Image processing-based purple sand ware feature acquisition method and device and electronic equipment

Abstract

The present application relates to purple sand ware image analysis technical field, especially in kind based on image processing's purple sand ware feature collection method, device and electronic equipment, the method includes: the acquisition of multi-light source purple sand ware image data and spectrum data, and pretreatment is carried out;Purple sand ware image data is carried out filtering processing, to obtain filter image;Filter image is carried out key point detection, to extract stable key point, and stable key point is carried out direction distribution, to obtain main direction and auxiliary direction;Purple sand ware sample data is collected, and based on purple sand ware sample data, purple sand ware image data, stable key point, main direction and auxiliary direction, feature extraction model is constructed, to obtain purple sand ware feature;According to spectrum data, purple sand ware image data is carried out spectrum marking, to update feature extraction model.The present application realizes the accurate extraction of purple sand ware feature.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of purple sand ware image analysis, and in particular to a purple sand ware feature collection method and device based on image processing and electronic equipment. BACKGROUND

[0002] Traditional purple sand ware production process is unique, and its image contains rich cultural and technological information. By extracting features such as color, texture and shape through image processing technology, it can assist in identification, classification and digital protection, and promote the integration of traditional technology and modern technology.

[0003] In the prior art, the identification of purple sand ware relies on single modal data, which has the following problems: the image feature specificity is insufficient, the traditional multi-scale key point is easily disturbed by light and angle, and the unique texture characteristics of the purple sand material are not considered; the fixed parameters lack adaptability, the Gaussian blur sigma value, the key point filtering threshold and other parameters depend on experience setting; efficiency and accuracy are difficult to balance, traditional methods require manual intervention or complex calculation, and it is difficult to meet the real-time identification requirements. SUMMARY

[0004] The purpose of the present application is to provide a purple sand ware feature collection method and device based on image processing and electronic equipment to solve at least one of the problems in the prior art.

[0005] To achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0006] A purple sand ware feature collection method based on image processing, comprising:

[0007] Collecting multi-light source purple sand ware image data and spectral data and performing preprocessing;

[0008] Filtering the purple sand ware image data to obtain a filtered image;

[0009] Detecting key points on the filtered image to extract stable key points, and assigning directions to the stable key points to obtain main directions and auxiliary directions;

[0010] Collecting purple sand ware sample data, and constructing a feature extraction model based on the purple sand ware sample data, the purple sand ware image data, the stable key points, the main directions and the auxiliary directions to obtain purple sand ware features;

[0011] According to the spectral data, the purple sand ware image data is spectrally labeled to update the feature extraction model.

[0012] Further, the purple sand ware image data is divided into multiple groups of a1x a1 pixel regions, and the image local entropy of each region is calculated, and a1 represents a pixel region parameter;

[0013] The image local entropy is compared with an entropy threshold value, if the image local entropy is less than the entropy threshold value, it is determined that the region type of the current region is a low-entropy region, otherwise, it is determined that the region type of the current region is a high-entropy region.

[0014] Further, the image data of the purple sand ware is filtered using a Gaussian filter, if the region type is a low-entropy region, the Gaussian distribution standard deviation is set as σ = 1.2, if the region type is a high-entropy region, the Gaussian distribution standard deviation is set as σ = 2, the Gaussian kernel size is set according to the Gaussian distribution standard deviation, the Gaussian kernel size = 6 x Gaussian distribution standard deviation + 1, (x, y) is the pixel point coordinates in the region, and the purple sand ware image data after the Gaussian filtering processing is taken as a filtered image.

[0015] Further, a Gaussian pyramid is constructed according to the filtered image, the filtered image is down-sampled to 2 / 5 of the original purple sand ware image data as an initial image, the down-sampling rate is set as 1.2, the number of layers is set as 4, to generate multi-scale images, and a DOG pyramid is constructed based on the Gaussian pyramid;

[0016] In the scale space of the DOG pyramid, the gray values of each pixel point and 8 pixel points in its neighborhood and 9 pixel points corresponding to the upper and lower two layers are compared, and the pixel points of the maximum and minimum values of 26 pixel points are reserved as extreme points;

[0017] The clustering algorithm is used to process the extreme points, the extreme points are divided into a plurality of pixel point clusters, the average value of the coordinates of the extreme points in the pixel point cluster is taken as a centroid, if the distance between the extreme point in the pixel point cluster and the centroid is less than or equal to m, the current analysis extreme point is extracted as a stable key point, otherwise, the current analysis extreme point is not extracted, wherein m represents a distance threshold value.

