Intelligent detection method for content of down feather

Abstract

The invention discloses an intelligent down content detection method, which uses a UNet model and fractal dimensions, realizes processing of two-dimensional image information through a non-down area identified by the UNet model, introduces depth features through the fractal dimensions, describes geometric complexity of down clusters, and comprehensively represents geometric features of down, so as to realize intelligent detection of down content. Namely, geometric complexity is introduced on the basis of image recognition, and the problems that a traditional convolutional neural network and a deep learning technology can only process two-dimensional image information, and recognition precision and accuracy are low are solved. Furthermore, the non-textured component pixel ratio and the fractal dimension are used as input features of the random forest algorithm, the random forest algorithm forms multiple decision trees according to the input features, the non-linear relationship among the input features can be effectively handled, and prediction results of the multiple decision trees are combined to improve the accuracy of the result.

Description

Technical Field

[0001] The invention relates to the technical field of textile product quality detection, and in particular to an intelligent detection method for down content. Background Art

[0002] In the quality inspection of down products, down content is a key factor affecting the quality of down. However, traditional inspection methods mostly rely on manual experience, which is not only time-consuming and labor-intensive, but also lacks stability in the inspection results. With the development of computer vision technology, such as progress in machine learning and other fields, down quality inspection has shifted from manual inspection to traditional algorithm inspection. Existing down content inspection methods are mostly used to distinguish down types, and there is no quantitative analysis of down content.

[0003] However, even with traditional convolutional neural networks and deep learning techniques, the recognition accuracy and precision are low due to the complexity of down images. Summary of the invention

[0004] In view of the above-mentioned defects or deficiencies in the prior art, it is desirable to provide an intelligent detection method for down content.

[0005] The present invention provides an intelligent detection method for down content, which specifically comprises the following steps:

[0006] S100. Collecting image information of the down to be tested by a scanner;

[0007] S200. Using a UNet model, the image information is divided into a plurality of secondary image information, and a non-velvet area of ​​each of the secondary image information is identified;

[0008] S300. Obtaining the non-velvet component pixel ratio by the first formula;

[0009] S400. Processing the image information by a box counting method to obtain a fractal dimension;

[0010] S500. Input the non-down component pixel ratio and the fractal dimension into a two-factor prediction model to obtain the down content.

[0011] According to the technical solution provided in the embodiment of the present application, the step S100 further includes pretreatment, and the pretreatment is to perform constant temperature and humidity treatment on the down to be tested.

[0012] According to the technical solution provided in the embodiment of the present application, the UNet model is obtained through deep convolutional network training, and a ReLU activation function is used after each convolutional layer in the UNet model.

[0013] According to the technical solution provided in the embodiment of the present application, the image information is a high-resolution picture of 8192×8192 pixels, and the secondary image information is a low-resolution image of 512×512 pixels. In step S200, the UNet model divides the image information into 256 pieces of secondary image information.

[0014] According to the technical solution provided in the embodiment of the present application, the first formula is:

[0015] P r =P c / P t

[0016] Among them, P r is the non-velvet component pixel ratio, P c is the pixel value of the non-velvet area, P t is the total pixel value.

[0017] According to the technical solution provided in the embodiment of the present application, the box counting method specifically includes the following steps:

[0018] S401. grayscale and binarize the image information to obtain a binary image, wherein the resolution of the binary image is α×α;

[0019] S402. Divide the binary image into M boxes, each box having a resolution of δ×δ, 2≤δ≤α / 2, and obtain N boxes containing down pixels;

[0020] S403. Calculate the fractal dimension according to the second formula, wherein the second formula is:

[0021]

[0022] Wherein, δ is the resolution of the box; N is the number of boxes containing feather pixels; Dim box (X) is the fractal dimension of the image information.

[0023] According to the technical solution provided in the embodiment of the present application, the method for creating the dual-factor prediction model includes the following steps:

[0024] S501. Obtain several samples and obtain the true value of the down content of each sample;

[0025] S502. Collecting sample image information of each sample through a scanner;

[0026] S503. Processing each of the sample image information through the UNet model to obtain the non-velvet component pixel ratio of each of the sample image information;

[0027] S504. Processing each of the sample image information by the box counting method to obtain the fractal dimension of each of the sample image information;

[0028] S505. Using the non-down component pixel ratios, the fractal dimensions and the down content true values ​​as input features of the random forest algorithm to form the dual-factor prediction model.

