Method and system for extracting fiber cross-section profile

Abstract

The present disclosure provides a method and a system for extracting fiber cross-section profile. The method comprises the following steps. An image reading step reads yarn cross-section images, wherein the yarn cross-section images comprise first type cross-sections and second type cross-sections. A first image preprocessing step provides a selection signal to select the first type cross-sections or the second type cross-sections. A second image preprocessing step segments the selected cross-sections from the yarn cross-section images as regions of interest. A feature extraction step calculates frequency domain features, gradient features and color features according to the selection signal. A cross-section profile extraction step extracts cross-section profiles of the regions of interest according to the frequency domain features, the gradient features and the color features. The present disclosure also provides a system for extracting fiber cross-section profile. Thus, the analysis time of fiber profile can be greatly saved, and the accuracy of fiber profile analysis can be significantly improved.

Description

Technical Field

[0001] This disclosure relates to an image recognition and image extraction technique, and in particular to a method and system for extracting fiber cross-sectional contours. Background Technology

[0002] In the textile industry, fiber image analysis often requires manual intervention to segment and fit fiber contours due to uneven fiber dyeing or compression of fiber cross-sections, resulting in low accuracy and high labor costs. Therefore, a new method and system for fiber cross-section contour analysis is needed to improve the current situation. Summary of the Invention

[0003] This disclosure provides a method for extracting fiber cross-sectional contours, comprising the following steps: An image reading step, reading a yarn cross-sectional image, wherein the yarn cross-sectional image includes a first type of cross-section and a second type of cross-section. A first image preprocessing step, providing a selection signal to select either the first type of cross-section or the second type of cross-section. A second image preprocessing step, segmenting the selected cross-section from the yarn cross-sectional image as a region of interest. A feature extraction step, calculating frequency domain features, gradient features, and color features based on the selection signal. A cross-sectional contour extraction step, extracting the cross-sectional contour of the region of interest based on the frequency domain features, gradient features, and color features.

[0004] In one embodiment, the method for extracting the fiber cross-sectional profile further includes performing superpixel segmentation calculations on the yarn cross-sectional image based on the selection signal to segment the selected cross-section.

[0005] In one embodiment, the method for extracting fiber cross-sectional contours further includes: performing multi-scale transformation calculations on yarn cross-sectional images based on selected signals to generate multiple multi-scale fiber images; and calculating the frequency domain features, gradient features, and color features of the region of interest through the multiple multi-scale fiber images.

[0006] In one embodiment, the method for extracting fiber cross-sectional contours further includes: performing Fourier transform on multiple multi-scale fiber images to generate multiple frequency domain data; and inputting the multiple frequency domain data into multiple first deep belief networks to calculate multiple frequency domain feature coding layers, and generating frequency domain features through the multiple frequency domain feature coding layers.

[0007] In one embodiment, the method for extracting fiber cross-sectional contours further includes: calculating multiple gradient data of the multiple multi-scale fiber images; and inputting the multiple gradient data into multiple second deep belief networks to calculate multiple gradient feature encoding layers, and generating gradient features through the multiple gradient feature encoding layers.

[0008] In one embodiment, multiple multi-scale fiber images are converted to the LUV color space to generate multiple color data; and the multiple color data are respectively input into multiple third-depth belief networks to compute multiple color feature coding layers, and a color feature layer is generated through the multiple color feature coding layers.

[0009] In one embodiment, the method for extracting fiber cross-sectional contours further includes: concatenating one of multiple frequency domain feature coding layers, one of multiple gradient feature coding layers, and one of multiple color feature coding layers to generate multiple concatenated feature coding layers; performing random forest regression on the multiple concatenated feature coding layers to generate multiple weighted similarities; generating similar heatmaps based on the multiple weighted similarities; and filtering at least one pixel whose weighted similarity to the similar heatmap is below a threshold to extract the cross-sectional contour of the region of interest.

