Method for improving image segmentation capability
By establishing the relationship between discrete image information and time signals, and utilizing Fourier transform and cusp catastrophe theory, the image segmentation process is optimized, solving the problem of insufficient image segmentation accuracy in existing technologies, and achieving high-precision image segmentation and depth analysis.
Patent Information
- Application Number
- CN202410849022.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-27
- Publication Date
- 2024-11-15
AI Technical Summary
Existing technologies cannot deeply interpret discrete information in images or reveal the intrinsic relationship between discrete information and thresholds, resulting in insufficient image segmentation accuracy, especially in color image processing.
By establishing the relationship between discrete image information and time signals, and utilizing Fourier transform and cusp catastrophe theory, a mathematical model is constructed to accurately locate the threshold and optimize the image segmentation process.
It improves the accuracy and effectiveness of image segmentation, is applicable to various image types, and enables in-depth analysis and high-precision segmentation of image information.
Smart Images

Figure CN118967732A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of image recognition technology, in particular to a method for improving image segmentation capability by converting discrete image information into a continuous time series. Background Art
[0002] Image threshold segmentation is widely used in various engineering applications due to its simple implementation and high segmentation efficiency. The basic principle of image threshold segmentation is to use the selected threshold to divide image pixels into targets or backgrounds. Therefore, how to choose a reasonable segmentation threshold is the key to the accuracy of image threshold segmentation. At present, according to the different types of information used, threshold segmentation algorithms can be divided into the following categories:
[0003] (1) Histogram shape-based methods: This type of method mainly divides pixels according to the shape attributes of the histogram, which can be divided into three categories: "convex hull" method, "peak and valley" method and shape modeling method.
[0004] (2) Entropy-based methods: This type of method uses the entropy information of grayscale distribution to divide pixels.
[0005] (3) Clustering-based threshold style method. This type of method can be divided into four categories: iterative method, clustering method, minimum error method and fuzzy clustering. It mainly obtains the threshold by clustering analysis of grayscale data. Among them, the clustering method obtains the optimal threshold by minimizing the weighted method of foreground and background. It is one of the more classic algorithms in threshold segmentation algorithms.
[0006] Taking the histogram shape-based method as an example, the discrete information histogram of an image can present peakless, unimodal, bimodal or multi-modal morphological characteristics, while most traditional threshold selection methods are only suitable for processing discrete information with unimodal or bimodal morphology. Therefore, it is urgent to develop a new method to replace the traditional threshold selection method.
[0007] At present, the improvement methods for threshold selection are mostly focused on the use of artificial intelligence processing such as machine learning and optimization algorithms. It mainly divides the image into non-overlapping regions, and the region is defined as a group of similar connected pixels relative to the selected attribute. There are several methods to define the homogeneity of the region, which are based on specific goals in the segmentation process, such as: image color, grayscale, texture, layer depth, etc. Regions are not allowed to overlap with each other, so each pixel belongs to only a single region. If the new merged region is homogeneous, two adjacent regions should be merged. Therefore, each region is expected to be as large as possible under its specific characteristics, and then the total number of regions is reduced.
[0008] Most histogram-based algorithms work on grayscale images, which are represented as one-dimensional (1D) histograms. The intensity range is assumed to be constant, and the histogram is considered to be a Gaussian probability density function. The segmentation problem is reformulated as parameter estimation followed by inverse pixel classification. However, if the problem involves color images, selecting a global threshold in a 3D histogram is a difficult task. For threshold selection of color images, extending the general 3D histogram to a homogeneous domain histogram is a new idea. A traditional histogram based on hue features is constructed for any region and a histogram analysis is performed. However, most of the histogram-based methods have been found to be image-specific and have achieved reasonable performance when the input features are noise-free and the number of regions is small.
[0009] Based on this principle, machine learning is used to perform deep learning processing on the regional data collected in each area, and the threshold is obtained after multiple layers of iteration, and the image is segmented based on the threshold information. Although this method can obtain relatively accurate threshold information, model training takes a long time. When the number of samples is insufficient, it will cause a large deviation in the output results, and it is impossible to continue to deeply interpret the discrete information of the image, and it is impossible to reveal the intrinsic relationship between discrete information and threshold. Summary of the invention
[0010] In order to solve the technical problem in the prior art that it is impossible to deeply interpret the discrete information of an image and reveal the intrinsic relationship between the discrete information and the threshold, the present invention provides a method for improving the image segmentation capability.