[0029] Compared with the prior art, the beneficial effects of the present invention are:

[0030] The present invention uses the UNet model and fractal dimension to realize the processing of two-dimensional image information through the non-down area identified by the UNet model, and then introduces deep features through the fractal dimension to describe the geometric complexity of down clusters and comprehensively characterize the geometric features of down. That is, geometric complexity is introduced on the basis of image recognition, solving the problem that traditional convolutional neural networks and deep learning technologies can only process two-dimensional image information and have low recognition precision and accuracy.

[0031] Furthermore, the non-velvet component pixel ratio and fractal dimension are used as input features of the random forest algorithm. The random forest algorithm forms multiple decision trees based on the input features, which can effectively deal with the nonlinear relationship between the input features and combine the prediction results of multiple decision trees to improve the accuracy of the results.

[0032] It should be understood that the contents described in the summary of the invention are not intended to limit the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0034] Figure 1 A flowchart of the steps of an intelligent detection method for down content provided in an embodiment of the present application;

[0035] Figure 2 A flowchart of the steps of creating a dual-factor prediction model in a method for intelligent detection of down content provided in an embodiment of the present application. DETAILED DESCRIPTION

[0036] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It should also be noted that, for ease of description, only the parts related to the invention are shown in the accompanying drawings.

[0037] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0038] Please refer to Figure 1 The embodiment of the present invention provides a method for intelligent detection of down content, which specifically includes the following steps:

[0039] S100. Collecting image information of the down to be tested by a scanner;

[0040] Optionally, the present invention uses a CanonScan 9000F Mark II scanner for high-resolution image acquisition, and the obtained picture is more accurate and has a higher resolution, wherein the weight of each sample is precisely controlled at 0.25 grams, the scanning resolution is 1200 DPI, and the generated image resolution is 8192×8192 pixels;

[0041] In some embodiments, the step S100 further includes pretreatment, wherein the pretreatment is to subject the down to be tested to a constant temperature and humidity treatment;

[0042] Among them, the constant temperature and humidity treatment is to place the down to be tested in a constant temperature and humidity environment and let it stand for several hours. Among them, the constant temperature and humidity environment refers to a temperature in the range of 20±2℃ and a relative humidity in the range of 65%±4%, to ensure the stability of the detection and reduce the detection error caused by temperature and humidity differences.

[0043] S200. Using a UNet model, the image information is divided into a plurality of secondary image information, and a non-velvet area of ​​each of the secondary image information is identified;

[0044] The non-down area refers to the area occupied by the non-down components in each secondary image information, and the non-down components include the parts other than the down and down fibers in the down aggregate (such as feather flakes, feather fibers, impurities, etc.);

[0045] In some embodiments, the UNet model is obtained by deep convolutional network training, and a ReLU activation function is used after each convolutional layer in the UNet model;

[0046] Among them, the UNet model is designed with an encoder-decoder architecture, and has shown wide application potential in the field of image segmentation. Specifically, the contraction path of the UNet model extracts the features of the down image through multiple 3×3 convolutional layers and maximum pooling layers, and the expansion path upsamples the features through transposed convolution. In addition, to ensure the integrity and accuracy of the information, each layer in the expansion path will be spliced ​​with the corresponding feature layer in the contraction path. This process significantly improves the UNet model's ability to capture and restore image detail information. Deep convolutional network training is a common training method, and the general process is as follows: first manually annotate the sample image, determine the target area (i.e., non-down area), and generate a high-quality annotated data set; input the annotated data set into the initial UNet model, the initial UNet model extracts features in the encoding stage, and then gradually restores the spatial resolution of the image in the decoding stage, and finally completes the training. After the training, it also includes verification and adjustment of initial parameters, which will not be repeated here.

