Blue-and-white porcelain feature acquisition device and breaking method

Abstract

The invention discloses a blue and white porcelain feature acquisition device and a breaking method. The acquisition device comprises a storage box, a dimming glass cover, a light source, a rotary support mounting table, a camera, a telescopic rod and an adjustable clamping mechanism, the method comprises the following steps: 1, constructing a blue and white porcelain pattern image database; 2, constructing a blue and white porcelain pattern declaration model; 3, training the blue and white porcelain pattern generation model; 4, testing the blue and white porcelain pattern generation model; 5, acquiring an actual blue and white porcelain pattern image; 6, preprocessing the actual blue and white porcelain pattern image; and 7, breaking the blue and white porcelain patterns. The method adapts to acquisition of porcelain images of different sizes and types, analyzes blue and white porcelain ornamentation through multi-feature fusion and a gradient lifting model, introduces a ShuffleNetV2 network model of a space attention mechanism to capture a long-distance dependency relationship in the image, and fuses a color threshold segmentation technology to extract a complete ornamentation pattern on a glaze surface, so as to obtain a high-precision image of the blue and white porcelain ornamentation. And unrelated interference information is successfully filtered.

Description

Technical Field

[0001] The invention belongs to the technical field of dating blue and white porcelain patterns, and specifically relates to a blue and white porcelain feature collection device and dating method. Background Art

[0002] In the current field of porcelain display and porcelain photo collection, the commonly used display methods mainly include static display stands, display cabinets, etc., and porcelain photos are mostly collected manually or with fixed cameras, which cannot be adaptively adjusted. First, static display stands and display cabinets can usually only display a single side of the porcelain, and the audience cannot fully appreciate the beauty of the porcelain, and cannot meet the display needs at different angles, and the display function is insufficient; second, the height of most display stands and display cabinets is fixed, and cannot be flexibly adjusted according to the size and display needs of the porcelain, and cannot adapt to porcelain of different sizes and types, which reduces the display effect, and there is a lack of effective fixing methods during the adjustment of porcelain; third, porcelain is a valuable item, and its storage needs to be concealed, confidential, and safe and anti-theft.

[0003] In addition, as one of the treasures of China's traditional cultural heritage, blue-and-white porcelain carries rich historical and cultural connotations, but the current research on its age identification using visual intelligence technology is relatively scarce. Therefore, it is of far-reaching significance to use machine vision dating methods to explore the cultural and historical background behind it. Dating blue-and-white porcelain usually requires experienced experts to rely on their own profound knowledge of porcelain appreciation to identify porcelain. With the development of modern scientific and technological instruments, scientific instruments can also be used to scan chemical components for approximate age judgment. Expert experience identification relies on the experience of identification experts in blue-and-white porcelain research for many years. Through the surface feel and visual perception of porcelain, the overall blue-and-white porcelain is roughly judged and analyzed to achieve the identification of its unearthed age. In principle, this dating method is to study and study the blue-and-white porcelain cultural relics that have been unearthed and have accurate data labels by comparing them with each other, so as to accumulate a lot of appreciation experience and form the expert's own dating experience and skills. Experts in the identification of blue and white porcelain have summarized the essence of the research on the overall age characteristics of porcelain by their research on porcelain for many years. Through contact with a large number of porcelains, they have proposed that the five characteristics of porcelain shape, glaze material, body soil, underglaze decoration and bottom text be used as the basis for judging the age of blue and white porcelain, and have made detailed research for the identification of blue and white porcelain. While the expert experience identification method is recognized, there are also some problems. The differences in experience, knowledge and moral quality of different appraisers may lead to different identification conclusions, and this dating method is difficult to describe the various characteristic attributes of blue and white porcelain decoration through quantitative language. The decoration of blue and white porcelain is complex and diverse, and the age characteristics are complex and changeable. It is difficult to grasp the identification elements of blue and white porcelain. Scientific and technological instrument identification uses scanning instruments, such as X-ray fluorescence analysis, to identify the chemical composition of porcelain and roughly determine the age range. As time goes by, the surface of some blue and white porcelain is severely weathered, resulting in inaccurate instrument judgment. And dating methods such as XRF will cause secondary damage to blue and white porcelain and destroy cultural relics, and this value is difficult to estimate. The first two dating methods each have their own drawbacks, fail to meet the expected requirements, and are unable to solve the problem of non-destructive testing of blue and white porcelain very well. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a blue and white porcelain feature collection device in view of the deficiencies in the above-mentioned prior art. The device has a novel and reasonable design and can realize 360-degree rotation of the porcelain through a slewing support platform to meet the audience's viewing needs at different angles. The height of the porcelain is adjusted by a first telescopic rod, the extension height of the second telescopic rod is adjusted according to the clamping height of the porcelain, and the extension length of the third telescopic rod is adjusted according to the clamping width of the porcelain. The device can adapt to porcelains of different sizes and types and has a good display effect. The dimming glass cover is converted between transparent and non-transparent to realize concealed and confidential storage of the porcelain, adding a barrier to the problem of theft. The porcelain is illuminated by a light source and an adjustable camera is used to collect porcelain images, which is convenient for promotion and use.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a blue and white porcelain feature collection device, characterized in that: it includes a storage box and a dimming glass cover arranged on the upper part of the storage box, a first telescopic rod is installed at the bottom of the storage box, the top telescopic end of the first telescopic rod is fixedly connected to the bottom of a swivel support mounting table for installing a swivel support platform, an adsorption pad for buffering and adsorbing porcelain is arranged on the top of the swivel support platform, a pair of adjustable clamping mechanisms for clamping and fixing porcelain are symmetrically arranged on both sides of the swivel support mounting table, and a fourth telescopic rod for adjusting the position of a camera, the adjustable clamping mechanism includes a second telescopic rod installed on the swivel support mounting table, a mounting block is fixedly installed at the top telescopic end of the second telescopic rod, a third telescopic rod is horizontally installed on the side of the mounting block facing the porcelain, a splint is fixedly installed on the horizontal telescopic end of the third telescopic rod, a buffer layer is arranged on the side of the splint facing the porcelain, a camera holder is installed on the top of the fourth telescopic rod, and light sources for illuminating the porcelain are arranged on the top and sides of the dimming glass cover.

