Fast shape recognition method based on dual feature reduction and adjustable wavelet transform

Abstract

The invention discloses a rapid shape recognition method based on dual feature reduction and adjustable wavelet transform, and solves the problems of intra-class difference and high feature structure complexity of same-class shapes in the prior art. The method comprises the following steps: defining a 2D shape contour point coordinate as an input; through first-layer feature structure reduction, only shape contour strong feature points (shape inflection points) are left, and then single-group distance angle features and distance angle group features are obtained respectively; mapping the occurrence time of the strong feature points to a time component of wavelet transform, and adjusting scale parameters according to different distributions of the strong feature points; obtaining more representative features through second-layer feature structure reduction, and only selecting minimum, average and maximum feature structure vectors as feature input in a shape matching stage; and performing shape recognition on the features after two times of feature structure reduction and adjustable wavelet transformation, and obtaining recognition accuracy and speed. The technology can obviously improve the omnibearing performance, and can better meet the requirements of practical application.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision, and relates to a fast shape recognition method based on dual feature reduction and adjustable wavelet transform. Background Art

[0002] How to quickly understand the target image obtained by a computer vision system has become a major and important challenge. Understanding an image mainly involves obtaining the features of the target image. With the continuous development of computer science, computer vision has become the eyes for a computer to perceive the outside world, playing an increasingly important role in the scientific field and being widely applied in human daily life and the national defense technology field. The object of computer vision is the image of an external target obtained by a sensor (SAR image, infrared image, optical image, etc.). After being processed by a computer, a series of parameter information of the target is obtained, including position, size, distance, etc. Therefore, a computer vision system perceives and understands an external target mainly based on the analysis and understanding of the image, and perceives the target through the features represented by the image, mainly including features such as shape, color, and texture.

[0003] However, a computer vision system generally needs to transmit the target image over a long distance to obtain and analyze it. Especially for remote sensing targets, some feature information of the target image obtained by the computer vision system may be lost due to natural factors, such as color, texture, etc. At this time, 2D shape has become the most important feature for a computer vision system to analyze and understand the target image, and is widely used in target detection and recognition. Therefore, shape recognition plays an increasingly important role in computer vision and image understanding, and is an important tool for a computer to perceive the target. Shape representation of target features is mainly divided into local shape features, global shape features, and hybrid shape features.

[0004] Currently, the recognition of 2D shapes mainly extracts the target shape features through traditional methods. With the emergence and development of deep learning, the 2D shape recognition method based on convolutional neural network has correspondingly developed. However, the neural network is relatively complex in structure, and the existing 2D shape datasets are small in scale and cannot meet the requirements of the deep neural network for the training volume. Therefore, it generally cannot be applied to the actual scenario. In the process of using distance and angle to represent the features of 2D shapes by the existing traditional methods, the following problems exist: First, the acquisition of contour points ignores the edge differences between shapes of the same category, resulting in low recognition accuracy; Second, complex feature descriptors are uniformly used for 2D shapes with simple and complex edge structures, resulting in high feature calculation costs and slow recognition speeds; Third, in the process of shape feature matching, in order to meet the invariance of features, the expression of the feature vector is deformed, which increases the computational burden on the basis of the feature structure with weak feature redundancy, and thus reduces the recognition speed. These three problems directly lead to poor 2D shape recognition performance, unable to meet the actual requirements for high recognition accuracy and real-time performance, and it is difficult to be applied in practice. Summary of the Invention

[0005] The present invention proposes a 2D shape recognition method based on dual feature reduction and adjustable wavelet transform. The first layer of feature structure simplification is to reduce the significant contour points. Since there are local deformations or differences in the contour edges of 2D shapes of the same category, there is feature redundancy in the significant contour points, resulting in low correct recognition rate and slow recognition speed. Therefore, the reduction strategy of using strong contour feature points instead of significant contour points overcomes this problem and improves the recognition performance. The adjustable wavelet transform is to solve the problem of different complexities of the edge structures of 2D shapes of the same category, which will lead to a slowdown in the shape recognition speed. Among them, the adjustment means that the scale of the wavelet adaptively changes according to the distribution quantity of the strong contour feature points within a certain time period, which can better match the changes in the shape edge structure. The second layer of feature structure simplification is the reduction of the feature structure after the adjustable wavelet transform. Only the strong feature vectors (minimum value, mean value, and maximum value) are selected to represent the shape features, which greatly reduces the matching cost and improves the shape recognition speed.