[0010] In one embodiment, the method for extracting fiber cross-sectional contours further includes: performing alpha shape calculation based on filtered similar heatmaps to obtain a first contour map; performing Fourier transform on the yarn cross-sectional image to generate a frequency domain image; identifying multiple high-frequency data in the frequency domain image and generating a second contour map based on the multiple high-frequency data; performing pixel overlay on the first and second contour maps and filling the pixels within the contours to generate a third contour map; and performing boundary calculation of octet-connected component analysis on the third contour map to extract the cross-sectional contour of the region of interest.

[0011] In one embodiment, the first type of cross-section is the core yarn, and the second type of cross-section is the sheath yarn.

[0012] This disclosure provides a system for extracting fiber cross-sectional profiles, comprising an optical microscope, a memory, a processor, and an input / output interface. The optical microscope is used to generate images of the yarn cross-section. The memory is used to store program code. The processor is coupled to the optical microscope and the memory and is used to execute the program code to perform the aforementioned method for extracting fiber cross-sectional profiles. The input / output interface is coupled to the processor and is used to display the yarn cross-sectional image, receive selection signals, and display the cross-sectional profile of a region of interest.

[0013] Based on the above, the method and system disclosed herein for extracting fiber cross-sectional contours can identify the contours of all fibers from a yarn cross-sectional image by extracting the gradient features, color features, and frequency domain features of individual fibers in the yarn. Compared with traditional methods, the method and system disclosed herein can significantly save the analysis time of fiber contours and significantly improve the accuracy of fiber contour analysis. Attached Figure Description

[0014] Figure 1 This is a flowchart of the method for extracting fiber cross-sectional profiles disclosed herein;

[0015] Figure 2 This is a schematic diagram of a model of multiple first deep belief networks, multiple second deep belief networks, multiple third deep belief networks, and random forest regression calculus according to some embodiments of this disclosure;

[0016] Figure 3 This is a schematic diagram of the system for extracting fiber cross-sectional profiles disclosed herein.

[0017] [Symbol Explanation]

[0018] S110~S150: Steps

[0019] IDBN1: The First Combined Deep Belief Network

[0020] IDBN2: The Second Combined Deep Belief Network

[0021] IDBN3: The Third Combined Deep Belief Network

[0022] RFM: Random Forest Regression Algorithm Model

[0023] 300: System

[0024] 310: Optical Microscope

[0025] 320: Memory

[0026] 330: Processor

[0027] 340: Input / output interface Detailed Implementation

[0028] Figure 1 This is a flowchart of the method for extracting fiber cross-sectional profiles disclosed herein. Please refer to... Figure 1 The method for extracting fiber cross-sectional contours disclosed herein includes an image reading step S110, a first image preprocessing step S120, a second image preprocessing step S130, a feature extraction step S140, and a cross-sectional contour extraction step S150.

[0029] First, in the image reading step S110, a yarn cross-section image is read, wherein the yarn cross-section image includes a first type of cross-section and a second type of cross-section. In some embodiments, the yarn may be, for example, a core-sheath type yarn, and the first type of cross-section may be, for example, a core yarn, and the second type of cross-section may be, for example, a sheath yarn.

[0030] Next, in the first image preprocessing step S120, a selection signal is provided to select a first type of cross-section or a second type of cross-section, wherein the selection signal may be, for example, a sample image of the first type of cross-section or the second type of cross-section. In some embodiments, when the yarn is a core-sheath type yarn and its first type of cross-section and second type of cross-section are the core yarn and the sheath yarn, respectively, the selection signal may be, for example, a sample image of the core yarn or the sheath yarn. In some embodiments, an inscribed circular mask may be generated based on the selection signal, and the first type of cross-section or the second type of cross-section may be selected by the inscribed circular mask to filter out the first type of cross-section or the second type of cross-section from the yarn cross-section image.

[0031] Next, in the second image preprocessing step S130, selected sections are segmented from the yarn cross-section image to serve as regions of interest (ROIs), and the ROI can be, for example, either a first type of section or a second type of section. In some embodiments, superpixel segmentation calculations can be performed on the yarn cross-section image based on a selection signal to segment the selected sections, and then the ROI can be segmented from the selected sections. For example, when the yarn is a core-sheath type yarn, one of the core yarn and the sheath yarn can be segmented from the selected sections of the yarn to serve as the ROI. In some embodiments, the superpixel segmentation calculation is, for example, a Simple Linear Iterative Clustering (SLIC) segmentation calculation, and the number of categories in the SLIC segmentation calculation can be set to 2, thereby dividing the first type of section (core yarn) and the second type of section (sheath yarn) into a first group and a second group, and selecting one of them as the ROI.