[0018] Further, in a 20x20 pixel region around the stable key point, the horizontal response and the vertical response of each pixel point are calculated, the direction angle and the response amplitude of each pixel point are calculated, 360 / c intervals are obtained by dividing 360° around the stable key point into an interval every c°, the interval weight of each interval is set as the sum of the amplitudes corresponding to the direction angles in the angle range corresponding to the interval, and c represents a division angle parameter.

[0019] The central angle corresponding to the interval with the largest interval weight is taken as the main direction of the stable key point, if there is an interval with an interval weight greater than 80% of the interval weight of the main direction, the central angle of the current analysis interval is taken as the auxiliary direction of the stable key point.

[0020] Further, a lightweight convolutional neural network is used to extract feature vectors from the purple sand ware image data to output 128-dimensional texture features, 32 stable key points with the smallest distance from the centroid are extracted as coding key points, a SURF algorithm is used to process the coding key points to obtain 54-dimensional set features, and the texture features and the set features are spliced to obtain a 192-dimensional hybrid feature vector, and principal component analysis is used to reduce the hybrid feature vector to 50 dimensions to generate the purple sand ware features.

[0021] Further, the element fluorescence intensity is subjected to Min-Max normalization, and the element fluorescence intensity before normalization is compared with an element relative proportion threshold to perform spectral marking, so as to mark the corresponding pixel points in the purple sand ware image data as chemical pixel points and natural pixel points.

[0022] Further, the stable key points are removed according to the spectral marking, the stable key points with the spectral marking as chemical pixel points are removed, the feature extraction model is updated, the purple sand ware features extracted by the updated feature extraction model are taken as updated features, the cosine similarity between the updated features and the purple sand ware features extracted by the feature extraction model before updating is calculated, if the cosine similarity is greater than a similarity threshold, the purple sand ware features extracted by the feature extraction model before updating are retained, otherwise, the purple sand ware features extracted by the feature extraction model before updating are not retained.

[0023] On the other hand, the application also provides a purple sand ware feature acquisition device based on image processing, comprising:

[0024] The acquisition processing unit is used to acquire the purple sand ware image data and the spectral data of the multiple light sources, and to perform preprocessing;

[0025] The filtering processing unit is used to perform filtering processing on the purple sand ware image data to obtain a filtered image;

[0026] The key point analysis unit is used to perform key point detection on the filtered image to extract stable key points, and to assign directions to the stable key points to obtain main directions and auxiliary directions;

[0027] The feature extraction unit is used to acquire purple sand ware sample data, and to construct a feature extraction model based on the purple sand ware sample data, the purple sand ware image data, the stable key points, the main directions and the auxiliary directions to obtain purple sand ware features;

[0028] The model updating unit is used to perform spectral marking on the purple sand ware image data according to the spectral data to update the feature extraction model.

[0029] On the other hand, the application also provides an electronic device, comprising:

[0030] One or more processors;

[0031] a storage device configured to store one or more programs;

[0032] When the one or more programs are executed by the one or more processors, the one or more processors implement the image processing-based feature acquisition method of a bing-ware as described in any one of the above.

[0033] The present application has the following advantages: The present application combines multi-light-source image data and X-ray fluorescence spectrum data to realize multi-dimensional analysis of the surface texture and material composition of bing-ware, significantly improves the comprehensiveness and accuracy of feature acquisition, realizes efficient feature extraction and fusion through the combination of multi-scale key point detection, lightweight convolutional neural network and SURF algorithm, and further optimizes the feature dimension through principal component analysis, balances the calculation efficiency and feature discrimination, and applies adaptive preprocessing, regional type division, DBSCAN clustering and spectral marking, effectively suppresses noise and abnormal data interference, enhances the robustness of the scheme, reduces manual intervention through automated image processing and machine learning technology, improves the analysis efficiency, and provides an intelligent solution for image feature acquisition of bing-ware. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0035] Figure 1 The flow of the image processing-based feature acquisition method of bing-ware of the present embodiment.

[0036] Figure 2 The flowchart of the bing-ware image filtering processing method of the present embodiment.

[0037] Figure 3 The flowchart of the feature extraction model updating method of the present embodiment.