[0006] The technical solution of the present invention is to provide a fast shape recognition method based on dual feature reduction and adjustable wavelet transform, which includes the following steps:

[0007] Step 1: Define the plane coordinates of the contour points in the 2D shape recognition process as its feature input;

[0008] Step 2: Organically select the contour points through the first layer of feature structure reduction strategy, and only leave the strong contour feature points (shape inflection points) of the shape as the feature input, and then obtain single-group distance-angle features and distance-angle group features respectively;

[0009] Step 3: Map the feature sequence to the frequency components of the wavelet transform while mapping the occurrence time of the strong feature points to the time component of the wavelet transform, and adjust the scale parameter of the wavelet transform according to the different shape edge feature distributions;

[0010] Step 4: Obtain a more representative feature structure vector as the final feature of the 2D shape through the second-layer feature structure reduction strategy, and only select the minimum feature structure vector, the average feature structure vector, and the maximum feature structure vector as the feature inputs in the shape matching stage;

[0011] Step 5: Use the features after two times of feature structure reduction and adjustable wavelet transform as 2D shape features for shape recognition and obtain the recognition accuracy and speed.

[0012] Preferably, in step 1, a binary image f(x, y) containing a shape is given, the shape region is denoted as D, and the image plane is represented as a set of pixels Assume that Ω is the set of all pixels on the outer contour line of the shape in D, which can be represented in the form of arc-length parameterization as: where Since Ω is a closed contour, then z(t + 1) = z(t) and z(t - 1) = z(t).

[0013] Preferably, step 2 includes the following steps:

[0014] Step 2.1: Based on the feature point reduction strategy, use strong feature points to represent the contour feature input points, then the original point p = z(t) becomes p s = z(nt). Where n ≥ 1, n represents the change in the time position of the reduced contour strong feature points compared to the original contour significant points, and it is a constant with a changing value, which changes according to the position change of the reduced contour strong feature points.

[0015] Step 2.2: Obtain a single set of distance-angle features. For the contour point p = z(nt), along and walk respectively, and their endpoints are p' = z(nt + s) and p'' = z(nt - s) respectively, where s ∈ (0, 0.5) is the normalized length of the contour part generated by moving counterclockwise from point p to the end point p'. The walking paths sometimes fall within the shape region and sometimes outside the shape region.

[0016] The length of walking within the shape region is and The length of walking outside the shape region is and They can be represented mathematically by the following formula:

[0017]

[0018] Among them, where l s and l -s , and are respectively the length and direction of the distance pair and .

[0019] The and defined above are the integrals of the shape image function f(x, y) over the distance pairs and , while and are the integrals of the complement of this function over the distance pairs and . This method defines a five-tuple variable W. By observing the shape while walking along the distance pairs, the invaded area on this shape is described by the following formula:

[0020]

[0021] Distance pair walking measurement is used as a primitive descriptor to capture the local geometric and structural features around the contour point p = z(nt). The first two variables reflect the characteristics of the neighboring area of size s on the left distance, while the last two variables reflect the characteristics of the neighboring areas of size s on the right and on the line. The nature of the left distance neighborhood of size s, and the last two units represent the nature of its right distance neighborhood of size s. The intermediate variable describes the curvature characteristics of the left and right distances, where takes values as described by the following formula:

[0022]

[0023] Therefore, the distance pair walking variable can not only capture the geometric feature information of the external shape contour, but also capture the structural information of the internal area of size s neighboring the contour point z(nt).

[0024]

[0025] Step 2.3: Obtain the distance angle group features based on Step 2.2.

[0026] The parameter s in -1 is a scale that controls the neighborhood size of the contour point p = z(nt). By changing the value of s to 2 -2 , 2 -K ,..., 2

[0026] K distance pair walks starting from the common point p can be obtained and combined into shape features, defined by the following formula:

[0027] Where K is the total number of neighborhoods in the distance angle group.

[0028] For the first scale k = 1, i.e., s = 2 -1 , the entire contour is taken as the neighborhood of point p, and the corresponding feature captures the coarsest global feature of the shape. As the scale index k increases, the neighborhood becomes smaller, and the corresponding feature can capture finer local details. Therefore, this feature has the ability to capture shape feature information in the entire range from coarse to fine.