[0047] In addition, the ReLU activation function is selected as the activation function of the UNet model, which not only enhances the nonlinear modeling ability of the model, but also retains the image details during the segmentation process, greatly reducing the information loss. Preferably, the UNet model uses a cross entropy loss function to optimize the network performance, ensuring that the UNet model can accurately distinguish between down components and non-down components, thereby achieving efficient distinction between the two types of components.

[0048] In some embodiments, the image information is a high-resolution image of 8192×8192 pixels, and the secondary image information is a low-resolution image of 512×512 pixels. In step S200, the UNet model divides the image information into 256 pieces of secondary image information.

[0049] In the prior art, inputting high-resolution images will increase the demand and pressure on computer hardware. The common practice is to compress an 8192×8192 pixel image into a 512×512 pixel image, and the tiny information of the image will be compressed. In this embodiment, the UNet model is used to segment the scanned high-resolution image, and there is no need to compress the image. The image information is not easily lost, thereby improving the recognition accuracy of the UNet model.

[0050] S300. Obtaining the non-velvet component pixel ratio by the first formula;

[0051] In some embodiments, the first formula is:

[0052] P r =P c / P t

[0053] Among them, Pr is the non-velvet component pixel ratio, P c is the pixel value of the non-velvet area, P t is the total pixel value;

[0054] Because the UNet model identifies the non-down areas of each secondary image information, it can obtain the pixel value of the non-down area, and then calculate the pixel ratio of the non-down component based on the total pixel value of the image information, thereby obtaining the proportion of the non-down component in the plane image. However, considering the staggered overlap of down, it cannot be directly output as accurate data on the down content.

[0055] S400. Processing the image information by a box counting method to obtain a fractal dimension;

[0056] In some embodiments, the box counting method specifically comprises the following steps:

[0057] S401. grayscale and binarize the image information to obtain a binary image, wherein the resolution of the binary image is α×α;

[0058] Among them, grayscale processing and binarization processing are commonly used image processing methods. Grayscale processing will present the entire image with an obvious black and white effect. By setting the binarization threshold, the part with lower grayscale value (specifically the scattered down) will be filtered out during the binarization processing, so as to better show the part with higher grayscale value in the image (including stacked down, feather fibers and impurities), and better reflect the geometric characteristics of the down.

[0059] S402. Divide the binary image into M boxes, each box having a resolution of δ×δ, 2≤δ≤α / 2, and obtain N boxes containing down pixels;

[0060] Among them, according to the arrangement of geometric progression, select δ = 2, 4, 8, ..., α in turn; the total number of boxes after each division M = (α / δ) 2 , select the boxes containing feather pixels from the M boxes and mark them as N.

[0061] S403. Calculate the fractal dimension according to the second formula, wherein the second formula is:

[0062]

[0063] Wherein, δ is the resolution of the box; N is the number of boxes containing down pixels; Dim box (X0 is the fractal dimension of the image information;

[0064] Among them, according to the value of δ as the X-axis, the corresponding value of N is the Y-axis. As δ increases, the value of N becomes smaller. The corresponding δ-N numerical relationship is displayed on the XOY coordinate system, and a curve is fitted to obtain a curve that gradually decreases along the positive direction of X, and Dim box (X) is the slope of the curve, that is, the fractal dimension, which makes it possible to describe the geometric complexity of the down clusters. In other words, the smaller the fractal dimension, that is, the smaller the slope, the more complex the down clusters. Furthermore, the more complex the down clusters are, the stronger the filling ability of the down in the picture is. It can also show the fluffiness of the down and the quality of the down in disguise, because the non-down component in the down aggregate is the least fluffi, so it can be used as a reference factor for predicting the down content.

[0065] S500. Inputting the non-velvet component pixel ratio and the fractal dimension into a two-factor prediction model to obtain the velvet content;

[0066] The two-factor prediction model is based on a model created by a random forest algorithm. Figure 2 After the non-down component pixel ratio and the fractal dimension are input into the model, the decision trees that meet the requirements will vote to pass and output a preliminary prediction value of down. The preliminary prediction values ​​of down output by all the decision trees that have passed the vote are fitted to obtain the prediction value of down content. Compared with the traditional neural convolutional network that only judges the down content by the non-down component pixel ratio, the introduced fractal dimension can take into account the geometric complexity of the down clusters and improve the accuracy.