[0011] The technical solution adopted by the present invention is: a method for improving image recognition and segmentation capabilities, comprising the following steps:
[0012] (1) Extraction of image information: Obtain the image to be segmented, classify the image, import the image into the threshold segmentation software, and obtain the image discrete information histogram;
[0013] (2) Establish the relationship between discrete time signals and discrete image information (histogram):
[0014] 1) Introduce an auxiliary function, which is an even function and its domain is f∈(-∞,+∞);
[0015] 2) Since the discrete information of the image is meaningless when the domain is f<0, according to the properties of the even function, the time series in the negative frequency range is converted to the positive frequency range to obtain the unilateral spectrum function;
[0016] 3) According to the image threshold segmentation software, the image sampling time interval, the number of sampling points and the total sampling time information are converted into the frequency difference between two adjacent points of the discrete Fourier transform to obtain its frequency resolution;
[0017] 4) Discretize the time series according to the obtained frequency resolution to obtain the time series curve of the image discrete information, complete the mathematical relationship between the image discrete time signal and the discrete form of image information, and prove that this type of image information meets the prerequisite of Fourier transform;
[0018] (3) Extract discrete information based on the histogram and import it into Origin or Matlab software for inverse Fourier transform to obtain the amplitude and phase diagram of the time series. According to the image type, it can be proved whether the time series is a stationary random process of various states and whether a random phase needs to be used.
[0019] (4) Perform mutation analysis on the obtained time series to reveal the intrinsic relationship between the discrete information of the image and the threshold.
[0020] Mutation analysis is performed on the obtained time series. The Tate expansion is used once in each segment and substituted into the cusp mutation bifurcation set to accurately locate the position of the state parameter mutation.
[0021] 1) In the cusp mutation model, there is one state variable x and two control variables u and v, and its standard potential function V(x) is:
[0022] V(x)=x 4 +qx 2 +px (1)
[0023] The equilibrium curve equation in the model satisfies V'(x)=0, and the singular point set equation satisfies V”(x)=0. Combining the two equations, we can get:
[0024] Δ=8q 3 +27p 2 (2)
[0025] 2) The equilibrium surface is divided into upper, middle and lower leaves. Assume that M(x, y, z) is a point representing the state of the system. When M moves along the upper and lower leaves, the system is in a stable state. When the motion trajectory of point M crosses the bifurcation set and is in the middle leaf of the equilibrium surface, the system undergoes a mutation. The above phenomenon is characterized by the eigenvalue V:
[0026] V>0, no mutation (stable state);
[0027] V = 0, critical state;
[0028] V<0, mutation (unstable state). (3)
[0029] 3) Since the obtained time domain curve is based on the IFFT method, its essence is a function based on a combination of multiple sine and cosine, and it does not have the properties of a complete time series. In order to obtain a complete time series that can be substituted into the cusp mutation bifurcation set, it is necessary to convert the time domain curve into a time series. First, an original sequence x is constructed based on the bonding area time domain curve. 0 :
[0030] x 0 =(x 0 (1),x 0 (2),…x 0 (n)) (4)
[0031] Superimposing equation (4) once, we get the AGO sequence, which is the one-dimensional time series ω corresponding to the bonding area. 1 (t):
[0032] x 1 =(x 1 (1),x 1 (2),…x 1 (n)) (5)
[0033] In formula (5):
[0034] 4) Expand the obtained AGO sequence using Taylor's formula, retain the first 5 items of the sequence, and obtain ω 1 The approximate expression of (t) is:
[0035]
[0036] Formula (6) is a one-dimensional time series obtained by gray cumulative calculation of the bonding density time domain curve, that is, the bonding area density time series with complete attributes, where A 0 , A 1 , …, A 5 is the unknown coefficient, which is determined by polynomial fitting of the bonding density time domain curve. After derivation of formula (6), the restored bonding area density time series is obtained:
[0037]
[0038] 5) Let a 0 =A 1 ,a 1 =2A 2 ,a 2 =3A 3 ,a 3 =4A 4 ,a 4 =5A 5, the standard potential function of equation (12) is constructed by using the parameter substitution method
[21] , when a 4 >0, let when a 4 <0, let q=a 3 / 4a 4 , we can get:
[0039] when a 4 >0:
[0040] V(z)=z 4 +uz 2 +vz+w (8)
[0041] Where:
[0042] when a 4 <0 hours:
[0043]
[0044] Then, by combining the cusp mutation theory (7), we can deduce:
[0045]
[0046] (5) Differentiate the image discrete information time series curve along the time axis, perform a mutation analysis in each time period, and obtain a curve of Δ changing with time series (hereinafter referred to as the mutation time series). Select the appropriate mutation point as the final threshold according to the change of the mutation time series.