[0029] To further enhance the distinguishability of the feature, the spatial relationship between distance pairs during walking is also considered in the design. For two pairs of adjacent distance pairs of walking and their spatial relationships can be respectively represented as the included angles of the left and right distances and defined by the following formula:

[0030]

[0031] Since there are a total of K distance pairs of walking in the distance angle group, there are K - 1 angle pairs. The spatial relationships of all distance pairs are represented by a combined angle pair as the following formula:

[0032]

[0033] Combining the geometric structure information and spatial information of the distance pairs, the distance feature group descriptor for a point p = z(nt) on the contour can be defined as a vector of size 5K - 3 + K - 1 = 6K - 4, represented by the following formula:

[0034]

[0035] Preferably, step 3 includes the following steps:

[0036] Step 3.1: The definition of the discrete orthogonal wavelet transform is as follows: Assume that ψ(t) ∈ L 2 (R) is an admissible wavelet, and its time-frequency dilation and translation system is represented by the following formula:

[0037]

[0038] The above constitutes an orthonormal basis of L 2 (R), then ψ(t) is an orthogonal wavelet, and the corresponding discrete wavelet transform is represented by the following formula:

[0039] WT f (a, τ) = <f(t), ψ a,τ (t)>

[0040] Based on the above - mentioned relevant rules and requirements of orthogonal wavelets, a typical decaying orthogonal wavelet function h(a, t′) is defined in the present invention, and the detailed expression of this function is defined as the following formula:

[0041]

[0042] Among them, a represents the scale parameter of the orthogonal wavelet. Since the distribution of strong feature points of the obtained shape contour is uneven, with strong feature points being dense in some areas and sparse in some areas, the present invention uses an adjustable wavelet to solve the problem of uneven distribution of strong feature points, and organically selects different wavelet scales for different feature distributions, so that wavelet transform can not only reflect simple contour features but also complex contour features.

[0043] Step 3.2: During a certain time period T′, the number of strong feature points p s = z(nt), nt ∈ T′ is not a fixed value. The more the number of strong feature points, the more obvious the detailed features of the shape contour in this time period. Therefore, a smaller - scale wavelet is needed to analyze the detailed features, which is manifested as a smaller frequency in the wavelet function. Therefore, the scale a of the wavelet and the number N s of strong feature points in this time period show an inverse - proportional relationship, expressed as Let When a = 1, it means that the number of strong feature points takes the average value in this time period. Therefore, k a = N avs , N avs represents the average number of strong feature points in the time period T, then The above formula can be rewritten as the following formula:

[0044]

[0045] The above formula represents the wavelet function in the adjustable orthogonal wavelet transform in the present invention. Based on the above - mentioned description, the features proposed in the present invention are transformed by the adjustable orthogonal wavelet during the time period T′ into which is expressed by the following formula:

[0046]

[0047] Among them, t′ = nt, T′ represents the fixed time period on the contour. Thus, a shape feature in the local time domain that satisfies translation, scale transformation, rotation, and mirror - image invariance is obtained. Since the entire shape contour is divided into N T time domains of T′, the final adjustable orthogonal wavelet transform feature is expressed by the following formula:

[0048]

[0049] Preferably, the second - layer feature structure reduction in step 4 is to selectively obtain strong feature structure vectors as shape features, including the minimum orthogonal wavelet transform feature vector, the average wavelet transform feature vector, and the maximum wavelet transform feature vector. The final shape feature vector obtained based on the above steps is:

[0050]

[0051] Preferably, in step 5, matching recognition is performed according to the final shape features to obtain the recognition accuracy rate and the matching time.

[0052] The present invention has the following advantages:

[0053] The first - layer feature structure reduction reduces the feature - significant points to strong feature points, effectively reducing the problems of the shape - edge difference and deformation of the same - category objects. On the basis of improving the feature robustness, the recognition accuracy rate is improved; the adjustable wavelet transform organically adjusts the scale of the wavelet transform according to the distribution of strong feature points, and then adapts to the different complexities of the shape - edge structures. Moreover, the orthogonal wavelet itself meets the requirement of the invariance of shape features, avoiding additional computational costs, enhancing the feature robustness while improving the recognition speed; the second - layer feature structure reduction removes the weak feature structure vectors by selection, leaving only the strong feature structure vectors with more feature representativeness, greatly accelerating the recognition speed.