[0038] Figure 4 The structure diagram of the image processing-based feature acquisition device of bing-ware of the present embodiment.

[0039] Figure 5 The structure diagram of the electronic device in the present embodiment. DETAILED DESCRIPTION

[0040] In order to make the present application clearer, further description will be made to the present application with preferred embodiments and the accompanying drawings. Like numerals will be used to represent like elements throughout the accompanying drawings. It should be understood by those skilled in the art that the following description is illustrative rather than restrictive and should not be construed as limiting the scope of the present application.

[0041] It should be noted that although the terms first, second, third, etc. may be used in the embodiments of the present application, these descriptions should not be limited to these terms. These terms are only used to distinguish the descriptions. For example, the first can also be referred to as the second, and similarly, the second can also be referred to as the first without departing from the scope of the embodiments of the present application.

[0042] Please refer to Figure 1 As shown in the figure, the image processing-based Zisha pot feature acquisition method of the present embodiment includes:

[0043] Step S1, collect Zisha pot image data and spectral data of multiple light sources, and perform preprocessing, the Zisha pot image data of multiple light sources is the Zisha pot image taken at different incident angles of light sources, the incident angles include 0°, 45° and 90°, the incident angle is the included angle between the light source position and the vertical direction of the center of the Zisha pot, the Zisha pot image data is a gray-scale image, and the spectral data is an X-ray fluorescence spectrum, which is obtained by using a fluorescence spectrometer device, including a two-dimensional distribution map of elements such as iron, manganese and chromium, the data of which is aligned with each pixel point of the Zisha pot image data,

[0044] Specifically, in step S1 of the present embodiment, adaptive histogram equalization and non-local mean filtering are used to preprocess the Zisha pot image data. When the Zisha pot image data is processed by adaptive histogram equalization, the Zisha pot image is divided into 32x32 pixel sub-blocks to limit the contrast, and clipLimit is set to 2x sub-block pixel number / 256 to enhance the light and dark differences between Zisha particles and highlight the texture details. When the Zisha pot image data is processed by non-local mean filtering, the search window is set to 21x21 pixels, the similar window is set to 7x7 pixels, and the attenuation parameter is set to h=12 to suppress Gaussian noise while retaining the sharp edges of the kiln texture.

[0045] Specifically, the image data is collected by multiple light sources in step S1 of the present embodiment to comprehensively capture the surface texture and details of the Zisha pot, and the X-ray fluorescence spectral data is combined to realize the alignment of the material composition and the image data, providing multi-dimensional information support for subsequent analysis.

[0046] Please continue to refer to Figure 1 As shown in the figure, the image processing-based Zisha pot feature acquisition method further includes:

[0047] Step S2, filtering the image data of the Zisha pot to obtain a filtered image.

[0048] Specifically, in step S2, by analyzing the image local entropy and dividing the region type, different Gaussian filter parameters can be used for different texture complexity regions, which not only preserves sharp edges, but also smooths noise and optimizes image quality, laying a foundation for subsequent key point detection.

[0049] Please refer to Figure 2 As shown in the figure, it is a Zisha pot image filtering method, comprising:

[0050] Step S21, analyzing the image local entropy according to the Zisha pot image data, and judging the region type.

[0051] Specifically, in step S21, the Zisha pot image data is divided into a plurality of a1x a1 pixel regions, and the image local entropy of each region is calculated, and the expression of the image local entropy is: In the formula, E represents the image local entropy, i represents the gray level, P(i) represents the probability distribution of gray level i in a1x a1 pixel region, a1 represents the pixel region parameter, and 12≤a1≤20.

[0052] Specifically, in step S21, the image local entropy is compared with the entropy threshold value, if the image local entropy is less than the entropy threshold value, the region type of the current region is determined as a low entropy region, otherwise, the region type of the current region is determined as a high entropy region.

[0053] Specifically, in this embodiment, the entropy threshold value is set to 4, and in this embodiment, the value of the entropy threshold value is not specifically limited, and those skilled in the art can freely set it as long as it meets the judgment of the region type, and the setting of the entropy threshold value should meet the condition that it belongs to [3, 5].

[0054] Please continue to refer to Figure 2 As shown in the figure, the Zisha pot image filtering method further comprises:

[0055] Step S22, filtering the Zisha pot image data according to the region type to obtain a filtered image.