[0047] (6) After obtaining the precise final threshold, it is necessary to input the threshold into the threshold segmentation software and re-segment the image to verify the accuracy of the new threshold. The accuracy is evaluated in a variety of ways depending on the image type.
[0048] The outstanding advantages of the present invention are:
[0049] 1. Compared with the threshold information based on traditional discrete information division, the optimized threshold information has the advantages of high accuracy and good segmentation effect, and is more capable of meeting the needs of high-precision image information recognition.
[0050] 2. It can convert discrete image information into continuous functions. By establishing a mathematical model, it can achieve in-depth analysis of image information. It can not only be used to divide thresholds, but also be applicable to other purposes of image analysis, providing a new idea for extracting and analyzing image information.
[0051] 3. The invention has a wide range of applicability and is not limited by image color, dimension, and type. As long as the image information histogram can be extracted, the method can be used for threshold analysis. BRIEF DESCRIPTION OF THE DRAWINGS
[0052] Figure 1 Schematic diagram of Fourier transform principle.
[0053] Figure 2 Schematic diagram of the cusp mutation bifurcation set.
[0054] Figure 3 The following is a schematic diagram of the test process and results of the tensile adhesion test of TSL materials. Figure 3 (d) shows the mechanical curve and bonding area diagram.
[0055] From the analysis of the tensile bonding mechanics curve in the figure, it can be seen that the tensile bonding strength of the sample reached 0.628MPa and the elongation rate reached 195%. This phenomenon of low strength and high elongation rate occurs because the curing time is short and the moisture in the TSL material has not completely evaporated, resulting in strong ductility of the TSL. It is worth noting that the mechanical curve of the specimen shows a certain degree of strain hardening. The main reason is that when the TSL material is thin, cracks are prone to propagate inside the material, causing crystal defects in the material. At this time, the grains are elongated along the direction of maximum deformation, and the lattice is distorted, thereby improving the material's ability to resist deformation.
[0056] Figure 4 This is a statistical chart of the tensile bonding area density of TSL materials.
[0057] The software can identify each pixel in the image and divide all pixels into 0 (black) and 1 (white) according to the contrast between the rock color and the TSL material color. Figure 4 (b) shows that the TSL material with a thickness of 6 mm after curing for 14 days mainly remains on the sandstone surface in the form of small particles.
[0058] Figure 5 It is the amplitude and phase diagram after conversion based on IFFT method.
[0059] After obtaining the phase and amplitude information, the Gauss function fitting method is used to obtain the time domain curve, such as Figure 5 Next, we need to analyze and demonstrate the mutation of the bonding area information time domain curve to provide the prerequisite for the subsequent use of the continuous cusp mutation theory.
[0060] Figure 6 It is a time domain curve diagram of the bonding area diagram.
[0061] Figure 7 Schematic diagram of the mutation time series.
[0062] The solid line running horizontally through the figure is the Δ=0 straight line. The area above the straight line represents Δ>0, and the area below the straight line represents Δ<0. Therefore, when the time series passes through the straight line from top to bottom, it means that the energy has undergone a mutation, and the mutation point is the threshold.
[0063] Figure 8 Comparison chart of precise threshold effects. In the figure, the automatic threshold range is 210-247, and the calculated threshold range is 228-249.
