the curvature of the boundary curve; x′ is the first derivative of x; y′ is the first derivative of y; x″ is the second derivative of x; y″ is the second derivative of y; ds is the micro-arc element of the curve, indicating the small step size in the integration process; σ is the standard deviation of Gaussian smoothing.

5. The intelligent blank label image recognition and composite coding system according to claim 4, characterized in that: The geometric invariant moment M of the target label image is calculated by the following formula: pqr : Wherein, V is a closed volume obtained by expanding the boundary curve of the target label image, which is used as the characteristic volume of the target label image; is the X-axis coordinate of the centroid of the feature volume; is the Y-axis coordinate of the centroid of the feature volume; is the Z-axis coordinate of the centroid of the characteristic volume; p is the order of the moment of the X-axis; q is the order of the moment of the Y-axis; r is the order of the moment of the Z-axis; is the divergence operator of the local curvature distribution K; H is the average curvature of the closed volume; K G is the Gaussian curvature; is the gradient of the local curvature distribution K; dV is the infinitesimal element of the volume integral.

6. The intelligent blank label image recognition and composite coding system according to claim 5, characterized in that: The topological features are calculated using the following formula: in, is a topological feature, an n-dimensional homology group, describing the topological space The n-th dimensional topological property of ; the value of n is 0, 1 or 2; is the number of independent connected parts in the target label image; is the number of holes in the target label image; is the number of cavities on the surface of the target labeled image; ker is the kernel space operator; im is the image space operator; Lap is the Laplace operator; is a one-dimensional boundary operator; is a two-dimensional boundary operator; is the integral differential element of the topological space; eig(Δ) is the eigenvalue sequence of the Laplace operator.

7. The intelligent blank label image recognition and composite coding system according to claim 6, characterized in that: Use the following formula to construct the feature popular coding E based on the topological features using the Gauss-Bonet theorem:

8. The intelligent blank label image recognition and composite coding system according to claim 6, characterized in that: Use the following formula and Hessian matrix to perform multi-scale fusion of scale space theory and obtain the fusion result: Among them, F is the fusion result.

9. The intelligent blank label image recognition and composite coding system according to claim 8, characterized in that: Use the following formula to generate a unique code based on differential topology: Among them, hash(·) is the hash mapping function; Ric(·) means calculating the Ricci curvature; ∧ is the exterior integral operator.

Images