[0056] Specifically, in step S22, the Zisha pot image data is filtered using a Gaussian filter, if the region type is a low entropy region, the Gaussian distribution standard deviation is set to σ=1.2, if the region type is a high entropy region, the Gaussian distribution standard deviation is set to σ=2, the Gaussian kernel size is set according to the Gaussian distribution standard deviation, and the Gaussian kernel size=6x Gaussian distribution standard deviation+1, and the Gaussian formula is: (x, y) is the pixel point coordinate in the region, and the Zisha pot image data filtered by the Gaussian filter is used as the filtered image.

[0057] Please continue to refer to Figure 1 As shown, the image processing-based Zisha ware feature acquisition method further comprises:

[0058] In step S3, key points are detected from the filtered image to extract stable key points, and the stable key points are assigned with directions to obtain main directions and auxiliary directions.

[0059] Specifically, in step S3, a Gaussian pyramid is constructed according to the filtered image, and the filtered image is down-sampled to 2 / 5 of the original Zisha ware image data as an initial image, the down-sampling rate is set to 1.2, and the number of layers is set to 4 to generate multi-scale images.

[0060] Specifically, the multi-scale images in this embodiment are 4 layers of images in the Gaussian pyramid. For example, if the original Zisha ware image data is an image of 5000x5000 pixels, the initial image after down-sampling is an image of 2000x2000 pixels, the second layer image is an image of 1566x1566 pixels, the third layer image is an image of 1388x1388 pixels, and the fourth layer is an image of 1157x1157 pixels.

[0061] Specifically, in step S3, a DOG pyramid is constructed based on the Gaussian pyramid, and the DOG pyramid is obtained by subtracting two adjacent layers of multi-scale images in the Gaussian pyramid.

[0062] Specifically, in step S3, in the scale space of the DOG pyramid, the gray values of each pixel point and 8 pixel points in its neighborhood and 9 pixel points corresponding to the upper and lower two layers are compared, and the pixel points of the maximum and minimum values of 26 pixel points are reserved as extreme value points.

[0063] Specifically, in step S3, the DBSCAN clustering algorithm is used to process the extreme value points, the neighborhood radius ε is set to b1 pixels, the minimum sample number is set to b2, and the extreme value points are divided into multiple pixel point clusters to eliminate isolated noise points in the extreme value points, b1 represents a first clustering parameter, 4≤b1≤7, and b2 represents a second clustering parameter, 3≤b2≤5.

[0064] Specifically, in this embodiment, the values of the clustering parameters are not specifically limited, and a person skilled in the art can freely set them as long as the clustering processing of the extreme value points is satisfied. The best values of the clustering parameters are b1=5 and b2=3.

[0065] Specifically, in step S3, the average of the coordinates of the extreme points in the pixel cluster is calculated as the centroid, and if the distance between the extreme point in the pixel cluster and the centroid is less than or equal to m, the current analysis extreme point is extracted as a stable key point; otherwise, the current analysis extreme point is not extracted, where m represents a distance threshold, 1≤m≤3. It can be understood that the value of the distance threshold is not specifically limited in the embodiment, and can be freely set by a person skilled in the art, as long as the extraction of the stable key point is met, and the optimal value of the distance threshold is m=2.

[0066] Specifically, in step S3, a 20x20 pixel region around the stable key point is constructed, a Gaussian weight is applied to each pixel point in the region, and a Haar wavelet template is constructed, the Haar wavelet template including a horizontal template and a vertical template, the horizontal template having a width of 6 pixels, and the vertical template having a height of 6 pixels, the Haar wavelet template is slid in the 20x20 window, the horizontal response and the vertical response of each pixel point are calculated, the expression of the horizontal response is the expression of the vertical response is dx represents the horizontal response, dy represents the vertical response, and I(x, y) is the integral value at point (x, y).

[0067] Specifically, in step S3, the direction angle and the response amplitude of each pixel point are calculated based on the horizontal response and the vertical response, the expression of the direction angle is θ=arctan2(dy, dx), and the expression of the response amplitude is In the formula, θ represents the direction angle, and V represents the amplitude.

[0068] Specifically, in step S3, 360° around the stable key point is divided into intervals of c° to obtain 360 / c intervals, the interval weight of each interval is set to the sum of the amplitudes corresponding to the direction angles in the angle range corresponding to the interval, and c represents a division angle parameter.