[0006] The above-mentioned blue and white porcelain feature collection device is characterized in that: the adsorption pad is a soft glue adsorption pad, the buffer layer is a memory foam buffer layer, the storage box is a non-transparent storage box, and a control terminal is arranged in the non-transparent storage box.

[0007] The above-mentioned blue and white porcelain feature collection device is characterized in that the first telescopic rod, the second telescopic rod, the third telescopic rod and the fourth telescopic rod are all electric telescopic rods.

[0008] The above-mentioned device for collecting characteristics of blue and white porcelain is characterized in that a pressure sensor is also provided on the side of the clamping plate facing the porcelain.

[0009] The above-mentioned blue and white porcelain feature collection device is characterized in that: the height of the storage box is 0.5m to 1.2m, the height of the dimming glass cover is 1m to 2m, and the light source is a warm light lamp.

[0010] At the same time, the present invention also discloses a method for extracting and fusing directional gradient histogram, HSV color histogram and local binary pattern features, introducing unsharp mask processing to enhance the texture features of the decoration, and using a gradient boosting model to perform dating analysis on the blue and white porcelain decoration, introducing a ShuffleNetV2 network model with a LAM spatial attention mechanism to capture long-distance dependencies in the image and enhance the expression ability of the model. For the decoration, the color threshold segmentation technology is integrated to effectively extract the complete decoration pattern on the glaze surface, and irrelevant interference information is successfully filtered out. After comprehensive improvement of the model, a method for dating blue and white porcelain with blue and white porcelain decoration has a good dating effect, and the dating method comprises the following steps:

[0011] Step 1: construct a blue and white porcelain decoration image database: randomly classify the blue and white porcelain decoration image database to obtain an image training data set and an image test data set;

[0012] Step 2: Construct a dating model for blue and white porcelain decoration: The blue and white porcelain decoration dating model includes a ShuffleNetV2 network model that introduces the LAM spatial attention mechanism, and the ShuffleNetV2 network model that introduces the LAM spatial attention mechanism is then subjected to a maximum pooling layer, global downsampling, and connected to a fully connected layer for category prediction;

[0013] Step 3: Train the blue-and-white porcelain pattern dating model: extract the enhanced blue-and-white porcelain pattern images from the image training data set and input them into the blue-and-white porcelain pattern dating model. Use the transfer learning strategy, RMSProp optimizer and cross entropy loss function to calculate the category loss, and output the category prediction results with category probabilities until the blue-and-white porcelain pattern dating model converges, thus completing the training process of the blue-and-white porcelain pattern dating model.

[0014] Step 4: Testing the blue-and-white porcelain pattern dating model: Retrieving blue-and-white porcelain pattern sample images from the image test data set, testing the blue-and-white porcelain pattern dating model until the test is completed, and obtaining a trained and tested blue-and-white porcelain pattern dating model;

[0015] Step 5: using a blue and white porcelain feature acquisition device to obtain an actual blue and white porcelain pattern image;

[0016] Step 6: Preprocess the actual blue and white porcelain pattern image. The process is as follows:

[0017] Step 601, scaling the actual blue and white porcelain decoration image to a size of 224×224, and then graying it;

[0018] Step 602, unsharp mask processing: use a low-pass filter to perform low-frequency filtering on the actual blue and white porcelain decoration image after grayscale processing to remove high-frequency part information; then subtract the low-frequency component image from the actual blue and white porcelain decoration image after grayscale processing to obtain a high-frequency change image, perform detail magnification processing on the high-frequency change image through different weight coefficients, overlap the high-frequency change image with magnified details and the actual blue and white porcelain decoration image after grayscale processing to obtain an unsharp mask enhanced image with more prominent edge and contour shapes of the image;

[0019] Step 603, HOG edge feature extraction: perform equal area segmentation on the unsharp mask enhanced image, with 8*8 pixels as a unit and 2*2 pixels as a region, divide the image into multiple regions, calculate the pixel gradient of each region, extract the undirected HOG gradient, divide the gradient direction into 9 intervals on average, and generate 9 histograms of corresponding intervals;

[0020] Step 604, LBP texture feature extraction: By comparing the grayscale values ​​of each single pixel in the image with the surrounding pixels, a binary code is formed, and the spatial distribution characteristics of these codes are counted to form a feature vector. The LBP operator operates in a 3×3 neighborhood, with the original single point grayscale value as the reference mark, and then the size is determined by the grayscale values ​​of the center pixel and the eight surrounding pixels. Finally, a single 8-bit binary number is generated by stacking. If the grayscale value of the field pixel is higher than the original single point pixel, the binary number of the digital will be set to 01. On the contrary, if it is lower than the center pixel, the corresponding binary number will be set to 00. Through a complete calculation, each 3×3 pixel block will be converted into a unique 8-bit binary code, which is used to characterize the texture feature signal of the area, and finally the local binary value corresponding to the original single pixel value is obtained through cascade arrangement.