[0054] All - around experimental results on the general dataset have shown the effectiveness of the method proposed in the present invention in terms of recognition performance compared with some state - of - the - art (SoTA) methods of the same feature type. On the basis of greatly improving the recognition accuracy rate, the recognition speed is more than 200 times faster than that of the same - type feature methods. Especially compared with the CBW method. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is the flowchart of the fast shape recognition method based on dual - feature reduction and adjustable wavelet transform of the present invention;

[0056] Figure 2 is the specific schematic diagram of the strong - feature contour points in the present invention;

[0057] Figure 3 is the schematic diagram of a single - group distance - angle feature in the present invention;

[0058] Figure 4 is the schematic diagram of the distance - angle group feature in the present invention;

[0059] Figure 5 is in the present invention Figure 2Schematic diagram of the average feature vector of a pair of shapes after two - layer feature reduction and adjustable wavelet transform. Detailed implementation manners

[0060] The present invention will be further described below in conjunction with the accompanying drawings and detailed implementation manners.

[0061] A fast shape recognition method based on dual feature reduction and adjustable wavelet transform includes the following steps:

[0062] As Figures 1 to 5 shown,

[0063] S1: Given a binary image f(x, y) containing a shape, the shape region is denoted as D, and the image plane is represented as a set of pixels Assume that Ω is the set of all pixels on the outer contour line of the shape in D, which can be represented in the form of arc - length parameterization as: where Since Ω is a closed contour, then z(t + 1) = z(t) and z(t - 1) = z(t).

[0064] S2: Select contour points organically through the first - layer feature structure reduction strategy, and only leave the strong feature points (shape inflection points) of the shape contour as the feature input, and then obtain single - group distance - angle features and distance - angle group features respectively.

[0065] S2.1: Use the strong feature point reduction strategy to represent the contour feature input points with strong feature points, then the original point p = z(t) becomes p s = z(nt). Where n≥1, n represents the change in the time position of the strong feature points of the reduced contour compared to the original contour significant points, and it is a constant with variable values, which changes according to the position change of the strong feature points of the reduced contour.

[0066] S2.2: Obtain single - group distance - angle features, as Figure 2 shown. For the contour point p = z(nt), walk along and respectively. Their endpoints are p′ = z(nt + s) and p″ = z(nt - s) respectively, where s∈(0, 0.5) is the normalized length of the contour part generated by moving counter - clockwise from point p to the end point p′. The walking paths sometimes fall within the shape region and sometimes fall outside the shape region.

[0067] The length of walking within the shape region is and The length of walking outside the shape region is and They can be expressed by the following mathematical formula:

[0068]

[0069] Among them, where l s and l -s , and are the length and direction of the distance pair and respectively.

[0070] The and defined above are the integrals of the shape image function f(x, y) over the distance pair and , while and are the integrals of the complement of the function over the distance pair and . This method defines a five-tuple variable W. By observing the shape while walking along the distance pair, the invaded area on the shape is described by the following formula:

[0071]

[0072] Distance pair walking measurement is used as a primitive descriptor to capture the local geometry and structural features around the contour point p = z(nt). The first two variables reflect the characteristics of the neighboring area of size s on the left distance pair, while the last two variables reflect the characteristics of the neighboring areas of size s on the right and on the line. The nature of the left distance neighborhood of size s, and the last two units represent the nature of its right distance neighborhood of size s. The intermediate variable describes the curvature characteristics of the left and right distances, where takes values as described by the following formula:

[0073]

[0074] Therefore, the distance pair walking variable can not only capture the geometric feature information of the external shape contour, but also capture the structural information of the internal area of size s adjacent to the contour point z(nt).

[0075] S2.3: Obtain the distance angle group features based on S2.2.

[0076] The parameter s in -1 , 2 -2 ,..., 2 -K is a scale that controls the neighborhood size of the contour point p = z(nt). By changing the value of s to 2

[0077]

[0078] Among them, K is the total number of neighborhoods in the distance-angle group.