[0032] Next, in feature extraction step S140, frequency domain features, gradient features, and color features are calculated based on the selected signal. In some embodiments, multi-scale transformation calculations can be performed on the yarn cross-section image based on the selected signal to generate multiple multi-scale fiber images, and the frequency domain features, gradient features, and color features of the region of interest can be calculated using these multiple multi-scale fiber images. More specifically, the yarn cross-section image can be randomly scaled into corresponding images at multiple scales based on the selected signal, and then feature extraction can be performed on the corresponding images to calculate the frequency domain features, gradient features, and color features of the region of interest.

[0033] In some implementations, Fourier transforms can be performed on multiple multi-scale fiber images to generate multiple frequency domain data. These multiple frequency domain data can then be input into multiple first deep belief networks to compute multiple frequency domain feature coding layers, thereby generating frequency domain features. These multiple frequency domain feature coding layers can each be output layers of the multiple first deep belief networks. In other words, multiple first deep belief networks can be used to reduce the dimensionality of the multiple frequency domain data to generate frequency domain features.

[0034] In some implementations, multiple gradient data points can be calculated from the multiple multi-scale fiber images. These gradient data points can then be input into multiple second deep belief networks to compute multiple gradient feature encoding layers, thereby generating gradient features through these multiple gradient feature encoding layers, where each gradient feature encoding layer can be an output layer of the multiple second deep belief networks. In other words, multiple second deep belief networks can be used to reduce the dimensionality of the multiple gradient data points to generate gradient features.

[0035] In some implementations, multiple multi-scale fiber images can be converted to the LUV color space to generate multiple color data. These multiple color data can then be input into multiple third-depth belief networks to compute multiple color feature encoding layers, thereby generating a color feature layer. Each of these color feature encoding layers can be an output layer of one of the multiple third-depth belief networks. In other words, multiple third-depth belief networks can be used to reduce the dimensionality of the multiple color data to generate color features.

[0036] Next, in the cross-sectional contour extraction step S150, the cross-sectional contour of the region of interest is extracted based on frequency domain features, gradient features, and color features. The frequency domain features, gradient features, and color features can be generated from the aforementioned multiple first to third deep belief networks. In some embodiments, one of the multiple frequency domain feature coding layers, one of the multiple gradient feature coding layers, and one of the multiple color feature coding layers can be concatenated to generate multiple concatenated feature coding layers. Each concatenated feature coding layer can be formed by normalizing and concatenating the same multi-scale fiber image's frequency domain feature coding layer, gradient feature coding layer, and color feature coding layer. Next, random forest regression is performed on the multiple concatenated feature coding layers to generate multiple weighted similarities, and similar heatmaps are generated based on these weighted similarities. Then, at least one pixel with a weighted similarity below a threshold corresponding to a similar heatmap is filtered out to extract the cross-sectional contour of the region of interest. In detail, the feature vectors output from multiple concatenated feature encoding layers and multiple multi-scale fiber images can be used as multiple training samples. Random forest regression is then performed on these training samples to generate a similarity recognition model. Next, the similarity recognition model generates multiple weighted similarities for pixels within the region of interest. Furthermore, these weighted similarities can be mapped to multiple color levels, and similarity heatmaps are generated based on these color levels. For example, a higher weighted similarity corresponds to a deeper red.