[0021] Step 605, HSV color histogram processing: by dividing the color space of the image into a series of unrelated single intervals, and counting the total number of pixels in each interval, a color histogram is constructed; in the processing, different color channels of the image are processed separately to obtain the distribution characteristics of the color in each channel, and a three-dimensional or one-dimensional histogram is formed by counting all color channels to intuitively represent the color distribution of the image. The three-dimensional or one-dimensional histogram uses color values ​​as abscissas, ranging from 0 to 255, and corresponds to hue H information, lightness V information and saturation S information in the color space. The three-dimensional or one-dimensional histogram uses probability as a ordinate, and through normalization processing, the total probability is 1;

[0022] Step 606: Fusion of multiple features from step 603 to step 605;

[0023] Step 607, gradient boosting: using the gradient boosting model to improve the ability of the fused multi-feature weak learning machine to a strong learner;

[0024] Step 7: Dating blue and white porcelain decorations. The process is as follows:

[0025] Step 701: Use color threshold segmentation technology to extract the main decorative pattern of blue and white porcelain, and shield the background and edge areas to avoid generating additional interference information;

[0026] Step 702: Input the main blue and white porcelain pattern image extracted in step 701 into the trained blue and white porcelain pattern dating model, and output the probability information of the age of the blue and white porcelain pattern result.

[0027] The above-mentioned method for dating blue and white porcelain by features is characterized in that: in step one, the blue and white porcelain decoration image is processed by random cropping and resizing, horizontal flipping, random rotation, and color perturbation to increase the diversity of the data; the blue and white porcelain decoration image includes blue and white porcelain decoration images of multiple dynasties, and the blue and white porcelain decoration includes figure blue and white porcelain decoration and dragon blue and white porcelain decoration.

[0028] The above-mentioned method for dating blue and white porcelain features is characterized in that: in step 2, the ShuffleNetV2 network model that introduces the LAM spatial attention mechanism includes a first convolutional layer, a first maximum pooling layer and a LAM spatial attention mechanism, the first maximum pooling layer performs downsampling to obtain a first feature map with a reduced size, the first feature map passes through Block-Conv1, Block-Conv2, Block-Conv3 and the second convolutional layer in sequence to output a second feature map, the second feature map introduces the LAM spatial attention mechanism and is input to the second maximum pooling layer, and finally is globally downsampled through a convolution kernel and connected to the fully connected layer for dating prediction; the Block-Conv1 performs one operation with a step size of 2 and three operations with a step size of 1; the Block-Conv2 performs one operation with a step size of 2 and seven operations with a step size of 1; the Block-Conv2 performs one operation with a step size of 2 and three operations with a step size of 1.

[0029] The above-mentioned method for dating blue and white porcelain by features is characterized in that: the size of the actual blue and white porcelain decoration image is 224×224×3, and after the color threshold segmentation technology is introduced and processed by 24 3×3 convolution kernels, the size of the main blue and white porcelain decoration image becomes 112×112×24, and the first maximum pooling layer performs downsampling to obtain a first feature map with a size reduced to 56×56×24, the size of the image after Block-Conv1 processing is 28×28×116, the size of the image after Block-Conv2 processing is 14×14×232, and the size of the image after Block-Conv3 processing is 7×7×464, the second convolution layer includes 1024 1×1 convolution kernels, the second feature map introduces the LAM spatial attention mechanism and is input into the second maximum pooling layer, the output image size is 7×7×1024, and finally a 7×7 convolution kernel is used for global downsampling to output the image size of 1×1×1024, and the fully connected layer is connected for dating prediction.

[0030] The above-mentioned method for dating blue and white porcelain by features is characterized in that: in step 701, the color threshold segmentation technology is to divide the colors in the image into different categories by setting the color range. The colors of the blue and white porcelain include blue and white. The blue patterns of the blue and white porcelain are separated from the white background by setting the corresponding color threshold to achieve preliminary extraction of the patterns. The blue patterns of different shades are extracted by adjusting the threshold. The adjustment of the color threshold captures the details more accurately, thereby obtaining a clearer and more complete pattern image.

[0031] Compared with the prior art, the present invention has the following advantages:

[0032] 1. The acquisition device adopted by the present invention can realize 360-degree rotation of the porcelain through a slewing support platform to meet the audience's viewing needs at different angles. The height of the porcelain is adjusted by the first telescopic rod, the extension height of the second telescopic rod is adjusted according to the clamping height of the porcelain, and the extension length of the third telescopic rod is adjusted according to the clamping width of the porcelain. It can adapt to porcelains of different sizes and types and has a good display effect. The dimming glass cover is converted between transparent and non-transparent to realize the concealed and confidential storage of porcelain, adding a barrier to the problem of theft. The light source is used to illuminate the porcelain, and an adjustable camera is used to collect porcelain images. It has the advantages of simple operation, strong applicability, and high safety.

[0033] 2. Most of the blue and white porcelain patterns in the present invention are relatively complex. By extracting and fusing the directional gradient histogram, HSV color histogram and local binary pattern features, an unsharp masking process is introduced to enhance the texture features of the patterns, enhance the sensitivity of the model, and ensure accurate learning of the features of complex patterns. The LBP feature is mainly used to extract the main vein texture in the patterns, and can effectively extract all the line information of blue and white porcelain. The HOG feature is used to extract the edge shape of the blue and white porcelain patterns, and mainly describes the contour shape information of the image through gradient calculation. The HSV color histogram describes the layout of each color in the blue and white porcelain patterns, and provides richer information for the color space features of the pattern image, and fully understands the surface characteristics of the blue and white porcelain pattern image. Through the fusion of the three multi-information features, the comprehensive extraction of the blue and white porcelain pattern features can be achieved. The color threshold segmentation technology is integrated to effectively extract the complete pattern on the glaze surface, and successfully filter out irrelevant interference information, avoid incorrect identification results, and facilitate popularization and use.