[0079] For the first scale k = 1, that is, s = 2 -1 , the entire contour is taken as the neighborhood of point p, and the corresponding feature captures the coarsest global feature of the shape. As the scale index k increases, the neighborhood becomes smaller, and the corresponding feature can capture finer local details. Therefore, this feature has the ability to capture shape feature information throughout the range from coarse to fine. Figure 4 An illustrative example of the distance-angle group feature is given.

[0080] To further enhance the discriminability of the feature, the spatial relationship between distances during walking is also considered in the design. For two pairs of adjacent distance pairs during walking and their spatial relationships can be respectively represented as the included angles of the left and right distances and defined by the following formula:

[0081]

[0082] Since there are a total of k distance pairs during walking in the distance-angle group, there are K - 1 angle pairs. The spatial relationships of all the distance pairs in the distance-angle group are represented by a combined angle pair as the following formula:

[0083]

[0084] Combining the geometric structure information and spatial information of the distance pairs, the distance feature group descriptor for a point p = z(t) on the contour can be defined as a vector of size 5K - 3 + K - 1 = 6K - 4, represented by the following formula:

[0085]

[0086] S3: Map the feature sequence to the frequency components of the wavelet transform while mapping the occurrence time of the strong feature points to the time component of the wavelet transform, and adjust the scale parameter of the wavelet transform according to different shape edge feature distributions.

[0087] S3.1: The definition of the discrete orthogonal wavelet transform is as follows: Assume ψ(t) ∈ L 2 (R) is an admissible wavelet, and its time-frequency dilation and translation system is represented by the following formula:

[0088]

[0089] The above formula constitutes L 2If {ψ(t - k)} is an orthonormal basis of (R), then ψ(t) is an orthogonal wavelet, and the corresponding discrete wavelet transform is expressed by the following formula:

[0090] WT f (a, τ) = <f(t), ψ a,τ (t)>

[0091] Based on the above rules and requirements of the orthogonal wavelet, a typical decaying orthogonal wavelet function h(a, t′) is defined in the present invention, and the detailed expression of this function is defined by the following formula:

[0092]

[0093] where a represents the scale parameter of the orthogonal wavelet. Since the distribution of the strong feature points obtained from the shape contour is uneven, as Figure 2 shown, the strong feature points are dense in some areas and sparse in some areas. Therefore, the present invention uses an adjustable wavelet to solve the problem of uneven distribution of strong feature points, and organically selects different wavelet scales according to different feature distributions, so that the wavelet transform can not only reflect simple contour features but also complex contour features.

[0094] S3.2: During a certain time period T′, the number of strong feature points p s = z(n′t), n′t ∈ T is not a fixed value. The more the number of strong feature points, the more obvious the detailed features of the shape contour in this time period. Therefore, a wavelet with a smaller scale is needed to analyze the detailed features, which is manifested as a smaller frequency in the wavelet function. Therefore, the scale a of the wavelet in a certain time period and the number N s of the strong feature points in this time period show an inverse relationship, expressed as Let When a = 1, it means that the number of strong feature points takes the average value in this time period. Therefore, k a = N avs , N avs represents the average number of strong feature points in the time period T, then The above formula can be rewritten as the following formula:

[0095]

[0096] The above formula represents the wavelet function in the adaptive orthogonal wavelet transform of the present invention. Based on the above description, the distance - angle group feature proposed in the present invention is adaptively orthogonally wavelet - transformed in the time period T′ into expressed by the formula as the following formula:

[0097]

[0098] where \(t' = nt\), and \(T'\) represents a fixed time period on the contour. Thus, a shape feature in the local time domain that satisfies translation, scale transformation, rotation, and mirror invariance is obtained. Since the entire shape contour is divided into \(N\) T time domains of \(T'\), the final adjustable orthogonal wavelet transform feature is represented by the following formula:

[0099]

[0100] S4: The second-layer feature structure reduction is to selectively obtain strong feature structure vectors as shape features, including the minimum orthogonal wavelet transform feature vector, the average wavelet transform feature vector, and the maximum wavelet transform feature vector. Among them, the schematic diagram of the average wavelet transform feature vector is as Figure 5 shown. The final shape feature vector obtained based on the above steps is:

[0101]

[0102] S5: Perform matching recognition according to the final shape feature to obtain the recognition accuracy rate and the matching time.