[0064] Fig. 9 Flowchart of the mathematical model for calculating the threshold information of image binary segmentation. DETAILED DESCRIPTION
[0065] The image example uses the bonding area diagram of the tensile failure of the viscous material as an embodiment, and the viscous material in the embodiment is TSL865 material. The step numbers in the embodiment are only set for the convenience of explanation, and there is no limitation on the order between the steps. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.
[0066] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments:
[0067] This embodiment is an example of the method for improving image recognition and segmentation capabilities according to the present invention, comprising the following steps:
[0068] (1) The tensile bonding residual area map of TSL material was obtained according to the tensile bonding strength test of TSL material, and the bonding area map was identified using threshold segmentation software to obtain a discrete information histogram.
[0069] (2) At this point, it is necessary to establish the relationship between the discrete time signal and the discrete form of the bonding area density distribution. Assume that the white pixels of the same radius in the image are generated at the same time, and the times of generation of the white pixels of different radii are independent of each other. Then x(t), t∈[0,+∞) can be assumed to be a stationary random process of various states, which represents the time domain information of the bonding residual area during the formation process. Obviously, it cannot meet the absolute integrability condition: So x(t) does not have a Fourier transform, so we introduce an auxiliary function x T (t):
[0070]
[0071] Obviously x T (t) satisfies the absolute integrability condition and has a Fourier transform, namely:
[0072]
[0073] When T→+∞, x T (t)→x(t), referring to the inference of the power spectrum density function in random vibration theory, it can be known that the bonding area density function when x(t), t∈[0,+∞) is:
[0074]
[0075] Where the domain is f∈(-∞,+∞).
[0076] The bonding area density analyzed above is a double-sided spectral density, that is, both positive and negative values of f are defined. In engineering practice, since f<0 is meaningless, it is often calculated based on S x The even function property of (f) converts the bonding area density in the negative frequency range to the positive frequency range, thereby obtaining the unilateral spectral density G x (f), namely:
[0077]
[0078] So we have:
[0079]
[0080] The binary segmentation measurement accuracy set in the present invention is 100 / 344, that is, 344 pixels are used to represent a length of 100 μm, so T is finite, so G x (f) can be expressed as:
[0081]
[0082] In practice, the time interval, number of sampling points and total sampling time of the image sampling software are respectively: Δt, N, T (T = NΔt), then the frequency difference between two adjacent points of the discrete Fourier transform, that is, the frequency resolution is:
[0083]
[0084] According to the obtained frequency resolution, equation (4) is discretized:
[0085]
[0086] Considering have:
[0087]
[0088] For a discrete time domain signal x with N data m For (m=0,1,…,N-1), its discrete Fourier transform is also N data, and the frequency difference between two adjacent data is Δf, x k or Gx (k) The frequency corresponding to the point is Visible G x The value of k in (k) should be:
[0089] (3) According to step (2), theoretically, the bonding area density information can be assigned a time attribute, so the bonding area distribution information is subjected to an inverse Fourier transform. It should be noted that the present invention does not use a random phase method to generate a time domain curve, because the TSL tensile process is a stationary random process of various states, and the random phase cannot reflect the randomness of the material in the tensile process. After obtaining the phase and amplitude information, the Gauss function fitting method is used to obtain the time domain curve.
[0090] (4) Differentiate the time domain curve domain, perform a cusp mutation process in each time period, and obtain a complete mutation time series. The final threshold can be obtained according to the change trend of the mutation time series, such as Figure 7 As shown in the figure, point A is the threshold point, point B is the threshold end point, and a complete threshold range is defined between AB, which corresponds to the threshold range of the image segmentation software. The threshold range calculated in this example is 228-249, and the threshold range of the image segmentation software automatic segmentation is 210-247. Substitute the recalculated threshold point into the image recognition software for re-segmentation calculation, and compare the exported image with the original image. The result is as follows: Figure 8 shown.