[0034] 3. The present invention uses the gradient boosting model to enhance the capability of the fused multi-feature weak learning machine to a strong learner. The gradient boosting model performs better in processing small data set classification problems. Its advantages are mainly reflected in processing nonlinear relationships, robustness, no need for complex parameter adjustment, adaptation to different types of features, and processing small sample data. It is very suitable for use in blue and white porcelain classification tasks.

[0035] 4. The dating method of the present invention has simple steps. It is based on the ShuffleNetV2 network model and introduces a spatial attention mechanism to improve the original network ShuffleNetV2 to capture long-distance dependencies in the image and enhance the expressiveness of the model. It also uses a transfer learning strategy, RMSProp optimizer and cross-entropy loss function to perform experimental verification on a blue and white porcelain dragon pattern dataset for accurate dating.

[0036] In summary, the design of the present invention is novel and reasonable. By extracting and fusing the directional gradient histogram, HSV color histogram and local binary pattern features, the unsharp masking process is introduced to enhance the texture features of the decoration, and the gradient boosting model is used to perform dating analysis on the blue and white porcelain decoration. The ShuffleNetV2 network model with the LAM spatial attention mechanism is introduced to capture the long-distance dependencies in the image and enhance the expressive ability of the model. For the decoration, the color threshold segmentation technology is integrated to effectively extract the complete decoration pattern on the glaze surface, and successfully filter out irrelevant interference information. After the comprehensive improvement of the model, the dating effect of the blue and white porcelain decoration is good, which is convenient for popularization and use.

[0037] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a schematic diagram of the structure of the collection device used in the present invention.

[0039] Figure 2 It is a flowchart of the dating method of the present invention.

[0040] Figure 3 It is a grayscale image of the comparison result of the unsharp mask processing in an embodiment of the present invention.

[0041] Figure 4 This is a grayscale image of the HOG edge feature extraction comparison result in an embodiment of the present invention.

[0042] Figure 5 This is a grayscale image of the LBP texture feature extraction comparison result in an embodiment of the present invention.

[0043] Figure 6 It is a grayscale image of the HSV color histogram processing comparison result in an embodiment of the present invention.

[0044] Figure 7 It is a feature map of 5 hidden layers in an embodiment of the present invention.

[0045] Figure 8 This is a grayscale image of the comparison result of Example 1 of the color threshold segmentation result in an embodiment of the present invention.

[0046] Fig. 9This is a grayscale image of the comparison result of Example 2 of the color threshold segmentation result in an embodiment of the present invention.

[0047] Description of reference numerals:

[0048] 1—storage box; 2—first telescopic rod; 3—slewing support mounting platform;

[0049] 4—slewing support platform; 5—adsorption pad; 6—porcelain;

[0050] 7—second telescopic rod; 8—mounting block; 9—third telescopic rod;

[0051] 10—plywood; 11—buffer layer; 12—dimming glass cover;

[0052] 13—fourth telescopic rod; 14—camera base; 15—camera;

[0053] 16—Light source. DETAILED DESCRIPTION

[0054] like Figure 1 As shown, a blue and white porcelain feature collection device of the present invention comprises a storage box 1 and a dimming glass cover 12 arranged on the upper part of the storage box 1, a first telescopic rod 2 is installed at the bottom of the storage box 1, the top telescopic end of the first telescopic rod 2 is fixedly connected to the bottom of a swivel support mounting table 3 for mounting a swivel support platform 4, an adsorption pad 5 for buffering and adsorbing porcelain 6 is arranged on the top of the swivel support mounting table 4, a pair of adjustable clamping mechanisms for clamping and fixing porcelain 6 and a camera 15 are symmetrically arranged on both sides of the swivel support mounting table 3. The fourth telescopic rod 13 is in the position, and the adjustable clamping mechanism includes a second telescopic rod 7 installed on the rotary support mounting table 3, the top telescopic end of the second telescopic rod 7 is fixedly installed with a mounting block 8, the side of the mounting block 8 facing the porcelain 6 is horizontally installed with a third telescopic rod 9, the horizontal telescopic end of the third telescopic rod 9 is fixedly installed with a clamping plate 10, the side of the clamping plate 10 facing the porcelain 6 is provided with a buffer layer 11, the top of the fourth telescopic rod 13 is installed with a camera base 14, and the top and side of the dimming glass cover 12 are provided with a light source 16 for lighting the porcelain 6.

[0055] It should be noted that the 360-degree rotation of the porcelain can be achieved through the slewing support platform to meet the audience's viewing needs at different angles. The height of the porcelain can be adjusted by the first telescopic rod, the extension height of the second telescopic rod can be adjusted according to the clamping height of the porcelain, and the extension length of the third telescopic rod can be adjusted according to the clamping width of the porcelain to adapt to porcelains of different sizes and types and achieve good display effects. The dimming glass cover can be converted between transparent and non-transparent to achieve concealed and confidential storage of porcelain, adding a barrier to the problem of theft. The light source is used to illuminate the porcelain, and an adjustable camera is used to capture the image of the porcelain. It has the advantages of simple operation, strong applicability and high safety.

[0056] In this embodiment, the adsorption pad 5 is a soft-glue adsorption pad, which is located at the part of the collection device that contacts the porcelain. Its main function is to provide a soft contact surface to protect the porcelain from damage during the fixing process. At the same time, the soft-glue adsorption pad has a certain adsorption force, which can stably fix the porcelain to prevent it from sliding during the display process.