[0103] Experimental Results and Analysis:

[0104] 1. Experimental Environment

[0105] The experimental development environment is: Intel(R) Core(TM) (America) i7-4850 2.30GHz CPU + NVIDIA GeForce GT 750M + Intel GMA HD 5200 GPU; the development tool is: Matlab.

[0106] 2. Experimental Data Set

[0107] To verify the effectiveness of the present invention, experiments were carried out using the leaf shape data set, including the ICL data set, the MEW2012 data set, and the CVIP100 data set. Among them, the ICL data set is an internationally common leaf shape data set, which has 220 categories, 26 to 108 shapes in each category, and a total of 16,846 leaf shapes; the Central European Woody Plants (MEW2012) is an open leaf data set, which includes native or frequently cultivated trees and shrubs in the Central European region. There are 9,745 leaf images in total, divided into 153 species, and each species has no less than 50 samples; the CVIP100 data set contains 1,200 leaf images of 100 plants, and each plant has 12 different leaf images.

[0108] 3. Experimental Results and Analysis

[0109] Tests were conducted on the ICL, MEW2012, and CVIP100 datasets respectively. The comparison methods are some methods of the same feature type in recent years, classic and representative shape recognition methods, and SoTA methods. It can be seen from the experimental results in the table that the 2D shape recognition method based on dual feature structure reduction and adjustable wavelet transform proposed in the present invention has significantly improved recognition accuracy and speed on multiple internationally common datasets. Since the present invention uses dual feature structure reduction to reduce the feature differences and redundant weak features of the same category of shapes, combined with the feature robustness of adjustable wavelet transform, the effectiveness of the method proposed in the present invention is more significant and the performance is stronger.

[0110] Experiment 1 was conducted on the entire ICL dataset, including the present invention and other classic high-performance methods, and the MAP (Minimum Average Precision) and average recognition time were obtained respectively. The results show that the present invention performs best with a MAP score of 79.2% and an average recognition time of 26.9 milliseconds, having both the highest accuracy and the fastest recognition speed. In contrast, the SC method has a MAP score of 35.0% and the longest recognition time, reaching 6200 milliseconds, with poor performance. Other methods such as IMTD and CBW also have relatively high MAP scores, but the recognition speeds are relatively slow, 1000 milliseconds and 5110 milliseconds respectively. Generally speaking, the present invention shows significant performance advantages in this comparative experiment.

[0111] Experiment 2 was conducted on the entire MEW2012 dataset, including the present invention and other classic high-performance methods, and the MAP (Minimum Average Precision) and average recognition time were obtained respectively. The results show that the present invention performs best with a MAP score of 79.2% and an average recognition time of 23.4 milliseconds, showing high accuracy and extremely fast processing speed. The IMTD method has the highest MAP score of 79.4% and a recognition time of 950 milliseconds, and its performance is also excellent. In contrast, the SC method has a MAP score of 48.0% and the longest recognition time of 6500 milliseconds. Other methods such as HSC perform outstandingly in terms of recognition speed, with an average recognition time of only 170 milliseconds, but the MAP score is 55.0%. Overall, the present invention shows obvious advantages in balancing recognition accuracy and speed.

[0112] Experiment 3 was conducted on the entire CVIP100 dataset, including the present invention and other classic high-performance methods, to obtain the MAP (Minimum Average Precision) and average recognition time respectively. The results show that the present invention demonstrates the highest accuracy and the fastest recognition speed with a MAP score of 95.2% and an average recognition time of 23.9 milliseconds. The IMTD method has a MAP score of 91.5% and a recognition time of 900 milliseconds, showing relatively excellent performance. The HSC method has a MAP score of 86.1% and a relatively fast recognition speed, with an average recognition time of 150 milliseconds. In contrast, the MAP scores of the SC and HF methods are 81.5% and 84.5% respectively, but their recognition times are relatively long, 8700 milliseconds and 9800 milliseconds respectively. Overall, the present invention shows obvious advantages in balancing recognition accuracy and efficiency.

[0113] In summary, the comprehensive experimental results on the general dataset all demonstrate the effectiveness of the method proposed by the present invention in terms of recognition performance compared with some SoTA methods of the same feature type. On the basis of greatly improved recognition accuracy, the recognition speed is more than 200 times faster than that of the same type of feature method. Especially compared with the CBW method.