[0067] In some embodiments, the method for creating the dual-factor prediction model comprises the following steps:

[0068] S501. Obtain several samples and obtain the true value of the down content of each sample;

[0069] S502. Collecting sample image information of each sample through a scanner;

[0070] S503. Processing each of the sample image information through the UNet model to obtain the non-velvet component pixel ratio of each of the sample image information;

[0071] S504. Processing each of the sample image information by the box counting method to obtain the fractal dimension of each of the sample image information;

[0072] S505. Using the non-velvet component pixel ratios, the fractal dimensions and the real values ​​of the velvet contents as input features of the random forest algorithm to form the dual-factor prediction model;

[0073] The sample in step S501 is obtained from the Textile Fiber Inspection Center of Tianjin Product Quality Supervision and Inspection Technology Research Institute, and is obtained by manual inspection according to the method of GB / T 14272-2021, and the true value of the down content is accurate and reliable; step S502 is the same as step S100, and step S501 and step S502 also include pretreatment. After pretreatment, the sample needs to be evenly shaken into a 17.6 cm × 17.6 cm square dark box to simulate the sample distribution state in the actual detection process. Further, step S503 is the same as step S300, and step S504 is the same as step S400;

[0074] In addition, step S505 can refer to Figure 2 , which specifically includes the following steps:

[0075] Firstly, several sub-training sample sets are created, each of which contains the non-velvet component pixel ratio P, fractal dimension X and the true value S of the velvet content of a sample image information;

[0076] Then the random forest algorithm randomly extracts multiple sub-training sample sets from the data set. Each sub-training sample set contains different X values ​​and S values ​​and corresponds to a target P value.

[0077] When training a decision tree, the random forest algorithm will automatically select the split point of the fractal dimension X and the true value S of the down content, so that the error of the sub-dataset of each split node in the pixel ratio P is minimized; the random forest algorithm will find the best split condition through different split combinations. For example, at a certain node, there may be a split condition X>1.3 or S<50, and the decision tree generates a predicted pixel ratio P;

[0078] Finally, all decision trees will generate multiple predicted pixel ratios P ​​based on different splitting conditions. For each set of X and S inputs, the decision tree that meets the splitting conditions will output the predicted pixel ratio P. The random forest will integrate all decision tree results and output the final pixel ratio P by averaging or fitting. In this way, based on different combinations of X and S values, the predicted P value of each combination is synthesized, and the complex relationship between P, X and S values ​​is automatically learned, and finally the two-factor prediction model is formed.

[0079] When the detection method provided in this embodiment implements step S500, each decision tree of the two-factor prediction model will make a decision based on the input non-down component pixel ratio P and fractal dimension X, and each decision tree will vote, that is, whether the split condition is met, and the voting results include not meeting the split condition and meeting the split condition, wherein the decision tree whose voting result meets the split condition will also output the predicted pixel ratio P, and the preliminary down prediction value is fitted to form the down content prediction value, wherein the fitting and training both adopt a sub-algorithm of the existing random forest algorithm, which will not be elaborated here.

[0080] In some embodiments, after step S505, the step further includes verifying whether the two-factor prediction model meets the prediction standard, that is, selecting a number of verification samples with known true values ​​of down content. The intelligent detection method for down down content provided in the above embodiment performs the detection step, and the parameters obtained are as follows:

[0081] Serial number Actual value of down content (%) Predicted value of down content (%) error(%) 1 48.6 48.7 0.19 2 52.8 52.8 0.03 3 90.2 90.1 -0.13 4 61.1 61.1 0.04 5 85.2 84.7 -0.56 6 68 68.1 0.19 7 70.3 69.5 -1.17 8 78.4 79.5 1.38