[0057] In this embodiment, the buffer layer 11 is a memory sponge buffer layer, which is wrapped around the contact portion between the splint and the porcelain. Its function is to buffer the pressure between the porcelain and the splint, and further protect the porcelain from possible damage caused by contact with hard objects.

[0058] A control terminal is arranged in the non-transparent storage box.

[0059] In this embodiment, the first telescopic rod 2, the second telescopic rod 7, the third telescopic rod 9 and the fourth telescopic rod 13 are all electric telescopic rods.

[0060] In this embodiment, a pressure sensor is further provided on the side of the clamping plate 10 facing the porcelain 6 to prevent the third telescopic rod 9 from extending too long and causing squeezing damage to the porcelain.

[0061] In this embodiment, the height of the storage box 1 is 0.5m-1.2m, the height of the dimming glass cover 12 is 1m-2m, and the light source 16 is a warm light lamp.

[0062] In this embodiment, the storage box 1 is a non-transparent storage box.

[0063] When the present invention is used, firstly, the porcelain is placed on the rotating support platform to ensure that the bottom of the porcelain is in stable contact with the adsorption pad on the rotating support platform; then, according to the height of the porcelain, the height of the rotating support mounting platform is adjusted by the first telescopic rod until the ideal display height is reached; then, the extension height of the second telescopic rod is adjusted according to the clamping height of the porcelain, and the extension length of the third telescopic rod is adjusted according to the clamping width of the porcelain, and the splint is gently placed on the porcelain bottle body. At this time, the memory foam buffer layer will automatically adapt to the shape of the bottle body, provide uniform pressure distribution, and avoid damage to the porcelain. The soft glue adsorption pad plays an additional fixing role at the contact point between the splint and the porcelain, ensuring that the porcelain will not shake during the rotation process; when in use, the rotating support platform can be easily rotated to display different sides of the porcelain. During the whole process, the bottom plate of the storage box provides stable support to ensure that the collection device is Adjust the height and stability during rotation. Through this implementation method, the collection device can not only realize multi-angle display of porcelain and improve the viewing experience, but also quickly and easily adjust the height according to the size and display requirements of different porcelains, while ensuring the safety of the porcelain during the display process, achieving a display effect that is both beautiful and practical. Finally, after the display is over, the porcelain is retracted into the storage box by lowering the first telescopic rod, and the dimming glass cover is changed to a non-transparent state, so that the items in the collection device cannot be seen from the outside, thereby realizing hidden and confidential storage of porcelain, adding a barrier to the problem of theft, and having the advantages of simple operation, strong applicability, and high safety. When taking pictures, adjust multiple warm light lamps, adjust the height of the camera according to the height of the porcelain, and use the camera mount to adjust the camera angle. The camera mount adopts a pan / tilt head, and the camera preferably adopts Canon EOS RP.

[0064] like Figures 2 to 8 The method for dating blue and white porcelain by characteristics shown comprises the following steps:

[0065] Step 1: construct a blue and white porcelain decoration image database: randomly classify the blue and white porcelain decoration image database to obtain an image training data set and an image test data set;

[0066] Step 2: Construct a dating model for blue and white porcelain decoration: The blue and white porcelain decoration dating model includes a ShuffleNetV2 network model that introduces the LAM spatial attention mechanism, and the ShuffleNetV2 network model that introduces the LAM spatial attention mechanism is then subjected to a maximum pooling layer, global downsampling, and connected to a fully connected layer for category prediction;

[0067] Step 3: Train the blue-and-white porcelain pattern dating model: extract the enhanced blue-and-white porcelain pattern images from the image training data set and input them into the blue-and-white porcelain pattern dating model. Use the transfer learning strategy, RMSProp optimizer and cross entropy loss function to calculate the category loss, and output the category prediction results with category probabilities until the blue-and-white porcelain pattern dating model converges, thus completing the training process of the blue-and-white porcelain pattern dating model.

[0068] Step 4: Testing the blue-and-white porcelain pattern dating model: Retrieving blue-and-white porcelain pattern sample images from the image test data set, testing the blue-and-white porcelain pattern dating model until the test is completed, and obtaining a trained and tested blue-and-white porcelain pattern dating model;

[0069] Step 5: using a blue and white porcelain feature acquisition device to obtain an actual blue and white porcelain pattern image;

[0070] Step 6: Preprocess the actual blue and white porcelain pattern image. The process is as follows:

[0071] Step 601, scaling the actual blue and white porcelain decoration image to a size of 224×224, and then graying it;

[0072] Step 602, unsharp mask processing: use a low-pass filter to perform low-frequency filtering on the actual blue and white porcelain decoration image after grayscale processing to remove high-frequency part information; then subtract the low-frequency component image from the actual blue and white porcelain decoration image after grayscale processing to obtain a high-frequency change image, perform detail magnification processing on the high-frequency change image through different weight coefficients, overlap the high-frequency change image with magnified details and the actual blue and white porcelain decoration image after grayscale processing to obtain an unsharp mask enhanced image with more prominent edge and contour shapes of the image;

[0073] Step 603, HOG edge feature extraction: perform equal area segmentation on the unsharp mask enhanced image, with 8*8 pixels as a unit and 2*2 pixels as a region, divide the image into multiple regions, calculate the pixel gradient of each region, extract the undirected HOG gradient, divide the gradient direction into 9 intervals on average, and generate 9 histograms of corresponding intervals;

[0074] Step 604, LBP texture feature extraction: By comparing the grayscale values ​​of each single pixel in the image with the surrounding pixels, a binary code is formed, and the spatial distribution characteristics of these codes are counted to form a feature vector. The LBP operator operates in a 3×3 neighborhood, with the original single point grayscale value as the reference mark, and then the size is determined by the grayscale values ​​of the center pixel and the eight surrounding pixels. Finally, a single 8-bit binary number is generated by stacking. If the grayscale value of the field pixel is higher than the original single point pixel, the binary number of the digital will be set to 01. On the contrary, if it is lower than the center pixel, the corresponding binary number will be set to 00. Through a complete calculation, each 3×3 pixel block will be converted into a unique 8-bit binary code, which is used to characterize the texture feature signal of the area, and finally the local binary value corresponding to the original single pixel value is obtained through cascade arrangement.