[0069] Specifically, in the embodiment, the division angle parameter is set to 10, and the value of the division angle parameter is not specifically limited in the embodiment, and the division angle parameter should satisfy the factor of 360 to obtain an integer number of intervals, for example, if c=10, 36 intervals are obtained, including 0°-10°, 10°-20°, etc.

[0070] Specifically, in step S3, the center angle corresponding to the interval with the maximum interval weight is taken as the main direction of the stable key point, and if there is an interval with an interval weight greater than 80% of the interval weight of the main direction, the center angle of the current analysis interval is taken as the auxiliary direction of the stable key point.

[0071] Specifically, the center angle in this embodiment is the mean value of the angle range corresponding to the interval, for example, if the interval is 30°-40°, the corresponding center angle is 35°, and the auxiliary direction retains at most 3 intervals with the largest interval weight.

[0072] Specifically, in step S3 of this embodiment, multi-scale key point detection is realized by constructing Gaussian pyramid and DOG pyramid, which can capture texture features of different sizes, and combined with DBSCAN clustering algorithm and centroid distance screening, effectively eliminates isolated noise points, extracts stable key points, accurately determines the main direction and auxiliary direction of the key points through Haar wavelet template and direction assignment, and enhances the robustness of feature description.

[0073] Please continue to refer to Figure 1 As shown in the figure, the image processing-based purple sand ware feature acquisition method further includes:

[0074] Step S4, acquiring purple sand ware sample data, and constructing a feature extraction model based on the purple sand ware sample data, the purple sand ware image data, the stable key points, the main direction and the auxiliary direction to obtain the features of the purple sand ware, the purple sand ware sample data includes purple sand ware genuine image and purple sand ware fake image, and the acquisition mode of the purple sand ware sample data is user interactive upload.

[0075] Specifically, in step S4 of this embodiment, a lightweight convolutional neural network is used to extract feature vectors from the purple sand ware image data, the input layer size of the lightweight convolutional neural network is set to 224x224x1, and the gray value of each pixel point in the purple sand ware image is standardized to [-1, 1], the formula for standardization is: B=(B1-127.5) / 127.5, where B represents the standardized gray value, B1 represents the gray value before standardization, the convolution kernel size in the basic convolution layer in the lightweight convolutional neural network is set to 3x3, the channel number is set to 32, the step is set to 2, the padding is set to 1, the ReLU function is used as the activation function, and the output size is 112x112x32, the convolution kernel size in the depth separable convolution layer in the lightweight convolutional neural network is set to 3x3, the channel number is set to 54, the grouping is set to 54, the convolution kernel size of point-by-point convolution is set to 1x1, the channel number is set to 54, and the output size is 56x56x54, the global average pooling is used to take the mean value of each feature map to output a feature vector of 1x1x54, and the node number in the full connection layer is set to 128 to output a 128-dimensional texture feature.

[0076] Specifically, in step S4, the true image of the purple sand pot is collected as an anchor sample, the image data of the purple sand pot is collected as a positive sample, the fake image of the purple sand pot is collected as a negative sample, the loss function is set as L = max (d (a, p) - d (a, n) + a, 0), d (a, p) represents the Euclidean distance between the anchor sample and the positive sample, d (a, n) represents the Euclidean distance between the anchor sample and the negative sample, a represents the anchor sample, p represents the positive sample, and n represents the negative sample.

[0077] Specifically, in the embodiment, the interval parameter is set to 0.5, and the value of the interval parameter is not limited in the embodiment. The interval parameter is used to limit the minimum distance difference between the positive sample and the negative sample.

[0078] Specifically, in step S4, the 32 stable key points with the minimum distance to the centroid are extracted as the coding key points, the SURF algorithm is used to process the coding key points to obtain 54-dimensional set features, the texture features and the set features are spliced to obtain a 192-dimensional hybrid feature vector, and principal component analysis is used to reduce the hybrid feature vector to 50 dimensions to generate the features of the purple sand pot.

[0079] Specifically, in step S4, the lightweight convolutional neural network is used to extract the texture features, and the SURF algorithm is used to extract the set features, so that efficient and comprehensive feature fusion is realized, the feature redundancy is reduced through principal component analysis, and the calculation efficiency is improved.