[0037] The following practical examples illustrate the structure of models with multiple first-level deep belief networks, multiple second-level deep belief networks, multiple third-level deep belief networks, and random forest regression calculus. Figure 2This is a schematic diagram illustrating multiple first deep belief networks, multiple second deep belief networks, multiple third deep belief networks, and a random forest regression calculus model according to some embodiments of this disclosure. Please refer to... Figure 2 Multiple first-level deep belief networks can form a first-level combined deep belief network (IDBN1), multiple second-level deep belief networks can form a second-level combined deep belief network (IDBN2), and multiple third-level deep belief networks can form a third-level combined deep belief network (IDBN3). Each of the first, second, and third deep belief networks consists of an input layer, a hidden layer, and an output layer. Furthermore, the first-level combined deep belief network (IDBN1), the second-level combined deep belief network (IDBN2), and the third-level combined deep belief network (IDBN3) can be connected to the Random Forest Regression Model (RFM).

[0038] Please refer to the reply. Figure 1 In some implementations, an alpha shape calculation can be performed based on filtered similar heatmaps to obtain a first contour map. Furthermore, a Fourier transform can be performed on the yarn cross-section image to generate a frequency domain image, identifying multiple high-frequency data points within the frequency domain image, and then generating a second contour map based on these multiple high-frequency data points. For example, pixels corresponding to high-frequency data in the region of interest can be set to white, while pixels in the remaining areas of the region of interest can be set to black to generate the second contour map. Next, the first and second contour maps can be pixel-stacked and the pixels within the contour can be filled to generate a third contour map. Finally, boundary calculations using octet-connected component analysis can be performed on the third contour map to extract the cross-sectional contour of the region of interest, where the cross-sectional contour may include multiple fiber contours.

[0039] In some implementations, the weighted similarity of multiple fiber contours can be integrated based on the first contour map to generate multiple integral values. Then, the cross-sectional contours are filtered based on the multiple integral values ​​and a preset integration threshold to generate filtered cross-sectional contours that display the actual fibers, thus determining whether the fiber contours are real fibers.

[0040] In some implementations, the geometric parameters of the fibers, such as the number of fibers, the fiber area of ​​each fiber, the fiber fineness (Dtex), the perimeter, and the shape ratio (the ratio of the longest diameter to the shortest diameter), can be calculated based on the profile of the filtered cross section.

[0041] Figure 3 This is a schematic diagram of the system for extracting fiber cross-sectional profiles disclosed herein. Please refer to... Figure 3This disclosure also provides a system 300 for extracting fiber cross-sectional profiles, which can be used to implement the method for extracting fiber cross-sectional profiles disclosed herein. Specifically, the system 300 for extracting fiber cross-sectional profiles includes an optical microscope 310, a memory 320, a processor 330, and an input / output interface 340. The optical microscope 310 is used to capture and generate the aforementioned yarn cross-sectional image. The memory 320 is used to store program code. The processor 330 is coupled to the optical microscope 310 and the memory 320, and the processor 330 is used to execute the program code to perform, for example... Figure 1 The method for extracting fiber cross-sectional contours is described above. An input / output interface 340 is coupled to a processor 330 and is used to display the yarn cross-sectional image, receive selection signals, and display the cross-sectional contour of the region of interest. In some embodiments, the optical microscope 310 can be a digital microscope or any optical element capable of converting optical images into digital images, the input / output interface 340 can be a smartphone, tablet computer, notebook computer, desktop computer, or any computing system capable of processing image-related data, and the processor 330 and memory 320 can be a CPU and cache memory configured in the input / output interface 340. Thus, the method for extracting fiber cross-sectional contours described above can be implemented using the system 300 for extracting fiber cross-sectional contours.

[0042] In summary, the method for extracting fiber cross-sectional contours disclosed in this paper can extract gradient features, color features, and frequency domain features of the region of interest from yarn cross-sectional images, thereby enabling rapid and accurate analysis of the fiber cross-sectional contours in the yarn cross-section. Compared to traditional methods, this significantly reduces the time required for fiber contour analysis and substantially improves its accuracy. Furthermore, the method for extracting fiber cross-sectional contours disclosed in this paper can also compress and weight the features extracted from the yarn cross-sectional images using superpixel segmentation, deep belief networks, and random forest regression.

[0043] Although this disclosure has been presented above with reference to embodiments, it is not intended to limit this disclosure. Anyone skilled in the art can make some modifications and refinements without departing from the spirit and scope of this disclosure. Therefore, the scope of protection of this disclosure shall be determined by the scope defined in the appended claims.