[0075] Step 605, HSV color histogram processing: by dividing the color space of the image into a series of unrelated single intervals, and counting the total number of pixels in each interval, a color histogram is constructed; in the processing, different color channels of the image are processed separately to obtain the distribution characteristics of the color in each channel, and a three-dimensional or one-dimensional histogram is formed by counting all color channels to intuitively represent the color distribution of the image. The three-dimensional or one-dimensional histogram uses color values ​​as abscissas, ranging from 0 to 255, and corresponds to hue H information, lightness V information and saturation S information in the color space. The three-dimensional or one-dimensional histogram uses probability as a ordinate, and through normalization processing, the total probability is 1;

[0076] Step 606: Fusion of multiple features from step 603 to step 605;

[0077] Step 607, gradient boosting: using the gradient boosting model to improve the ability of the fused multi-feature weak learning machine to a strong learner;

[0078] Step 7: Dating blue and white porcelain decorations. The process is as follows:

[0079] Step 701: Use color threshold segmentation technology to extract the main decorative pattern of blue and white porcelain, and shield the background and edge areas to avoid generating additional interference information;

[0080] Step 702: Input the main blue and white porcelain pattern image extracted in step 701 into the trained blue and white porcelain pattern dating model, and output the probability information of the age of the blue and white porcelain pattern result.

[0081] It should be noted that most blue and white porcelain patterns are relatively complex. By extracting and fusing the directional gradient histogram, HSV color histogram and local binary pattern features, an unsharp masking process is introduced to enhance the texture features of the patterns, enhance the sensitivity of the model, and ensure accurate learning of the features of complex patterns; the LBP feature is mainly used to extract the main vein texture in the pattern, and can effectively extract all the line information of blue and white porcelain. The HOG feature is used to extract the edge shape of the blue and white porcelain pattern, and mainly describes the contour shape information of the image through gradient calculation. The HSV color histogram describes the layout of each color in the blue and white porcelain pattern, which provides richer information for the color space characteristics of the pattern image, and fully understands the surface characteristics of the blue and white porcelain pattern image. Through the fusion of the three multi-information features, the comprehensive extraction of blue and white porcelain pattern features can be achieved; the color threshold segmentation technology is integrated to effectively extract The complete decorative pattern on the glaze surface was obtained, and irrelevant interference information was successfully filtered out to avoid incorrect identification results; the gradient boosting model was used to enhance the ability of the fused multi-feature weak learning machine to a strong learner. The gradient boosting model performed better in dealing with small data set classification problems. Its advantages are mainly reflected in dealing with nonlinear relationships, robustness, no need for complex parameter adjustment, adaptability to different types of features, and processing small sample data. It is very suitable for use in blue and white porcelain classification tasks; based on the ShuffleNetV2 network model, the spatial attention mechanism was introduced to improve the original network ShuffleNetV2 to capture long-distance dependencies in the image and enhance the expressive power of the model. The transfer learning strategy, RMSProp optimizer and cross entropy loss function were experimentally verified on the blue and white porcelain dragon pattern dataset for accurate age identification.

[0082] In this embodiment, in step one, the blue and white porcelain decoration image is processed by random cropping and resizing, horizontal flipping, random rotation, and color disturbance to increase the diversity of the data.

[0083] In this embodiment, in step 2, the ShuffleNetV2 network model that introduces the LAM spatial attention mechanism includes a first convolutional layer, a first maximum pooling layer and a LAM spatial attention mechanism. The first maximum pooling layer performs downsampling to obtain a first feature map with a reduced size. The first feature map passes through Block-Conv1, Block-Conv2, Block-Conv3 and the second convolutional layer in sequence to output a second feature map. The second feature map introduces the LAM spatial attention mechanism and is input into the second maximum pooling layer. Finally, it is globally downsampled through a convolution kernel and connected to the fully connected layer for generation prediction; Block-Conv1 performs one operation with a step size of 2 and three operations with a step size of 1; Block-Conv2 performs one operation with a step size of 2 and seven operations with a step size of 1; Block-Conv2 performs one operation with a step size of 2 and three operations with a step size of 1.

[0084] In this embodiment, the size of the actual blue and white porcelain decoration image is 224×224×3. After the color threshold segmentation technology is introduced and processed by 24 3×3 convolution kernels, the size of the main blue and white porcelain decoration image becomes 112×112×24. The first maximum pooling layer performs downsampling to obtain a first feature map with a size reduced to 56×56×24. The size of the image after Block-Conv1 processing is 28×28×116, the size of the image after Block-Conv2 processing is 14×14×232, and the size of the image after Block-Conv3 processing is 7×7×464. The second convolution layer includes 1024 1×1 convolution kernels. The second feature map introduces the LAM spatial attention mechanism and is input into the second maximum pooling layer. The output image size is 7×7×1024. Finally, a 7×7 convolution kernel is used for global downsampling, and the output image size is 1×1×1024, and the fully connected layer is connected for period prediction.