[0080] Please continue to refer to Figure 1 As shown in the figure, the image processing-based feature acquisition method of the purple sand pot further comprises:

[0081] Step S5, performing spectral labeling on the image data of the purple sand pot according to the spectral data to update the feature extraction model.

[0082] Please refer to Figure 3 As shown in the figure, it is a feature extraction model updating method, comprising:

[0083] Step S51, performing spectral labeling on the image data of the purple sand pot according to the spectral data.

[0084] Specifically, in step S51, the element fluorescence intensity is subjected to Min-Max normalization, and the normalization formula is U (v) = [u (v) - u (v) min ] / [u (v) max -u (v) min ], wherein U (v) represents the normalized element fluorescence intensity, v represents the element, u (v) represents the element fluorescence intensity before normalization, u (v) min represents the minimum value of the element fluorescence intensity before normalization, and u (v)max represents the maximum value of the element fluorescence intensity before normalization.

[0085] Specifically, in step S51, the element fluorescence intensity before normalization is compared with the element relative proportion threshold to perform spectral marking. If G(v)(x, y) > g(v), the corresponding pixel point in the image data of the boccaro is marked as a chemical pixel point; otherwise, the corresponding pixel point in the image data of the boccaro is marked as a natural pixel point; wherein G(v)(x, y) represents the element proportion parameter, g(v) represents the element relative proportion threshold, and U(v)(x, y) represents the element fluorescence intensity of the pixel point after normalization.

[0086] Specifically, the elements in the embodiment include but are not limited to iron, manganese, chromium and the like. In the embodiment, the element relative proportion threshold of the iron element is set to 0.05, the element relative proportion threshold of the manganese element is set to 0.03, and the element relative proportion threshold of the chromium element is set to 0.02. The element relative proportion threshold is not specifically limited in the embodiment, and can be freely set by those skilled in the art.

[0087] Please continue to refer to Figure 3 As shown in the figure, the feature extraction model updating method further comprises:

[0088] Step S52, updating the feature extraction model according to the spectral marking.

[0089] Specifically, in step S52, the stable key points are removed according to the spectral marking. The stable key points with the chemical pixel points are removed to update the feature extraction model. The features extracted by the updated feature extraction model are taken as the updated features. The cosine similarity between the updated features and the features extracted by the feature extraction model before updating is calculated. If the cosine similarity is greater than the similarity threshold, the features extracted by the feature extraction model before updating are retained. Otherwise, the features extracted by the feature extraction model before updating are not retained.

[0090] Specifically, in the embodiment, the similarity threshold is set to 0.7. The setting of the similarity threshold is not specifically limited in the embodiment, and can be freely set by those skilled in the art. The setting of the similarity threshold should meet [0.7, 0.9].

[0091] Specifically, in step S5, the stable key points with material abnormalities can be identified and removed through spectral marking, which improves the accuracy of the features. The cosine similarity comparison ensures the rationality of the model updating, avoids invalid updating, and further optimizes the performance of the feature extraction model.

[0092] Please refer to Figure 4As shown, it is the image processing-based Yixi pot feature acquisition device of the embodiment, including

[0093] The acquisition processing unit is used for acquiring Yixi pot image data and spectrum data of multiple light sources and performing preprocessing;

[0094] The filtering processing unit is used for filtering Yixi pot image data to obtain a filtered image;

[0095] The key point analysis unit is used for detecting key points of the filtered image to extract stable key points and assigning directions to the stable key points to obtain main directions and auxiliary directions;

[0096] The feature extraction unit is used for acquiring Yixi pot sample data and constructing a feature extraction model based on the Yixi pot sample data, Yixi pot image data, stable key points, main directions and auxiliary directions to obtain Yixi pot features;

[0097] The model updating unit is used for performing spectrum labeling on Yixi pot image data according to spectrum data to update the feature extraction model.

[0098] Please refer to Figure 5 As shown, it is a structural schematic diagram of an electronic device in the embodiment, and the electronic device 50 in the embodiment can include but is not limited to mobile terminals such as mobile phones, notebook computers, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablets), PMPs (portable multimedia players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), wearable electronic devices and the like, and fixed terminals such as digital TVs, desktop computers, smart home devices and the like. Figure 5 The electronic device shown is only an example, and should not bring any limitation to the functions and use range of the embodiment of the present application.