Claims
1. A method for improving image recognition and segmentation capabilities, It is characterized in that The following steps are involved: (1) Extraction of image information: Obtain the image to be segmented, classify the image, import the image into the threshold segmentation software, and obtain the image discrete information histogram; (2) Establish the relationship between discrete time signals and discrete image information (histogram): 1) Introduce an auxiliary function, which is an even function and its domain is f∈(-∞,+∞); 2) Since the discrete information of the image is meaningless when the domain is f<0, according to the properties of the even function, the time series in the negative frequency range is converted to the positive frequency range to obtain the unilateral spectrum function; 3) According to the image threshold segmentation software, the image sampling time interval, the number of sampling points and the total sampling time information are converted into the frequency difference between two adjacent points of the discrete Fourier transform to obtain its frequency resolution; 4) Discretize the time series according to the obtained frequency resolution to obtain the time series curve of the image discrete information, complete the mathematical relationship between the image discrete time signal and the discrete form of image information, and prove that this type of image information meets the prerequisite of Fourier transform; (3) Extract discrete information based on the histogram and import it into Origin or Matlab software for inverse Fourier transform to obtain the amplitude and phase diagram of the time series. According to the image type, it can be proved whether the time series is a stationary random process of various states and whether a random phase needs to be used. (4) Perform mutation analysis on the obtained time series to reveal the intrinsic relationship between the discrete information of the image and the threshold. Mutation analysis is performed on the obtained time series. The Tate expansion is used once in each segment and substituted into the cusp mutation bifurcation set to accurately locate the position of the state parameter mutation. 1) In the cusp mutation model, there is one state variable x and two control variables u and v, and its standard potential function V(x) is: V(x)=x 4 +qx 2 +px (1) The equilibrium curve equation in the model satisfies V'(x)=0, and the singular point set equation satisfies V”(x)=0. Combining the two equations, we can get: Δ=8q 3 +27p 2 (2) 2) The equilibrium surface is divided into upper, middle and lower leaves. Assume that M(x, y, z) is a point representing the state of the system. When M moves along the upper and lower leaves, the system is in a stable state. When the motion trajectory of point M crosses the bifurcation set and is in the middle leaf of the equilibrium surface, the system undergoes a mutation. The above phenomenon is characterized by the eigenvalue V: V>0, no mutation (stable state); V = 0, critical state; V<0, mutation (unstable state); (3) 3) Since the obtained time domain curve is based on the IFFT method, its essence is a function based on a combination of multiple sine and cosine, and it does not have the properties of a complete time series. In order to obtain a complete time series that can be substituted into the cusp mutation bifurcation set, it is necessary to convert the time domain curve into a time series. First, an original sequence x is constructed based on the bonding area time domain curve. 0 : x 0 =(x 0 (1),x 0 (2),…x 0 (n)) (4) Superimposing equation (4) once, we get the AGO sequence, which is the one-dimensional time series ω corresponding to the bonding area. 1 (t): x 1 =(x 1 (1),x 1 (2),…x 1 (n)) (5) In formula (5): 4) Expand the obtained AGO sequence using Taylor's formula, retain the first 5 items of the sequence, and obtain ω 1 The approximate expression of (t) is: Formula (6) is a one-dimensional time series obtained by gray cumulative calculation of the bonding density time domain curve, that is, the bonding area density time series with complete attributes, where A 0 , A 1 , …, A 5 is the unknown coefficient, which is determined by polynomial fitting of the bonding density time domain curve. After derivation of formula (6), the restored bonding area density time series is obtained: 5) Let a 0 =A 1 ,a 1 =2A 2 ,a 2 =3A 3 ,a 3 =4A 4 ,a 4 =5A 5 , the standard potential function of equation (12) is constructed by using the parameter substitution method [21], when a 4 >0, let when a 4 <0, let q=a 3 / 4a 4 , we can get: when a 4 >0: V(z)=z 4 +uz 2 +vz+w (8) Where: when a 4 <0 hours: Then, by combining the cusp mutation theory (7), we can deduce: (5) Differentiate the image discrete information time series curve along the time axis, perform a mutation analysis in each time period, and obtain the curve of Δ changing with time series (hereinafter referred to as the mutation time series). Select the appropriate mutation point as the final threshold according to the change of the mutation time series. (6) After obtaining the precise final threshold, it is necessary to input the threshold into the threshold segmentation software and re-segment the image to verify the accuracy of the new threshold. The accuracy is evaluated in a variety of ways depending on the image type.
Citation Information
Patent Citations
Building extraction method and system for disaster video data and computer equipment
CN117333767A
Method for determining segmentation threshold of digital image of rock-soil material
US20210201497A1