[0085] In this embodiment, in step 701, the color threshold segmentation technology is to divide the colors in the image into different categories by setting the color range. The colors of the blue and white porcelain include blue and white. The blue patterns of the blue and white porcelain are separated from the white background by setting the corresponding color threshold to achieve preliminary extraction of the patterns. The blue patterns of different shades are extracted by adjusting the threshold. The adjustment of the color threshold captures the details more accurately, thereby obtaining a clearer and more complete pattern image.

[0086] It should be noted that in order to fully identify the blue-and-white porcelain decoration, it is necessary to select appropriate color thresholds (including upper thresholds and lower thresholds) based on the HSV color model, and use the cv2.inRange function to distinguish the blue-and-white porcelain decoration pattern color from the background domain of other colors in the image. Preferably, when the upper threshold upper_blue=np.array([130,255,255]) and the lower threshold lower_blue=np.array([90,50,50]) are set, the decoration and background image areas can be completely separated. In this experimental image, when the HSV value is between upper_blue and lower_blue, it is considered to be the blue-and-white porcelain decoration in the target area, and the original color is used by default to ensure complete extraction without changing the blue material color information. If it is not within this range, it is the background domain by default and is set to black.

[0087] In this embodiment, the blue and white porcelain decoration images include blue and white porcelain decoration images of multiple dynasties, and the blue and white porcelain decorations include figure blue and white porcelain decorations and dragon blue and white porcelain decorations.

[0088] When the present invention is used, the initial 5 hidden layer feature maps of the ShuffleNetV2 network model that introduces the LAM spatial attention mechanism are visualized, revealing the feature extraction mechanism of the ShuffleNetV2 network model that introduces the LAM spatial attention mechanism in the dating of blue and white porcelain, and analyzing the decorative features of blue and white porcelain of different eras learned by the network during the training process.

[0089] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any simple modification, change and equivalent structural change made to the above embodiment based on the technical essence of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A blue and white porcelain feature collection device, characterized in that: The invention comprises a storage box (1) and a dimming glass cover (12) arranged on the upper part of the storage box (1); a first telescopic rod (2) is installed at the bottom of the storage box (1); the top telescopic end of the first telescopic rod (2) is fixedly connected to the bottom of a slewing support mounting table (3) for mounting a slewing support platform (4); an adsorption pad (5) for buffering and adsorbing porcelain (6) is arranged on the top of the slewing support platform (4); a pair of adjustable clamping mechanisms for clamping and fixing the porcelain (6) and a fourth telescopic rod (13) for adjusting the position of a camera (15) are symmetrically arranged on both sides of the slewing support mounting table (3); The adjustable clamping mechanism comprises a second telescopic rod (7) mounted on a slewing support mounting platform (3); a mounting block (8) is fixedly mounted on the top telescopic end of the second telescopic rod (7); a third telescopic rod (9) is horizontally mounted on the side of the mounting block (8) facing the porcelain (6); a clamping plate (10) is fixedly mounted on the horizontal telescopic end of the third telescopic rod (9); a buffer layer (11) is arranged on the side of the clamping plate (10) facing the porcelain (6); a camera mount (14) is mounted on the top of the fourth telescopic rod (13); and a light source (16) for lighting the porcelain (6) is arranged on the top and side of the dimming glass cover (12).

2. A blue and white porcelain feature collection device according to claim 1, characterized in that: The absorption pad (5) is a soft-gel absorption pad, the buffer layer (11) is a memory foam buffer layer, the storage box (1) is a non-transparent storage box, and a control terminal is arranged inside the non-transparent storage box.

3. A blue and white porcelain feature collection device according to claim 1, characterized in that: The first telescopic rod (2), the second telescopic rod (7), the third telescopic rod (9) and the fourth telescopic rod (13) are all electric telescopic rods.

4. A blue and white porcelain feature collection device according to claim 1, characterized in that: A pressure sensor is also provided on the side of the clamping plate (10) facing the porcelain (6).

5. A blue and white porcelain feature collection device according to claim 1, characterized in that: The height of the storage box (1) is 0.5m to 1.2m, the height of the dimming glass cover (12) is 1m to 2m, and the light source (16) is a warm light lamp.