[0099] As Figure 5As shown, the electronic device 50 can include a processing device (e.g., a central processor, a graphics processor, etc.) 51 that can perform various appropriate actions and processes to implement the methods of the embodiments as described in the present application according to programs stored in a read-only memory (ROM) 52 or loaded into a random access memory (RAM) 53 from a storage device 58. Various programs and data required by the electronic device 50 to operate are also stored in the RAM 53. The processing device 51, the ROM 52, and the RAM 53 are connected to each other through a bus 54. An input / output (I / O) interface 55 is also connected to the bus 54. Generally, the following devices can be connected to the I / O interface 55: input devices 56 including, for example, a touch screen, a touch pad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; output devices 57 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; storage devices 58 including, for example, a magnetic tape, a hard disk, etc.; and communication devices 59. The communication devices 59 can allow the electronic device 50 to communicate with other devices wirelessly or through wires to exchange data. Although Figure 5 The electronic device 50 is shown with various devices, but it should be understood that not all of the shown devices are required to be implemented or present. More or fewer devices can alternatively be implemented or present.

[0100] In particular, the processes described above with reference to the flowcharts can be implemented as a computer software program according to embodiments of the present application. For example, embodiments of the present application include a computer program product comprising a computer program carried on a non-transitory computer readable medium, the computer program containing program code for performing the methods shown in the flowcharts, thereby implementing the methods as described above. In such embodiments, the computer program can be downloaded and installed from a network through the communication devices 59, or installed from the storage devices 58, or installed from the ROM 52. When the computer program is executed by the processing device 51, the above-described functions defined in the methods of the embodiments of the present application are performed.

[0101] In particular, the computer readable medium described in the present embodiments can be included in the above-described electronic device; or can exist separately without being assembled into the electronic device.

[0102] Specifically, the computer readable medium in the embodiment carries one or more programs, when the one or more programs are executed by the electronic device, the electronic device is caused to: collect image data and spectral data of a multi-light source purple sand ware, and perform preprocessing; filter the image data of the purple sand ware to obtain a filtered image; perform key point detection on the filtered image to extract stable key points, and perform direction assignment on the stable key points to obtain a main direction and an auxiliary direction; collect sample data of the purple sand ware, and construct a feature extraction model based on the sample data of the purple sand ware, the image data of the purple sand ware, the stable key points, the main direction and the auxiliary direction to obtain features of the purple sand ware; and perform spectral labeling on the image data of the purple sand ware according to the spectral data to update the feature extraction model.

[0103] Obviously, the above embodiments of the present application are only examples for clearly illustrating the present application, and are not intended to limit the implementation manners of the present application. For those skilled in the art, on the basis of the above description, other different forms of changes or variations can also be made, and it is impossible to enumerate all the implementation manners here. Any obvious changes or variations derived from the technical solutions of the present application still fall within the protection scope of the present application.

Claims

1. A method for acquiring features of Zisha teapots based on image processing, characterized in that, include: Collect image and spectral data of Zisha teapots from multiple light sources and perform preprocessing; The image data of the Zisha teapot is filtered to obtain a filtered image; Key point detection is performed on the filtered image to extract stable key points, and orientation assignment is performed on the stable key points to obtain the main orientation and the auxiliary orientation; Collect Zisha teapot sample data, and construct a feature extraction model based on Zisha teapot sample data, Zisha teapot image data, stable key points, main direction and auxiliary direction to obtain Zisha teapot features; The image data of Zisha teapots are spectrally labeled based on spectral data in order to update the feature extraction model; The feature extraction model update method includes: performing Min-Max normalization on the element fluorescence intensity, comparing the element fluorescence intensity before normalization with the element relative proportion threshold, and performing spectral labeling to label the corresponding pixels in the Zisha pottery image data as chemical pixels and natural pixels. Based on spectral marking, stable key points are removed. Stable key points marked as chemical pixels are removed to update the feature extraction model. The Zisha pottery features extracted by the updated feature extraction model are used as the updated features. The cosine similarity between the updated features and the Zisha pottery features extracted by the original feature extraction model is calculated. If the cosine similarity is greater than the similarity threshold, the Zisha pottery features extracted by the original feature extraction model are retained; otherwise, the Zisha pottery features extracted by the original feature extraction model are not retained.