Claims

1. A fast shape recognition method based on dual feature reduction and adjustable wavelet transform, characterized in that: It includes the following steps: Step 1: Define the plane coordinates of the contour points in the 2D shape recognition process as its feature input; Step 2: Organically select contour points through the first-layer feature structure reduction strategy, only leaving the strong feature points (shape inflection points) of the shape contour as the feature input, and then obtain single-group distance-angle features and distance-angle group features respectively; Step 3: Map the feature sequence to the frequency components of wavelet transform while mapping the occurrence time of strong feature points to the time components of wavelet transform, and adjust the scale parameter of wavelet transform according to different shape edge feature distributions; Step 4: Obtain a more representative feature structure vector as the final feature of the 2D shape through the second-layer feature structure reduction strategy, and only select the minimum feature structure vector, average feature structure vector and maximum feature structure vector as the feature input in the shape matching stage; Step 5: Use the features after two-layer feature structure reduction and adjustable wavelet transform as 2D shape features for shape recognition and obtain the recognition accuracy and speed; In step 2, the feature structure reduction strategy is used to only leave strong feature points as the input points of the shape contour, that is, the original point p = z(t) is changed to p s = z(nt), where n ≥ 1, n represents the change in the time position of the strong feature points of the reduced contour compared to the significant points of the original contour, and it is a constant with variable values, which changes according to the position change of the strong feature points of the reduced contour; The shape distance-angle group feature after feature structure reduction in Step 2 is: In Step 3, while mapping the feature sequence to the frequency components of wavelet transform, map the occurrence time of strong feature points to the time components of wavelet transform, and adjust the scale parameter of wavelet transform according to different shape edge feature distributions; it includes the following steps: Step 3.

1. The definition of the discrete orthogonal wavelet transform is as follows: Assume that ψ(t) ∈ L 2 (R) is an admissible wavelet, and its time-frequency dilation and translation system is expressed as follows: The above formula constitutes L 2 (R)'s orthonormal basis, then ψ(t) is an orthogonal wavelet, and the corresponding discrete wavelet transform is expressed by the following formula: WT f (a,τ) = <f(t), ψ a,τ (t)> Based on the above relevant rules and requirements of orthogonal wavelets, a typical decaying orthogonal wavelet function h(a, t ′ ) is defined in the present invention, and the detailed expression of this function is defined as the following formula: Among them, a represents the scale parameter of the orthogonal wavelet; since the distribution of the obtained strong feature points of the shape contour is uneven, with strong feature points dense in some areas and sparse in some areas, the present invention uses an adaptive wavelet to solve the problem of uneven distribution of strong feature points, and adaptively selects different wavelet scales according to different feature distributions, so that the wavelet transform can not only reflect simple contour features but also complex contour features; Step 3.2, within a certain time period T ′ Inside, the number of strong feature points p s = z(n ′ t), n ′ t ∈ T ′ is not a fixed value. The more the number of strong feature points, the more obvious the detailed features of the shape contour in this time period. Therefore, wavelet analysis with a smaller scale is required for detailed features, which is manifested as a smaller frequency in the wavelet function. Therefore, within a certain time period, the scale a of the wavelet and the number N s of strong feature points in this time period show an inverse relationship, expressed as Let When a = 1, it means that the number of strong feature points takes the average value in this time period. Therefore k a = N avs , N avs represents the average number of strong feature points within the time period T, then The above formula can be rewritten as the following formula: The above formula represents the wavelet function in the adaptive orthogonal wavelet transform of the present invention; based on the above description, the distance-angle group feature proposed by the present invention in the time period T ′ is adaptively orthogonally wavelet-transformed into which is expressed by the following formula: where t ′ = nt, T ′ represents a fixed time period on the contour. So far, a shape feature of the local time domain that satisfies translational, scale transformation, rotational, and mirror invariance is obtained. Since the entire shape contour is divided into N T T ′ time domains, the final adjustable orthogonal wavelet transform feature is expressed by the following formula: The second-layer feature structure reduction in Step 4 is to selectively obtain strong feature structure vectors as shape features, including the minimum orthogonal wavelet transform feature vector, average wavelet transform feature vector and maximum wavelet transform feature vector. The final shape feature vector obtained on the basis of the above steps is:

Images