6. A method for dating blue and white porcelain using the acquisition device as claimed in claim 1, characterized in that: The dating method includes the following steps: Step 1: construct a blue and white porcelain decoration image database: randomly classify the blue and white porcelain decoration image database to obtain an image training data set and an image test data set; Step 2: Construct a dating model for blue and white porcelain decoration: The blue and white porcelain decoration dating model includes a ShuffleNetV2 network model that introduces the LAM spatial attention mechanism, and the ShuffleNetV2 network model that introduces the LAM spatial attention mechanism is then subjected to a maximum pooling layer, global downsampling, and connected to a fully connected layer for category prediction; Step 3: Train the blue-and-white porcelain pattern dating model: extract the enhanced blue-and-white porcelain pattern images from the image training data set and input them into the blue-and-white porcelain pattern dating model. Use the transfer learning strategy, RMSProp optimizer and cross entropy loss function to calculate the category loss, and output the category prediction results with category probabilities until the blue-and-white porcelain pattern dating model converges, thus completing the training process of the blue-and-white porcelain pattern dating model. Step 4: Testing the blue-and-white porcelain pattern dating model: Retrieving blue-and-white porcelain pattern sample images from the image test data set, testing the blue-and-white porcelain pattern dating model until the test is completed, and obtaining a trained and tested blue-and-white porcelain pattern dating model; Step 5: using a blue and white porcelain feature acquisition device to obtain an actual blue and white porcelain pattern image; Step 6: Preprocess the actual blue and white porcelain pattern image. The process is as follows: Step 601, scaling the actual blue and white porcelain decoration image to a size of 224×224, and then graying it; Step 602, unsharp mask processing: use a low-pass filter to perform low-frequency filtering on the actual blue and white porcelain decoration image after grayscale processing to remove high-frequency part information; then subtract the low-frequency component image from the actual blue and white porcelain decoration image after grayscale processing to obtain a high-frequency change image, perform detail magnification processing on the high-frequency change image through different weight coefficients, overlap the high-frequency change image with magnified details and the actual blue and white porcelain decoration image after grayscale processing to obtain an unsharp mask enhanced image with more prominent edge and contour shapes of the image; Step 603, HOG edge feature extraction: perform equal area segmentation on the unsharp mask enhanced image, with 8*8 pixels as a unit and 2*2 pixels as a region, divide the image into multiple regions, calculate the pixel gradient of each region, extract the undirected HOG gradient, divide the gradient direction into 9 intervals on average, and generate 9 histograms of corresponding intervals; Step 604, LBP texture feature extraction: By comparing the grayscale values ​​of each single pixel in the image with the surrounding pixels, a binary code is formed, and the spatial distribution characteristics of these codes are counted to form a feature vector. The LBP operator operates in a 3×3 neighborhood, with the original single point grayscale value as the reference mark, and then the size is determined by the grayscale values ​​of the center pixel and the eight surrounding pixels. Finally, a single 8-bit binary number is generated by stacking. If the grayscale value of the field pixel is higher than the original single point pixel, the binary number of the digital will be set to 01. On the contrary, if it is lower than the center pixel, the corresponding binary number will be set to 00. Through a complete calculation, each 3×3 pixel block will be converted into a unique 8-bit binary code, which is used to characterize the texture feature signal of the area, and finally the local binary value corresponding to the original single pixel value is obtained through cascade arrangement. Step 605, HSV color histogram processing: by dividing the color space of the image into a series of unrelated single intervals, and counting the total number of pixels in each interval, a color histogram is constructed; in the processing, different color channels of the image are processed separately to obtain the distribution characteristics of the color in each channel, and a three-dimensional or one-dimensional histogram is formed by counting all color channels to intuitively represent the color distribution of the image. The three-dimensional or one-dimensional histogram uses color values ​​as abscissas, ranging from 0 to 255, and corresponds to hue H information, lightness V information and saturation S information in the color space. The three-dimensional or one-dimensional histogram uses probability as a ordinate, and through normalization processing, the total probability is 1; Step 606: Fusion of multiple features from step 603 to step 605; Step 607, gradient boosting: using the gradient boosting model to improve the ability of the fused multi-feature weak learning machine to a strong learner; Step 7: Dating blue and white porcelain decorations. The process is as follows: Step 701: Use color threshold segmentation technology to extract the main decorative pattern of blue and white porcelain, and shield the background and edge areas to avoid generating additional interference information; Step 702: Input the main blue and white porcelain pattern image extracted in step 701 into the trained blue and white porcelain pattern dating model, and output the probability information of the age of the blue and white porcelain pattern result.

7. The method for dating blue and white porcelain by characteristics according to claim 6, characterized in that: In step one, the blue and white porcelain decoration images are processed by random cropping and resizing, horizontal flipping, random rotation, and color perturbation to increase data diversity; the blue and white porcelain decoration images include blue and white porcelain decoration images of multiple dynasties, and the blue and white porcelain decorations include figure blue and white porcelain decorations and dragon blue and white porcelain decorations.

8. The method for dating blue and white porcelain by characteristics according to claim 6, characterized in that: In step 2, the ShuffleNetV2 network model that introduces the LAM spatial attention mechanism includes a first convolutional layer, a first maximum pooling layer and a LAM spatial attention mechanism. The first maximum pooling layer performs downsampling to obtain a first feature map with a reduced size. The first feature map passes through Block-Conv1, Block-Conv2, Block-Conv3 and the second convolutional layer in sequence to output a second feature map. The second feature map introduces the LAM spatial attention mechanism and is input into the second maximum pooling layer. Finally, it is globally downsampled through a convolution kernel and connected to the fully connected layer for generation prediction; Block-Conv1 performs one operation with a step size of 2 and three operations with a step size of 1; Block-Conv2 performs one operation with a step size of 2 and seven operations with a step size of 1; Block-Conv2 performs one operation with a step size of 2 and three operations with a step size of 1.

9. The method for dating blue and white porcelain by characteristics according to claim 8, characterized in that: The size of the actual blue and white porcelain decoration image is 224×224×3. After the color threshold segmentation technology is introduced and processed by 24 3×3 convolution kernels, the size of the main blue and white porcelain decoration image becomes 112×112×24. The first maximum pooling layer performs downsampling to obtain the first feature map with a size reduced to 56×56×24. The size of the image after Block-Conv1 processing is 28×28×116, the size of the image after Block-Conv2 processing is 14×14×232, and the size of the image after Block-Conv3 processing is 7×7×464. The second convolution layer includes 1024 1×1 convolution kernels. The second feature map introduces the LAM spatial attention mechanism and is input into the second maximum pooling layer. The output image size is 7×7×1024. Finally, a 7×7 convolution kernel is used for global downsampling, and the output image size is 1×1×1024, and it is connected to the fully connected layer for period prediction.

10. The method for dating blue and white porcelain by characteristics according to claim 6, characterized in that: In step 701, the color threshold segmentation technology is to divide the colors in the image into different categories by setting the color range. The colors of the blue and white porcelain include blue and white. The blue patterns of the blue and white porcelain are separated from the white background by setting the corresponding color threshold to achieve preliminary extraction of the patterns. The blue patterns of different shades are extracted by adjusting the threshold. The adjustment of the color threshold captures the details more accurately, thereby obtaining a clearer and more complete pattern image.

Images