2. The image processing-based method for acquiring features of Zisha teapots according to claim 1, characterized in that, The image data of Zisha teapots is divided into multiple a1×a1 pixel regions, and the local entropy of each region is calculated, where a1 represents the pixel region parameter. The local entropy of the image is compared with the entropy threshold. If the local entropy is less than the entropy threshold, the current region is determined to be a low-entropy region; otherwise, the current region is determined to be a high-entropy region.

3. The method for acquiring features of Zisha teapots based on image processing according to claim 2, characterized in that, A Gaussian filter is used to filter the Zisha teapot image data. If the region type is a low-entropy region, the standard deviation of the Gaussian distribution is set to σ=1.

2. If the region type is a high-entropy region, the standard deviation of the Gaussian distribution is set to σ=2. The Gaussian kernel size is set according to the standard deviation of the Gaussian distribution. The Gaussian kernel size = 6 × standard deviation of the Gaussian distribution + 1. (x,y) are the pixel coordinates within the region. The Zisha teapot image data after Gaussian filtering is used as the filtered image.

4. The image processing-based method for acquiring features of Zisha teapots according to claim 3, characterized in that, A Gaussian pyramid was constructed based on the filtered image. The filtered image was downsampled to 2 / 5 of the original Zisha pottery image data as the initial image. The downsampling rate was set to 1.2 and the number of layers was 4 to generate multi-scale images. A DOG pyramid was then constructed based on the Gaussian pyramid. In the DOG pyramid scale space, the gray values ​​of each pixel are compared with the gray values ​​of its 8 neighboring pixels and the 9 corresponding pixels in the two layers above and below it, and the pixels with the maximum and minimum values ​​in the middle of the 26 pixels are retained as extreme points. Clustering algorithms are used to process extreme points, dividing them into multiple pixel clusters. The average coordinates of each extreme point in a pixel cluster are calculated as the centroid. If the distance between the extreme point in a pixel cluster and the centroid is less than or equal to m, the current extreme point is extracted as a stable key point. Conversely, the extreme points of the current analysis are not extracted, where m represents the distance threshold.

5. The image processing-based method for acquiring features of Zisha teapots according to claim 4, characterized in that, In a 20×20 pixel region around the stable keypoint, calculate the horizontal and vertical responses of each pixel. For the horizontal and vertical responses of each pixel, calculate the orientation angle and response amplitude. Divide the 360° region around the stable keypoint into intervals of c° to obtain 360 / c intervals. Set the interval weight of each interval to the sum of the amplitudes of the orientation angles within the corresponding angle range of the interval, where c represents the division angle parameter. The center angle corresponding to the interval with the largest interval weight is taken as the main direction of the stable key point. If there is an interval whose interval weight is greater than the interval weight of the main direction multiplied by 80%, then the center angle of the current analysis interval is taken as the auxiliary direction of the stable key point.

6. The image processing-based method for acquiring features of Zisha teapots according to claim 5, characterized in that, A lightweight convolutional neural network was used to extract feature vectors from the Zisha teapot image data to output 128-dimensional texture features. The 32 stable keypoints with the smallest distance from the centroid were extracted as encoding keypoints. The SURF algorithm was used to process the encoding keypoints to obtain 54-dimensional ensemble features. The texture features and ensemble features were then concatenated to obtain a 192-dimensional hybrid feature vector. Principal component analysis was used to reduce the dimensionality of the hybrid feature vector to 50 dimensions to generate Zisha teapot features.

7. A Zisha teapot feature acquisition device based on image processing, applied to the Zisha teapot feature acquisition method based on image processing as described in any one of claims 1-6, characterized in that, include: The acquisition and processing unit is used to acquire image data and spectral data of Zisha teapots from multiple light sources and to perform preprocessing. The filtering processing unit is used to filter the Zisha teapot image data to obtain a filtered image; The key point analysis unit is used to detect key points in the filtered image to extract stable key points and assign directions to the stable key points to obtain the main direction and the auxiliary direction. The feature extraction unit is used to collect Zisha teapot sample data and construct a feature extraction model based on the Zisha teapot sample data, Zisha teapot image data, stable key points, main direction and auxiliary direction to obtain the features of Zisha teapot. The model update unit is used to perform spectral labeling on the Zisha teapot image data based on spectral data in order to update the feature extraction model.

8. An electronic device, characterized in that, The electronic device includes: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the image processing-based Zisha teapot feature acquisition method as described in any one of claims 1-6.

Images