Claims

1. A method for extracting fiber cross-sectional profiles, characterized in that, include: The image reading step involves reading a yarn cross-sectional image, wherein the yarn cross-sectional image includes a first type of cross-section and a second type of cross-section. The first image preprocessing step provides a selection signal to select the first type of cross-section or the second type of cross-section; The second image preprocessing step involves segmenting the selected cross-section from the yarn cross-section image to serve as the region of interest. The feature extraction step involves calculating frequency domain features, gradient features, and color features based on the selected signal. The cross-sectional contour extraction step extracts the cross-sectional contour of the region of interest based on the frequency domain features, the gradient features, and the color features. One of multiple frequency domain feature coding layers, one of multiple gradient feature coding layers, and one of multiple color feature coding layers are concatenated to generate multiple concatenated feature coding layers. Random forest regression is performed on the multiple concatenated feature encoding layers to generate multiple weighted similarities; Based on the multiple weighted similarities, a similar heatmap is generated; as well as Filter at least one pixel whose weighted similarity to the similar heatmap is below a threshold to extract the cross-sectional contour of the region of interest.

2. The method for extracting fiber cross-sectional contours according to claim 1, characterized in that, It also includes performing superpixel segmentation calculations on the yarn cross-section image based on the selection signal to segment out the selected cross-section.

3. The method for extracting fiber cross-sectional contours according to claim 1, characterized in that, Also includes: The yarn cross-sectional image is subjected to multi-scale transformation calculation based on the selected signal to generate multiple multi-scale fiber images. as well as The frequency domain features, gradient features, and color features of the region of interest are calculated using the multiple multi-scale fiber images.

4. The method for extracting fiber cross-sectional contours according to claim 3, characterized in that, Also includes: Fourier transform is performed on the multiple multi-scale fiber images to generate multiple frequency domain data; as well as The multiple frequency domain data are respectively input into multiple first deep belief networks to calculate the multiple frequency domain feature coding layers, and the frequency domain features are generated through the multiple frequency domain feature coding layers.

5. The method for extracting fiber cross-sectional contours according to claim 4, characterized in that, Also includes: Calculate multiple gradient data of the multiple multi-scale fiber images; as well as The multiple gradient data are respectively input into multiple second deep belief networks to calculate the multiple gradient feature encoding layers, and the gradient features are generated through the multiple gradient feature encoding layers.

6. The method for extracting fiber cross-sectional contours according to claim 5, characterized in that, Also includes: The multiple multi-scale fiber images are converted to the LUV color space to generate multiple color data. as well as The multiple color data are respectively input into multiple third-depth belief networks to calculate the multiple color feature coding layers, and the color feature layer is generated through the multiple color feature coding layers.

7. The method for extracting fiber cross-sectional contours according to claim 6, characterized in that, Also includes: Alpha shape calculation is performed based on the filtered similar heatmap to obtain a first contour map; The cross-sectional image of the yarn is subjected to Fourier transform to generate a frequency domain image; Identify multiple high-frequency data in the frequency domain image and generate a second contour map based on the multiple high-frequency data; The first contour image and the second contour image are pixel-overlayed and the pixels within the contour are filled to generate a third contour image; and Boundary calculations of eight-connected component analysis are performed on the third contour map to extract the cross-sectional contour of the region of interest.

8. The method for extracting fiber cross-sectional contours according to claim 1, characterized in that, The first type of cross-section is core yarn, and the second type of cross-section is sheath yarn.

9. A system for extracting fiber cross-sectional profiles, characterized in that, For implementing the method for extracting fiber cross-sectional profiles as described in claim 1, the system for extracting fiber cross-sectional profiles comprises: An optical microscope is used to produce cross-sectional images of the yarn. Memory, used to store program code; A processor, coupled to the optical microscope and the memory, is used to execute the program code; and The input / output interface, coupled to the processor, is used to display the yarn cross-sectional image, receive the selection signal, and display the cross-sectional outline of the region of interest.

Images