[0082] Table 1 Comparison between predicted down content and actual down content

[0083] The actual value of the above-mentioned down content comes from the Textile Fiber Inspection Center of Tianjin Product Quality Supervision and Inspection Technology Research Institute, which is obtained by manual testing using the GB / T 14272-2021 method and meets national standards. Furthermore, the determination coefficient (R 2 ) and mean absolute percentage error (M e ) statistical indicators are used to characterize it. The specific calculation method is shown in the following formula:

[0084]

[0085] Among them, x i represents the true value of the down content of the i-th verification sample; y i represents the predicted value of the down content of the i-th verification sample; Indicates the average value of the measured down content; represents the average down content predicted by the model; n represents the number of down samples;

[0086] After calculation, the determination coefficient (R 2 ) reached 0.95, and the mean absolute percentage error (M e ) is 1.25%, which belongs to the range of small deviation, so the prediction error is small.

[0087] In some embodiments, the UNet model also needs to obtain samples when training through a deep convolutional network, so it can be performed simultaneously with step S501 and step S502. When the UNet model and the two-factor prediction model in the detection method provided in this embodiment are created, a total of 450 first-class images of 512×512 pixels and 246 second-class images of 8192×8192 pixels are collected through steps S501 and S502, wherein the first-class images are manually annotated, and the specific annotation content includes covering various substances such as feathers, feather fibers, land bird hair and impurities in the down samples, and then the first-class images are used for training the UNet model; and the second-class images are used as sample image information for the creation of the two-factor prediction model.

[0088] In the description of this specification, the terms "connection", "installation", "fixation" and the like should be understood in a broad sense. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0089] In the description of this specification, the description of the terms "one embodiment", "some embodiments", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0090] The above are only preferred embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An intelligent detection method for down content, characterized in that: The specific steps include: S100. Collecting image information of the down to be tested by a scanner; S200. Using a UNet model, the image information is divided into a plurality of secondary image information, and a non-velvet area of ​​each of the secondary image information is identified; S300. Obtaining the non-velvet component pixel ratio by the first formula; S400. Processing the image information by a box counting method to obtain a fractal dimension; S500. Input the non-down component pixel ratio and the fractal dimension into a two-factor prediction model to obtain the down content.

2. The intelligent detection method for down content according to claim 1, characterized in that: The step S100 also includes pre-processing, wherein the pre-processing is to subject the down to be tested to a constant temperature and humidity treatment.

3. The intelligent detection method for down content according to claim 1, characterized in that: The UNet model is obtained through deep convolutional network training, and a ReLU activation function is used after each convolutional layer in the UNet model.

4. The intelligent detection method for down content according to claim 3 is characterized in that: The image information is a high-resolution image of 8192×8192 pixels, and the secondary image information is a low-resolution image of 512×512 pixels. In step S200, the UNet model divides the image information into 256 pieces of secondary image information.

5. The intelligent detection method for down content according to claim 1, characterized in that: The first formula is: P r =P c / P t Among them, P r is the non-velvet component pixel ratio, P c is the pixel value of the non-velvet area, P t is the total pixel value.

6. The intelligent detection method for down content according to claim 5, characterized in that: The box counting method specifically comprises the following steps: S401. grayscale and binarize the image information to obtain a binary image, wherein the resolution of the binary image is α×α; S402. Divide the binary image into M boxes, each box having a resolution of δ×δ, 2≤δ≤α / 2, and obtain N boxes containing down pixels; S403. Calculate the fractal dimension according to the second formula, wherein the second formula is: Wherein, δ is the resolution of the box; N is the number of boxes containing feather pixels; Dim box (X0 is the fractal dimension of the image information.

7. The intelligent detection method for down content according to claim 6, characterized in that: The method for creating the dual-factor prediction model comprises the following steps: S501. Obtain several samples and obtain the true value of the down content of each sample; S502. Collecting sample image information of each sample through a scanner; S503. Processing each of the sample image information through the UNet model to obtain the non-velvet component pixel ratio of each of the sample image information; S504. Processing each of the sample image information by the box counting method to obtain the fractal dimension of each of the sample image information; S505. Taking the non-down component pixel ratios, the fractal dimensions and the down content true values ​​as input features of the random forest algorithm to form the two-factor prediction model.

Images