Intelligent detection method for down loft

Abstract

The application discloses the technical field of down detection, and discloses a kind of down fluffiness intelligent detection method, comprising the following steps:1) down sample pretreatment and standardization preparation;2) multi-modal data acquisition;3) down area segmentation and multidimensional feature extraction;4) multidimensional feature fusion and digital quantization representation;5) multi-factor optimization regression model training and fluffiness prediction.The detection method of the application is aimed at the three-dimensional space characteristics of down fluffiness, analyzes the correlation between down microstructure and fluffiness, enhances the explainability of the detection result at the micro level, and realizes the rapid, accurate and automated detection of down fluffiness.

Description

Technical Field

[0001] This invention relates to the field of down detection technology, and in particular to an intelligent method for detecting down loft. Background Technology

[0002] The loft of down is a core physical indicator for measuring the warmth retention and quality grade of down. It is defined as the volume occupied by a unit mass of down under a specific pressure, which directly determines the insulation efficiency, comfort and market value of down products. Its accurate quantification is of great significance for product quality grading and market pricing.

[0003] With advancements in computer vision and machine learning, down quality inspection has shifted from manual inspection to traditional algorithm-based detection. However, existing image-based intelligent down detection technologies mostly focus on qualitative judgments of down type or down content, lacking a quantitative analysis system for the key physical property of loft. Furthermore, even using traditional Deep Convolutional Neural Networks (DCNNs) and deep learning techniques, the specific characteristics of loft detection and the complexity of down images make it difficult to achieve the required recognition accuracy and quantification. Therefore, developing an automated, high-precision intelligent method for down loft detection is particularly important. Summary of the Invention

[0004] In view of the above-mentioned defects or deficiencies in the existing technology, it is desirable to provide an intelligent detection method for down loft, which analyzes the correlation between down microstructure and loft based on the three-dimensional spatial characteristics of down loft, enhances the interpretability of detection results at the micro level, and realizes rapid, accurate and automated detection of down loft.

[0005] The present invention provides an intelligent detection method for down loft, comprising the following steps: S1, Pretreatment and standardization of down samples; S2, Multimodal data acquisition; Simultaneously acquire 2D RGB images and 3D point cloud images of down samples, and divide the 2D RGB images into down cluster instance segmentation datasets and down background coarse segmentation datasets; S3, down region segmentation and multi-dimensional feature extraction; S31, a deep learning-based instance segmentation model is trained using a down clump instance segmentation dataset. The trained model obtains the instance mask for each independent down clump, calculates the core morphological features of each down clump instance mask, and statistically analyzes the distribution parameters of these core morphological features across the entire image, forming a contour morphological feature set F. contour ; S32, a deep learning-based semantic segmentation model is trained using a coarse segmentation dataset of down background. The trained semantic segmentation model is then used to obtain a down region mask to distinguish down from the background. Core texture features within the down region mask are extracted to form a texture complexity feature set F. texture ; S33, convert the 3D point cloud image into a voxel mesh, segment the pore voxel region based on the voxel gray value, separate connected pore regions, calculate the core pore features, and form a pore distribution feature set F. pore ; S4, the contour morphology feature set F contour Texture complexity feature set F texture Pore ​​distribution feature set F pore The features are concatenated to form a basic feature vector; then, core cross features are constructed, and the basic feature vector and core cross features are merged to form a comprehensive feature vector F. final ; S5. Select down samples of different varieties and different loft grades, and determine the true loft value P of each down sample. true,i Obtain the comprehensive feature vector F for each down sample according to S1-S4. final,i and the comprehensive feature vector F of the down sample to be tested final,new , will F final,i With P true,i A one-to-one correspondence was used to construct a fluffiness dataset, train and validate the regression model, obtain the optimal regression model, and combine the comprehensive feature vector F. final,new Input the optimal regression model and output the predicted fluffiness value P. pred .

[0006] Furthermore, in step S31, obtaining the instance mask for each independent down cluster includes: dividing the down cluster instance segmentation dataset into a training set a, a validation set a, and a test set a; training the deep learning-based instance segmentation model using the training set a and the validation set a; and inferring the 2D RGB image of the test set a using the trained instance segmentation model to output the instance mask for each independent down cluster.

[0007] Furthermore, in step S32, obtaining the down region mask includes: dividing the down background coarse segmentation dataset into a training set b, a validation set b, and a test set b; training the deep learning-based semantic segmentation model using the training set b and the validation set b; and performing coarse segmentation on the 2D RGB image in the test set b using the trained semantic segmentation model to obtain a down region mask that distinguishes only the down from the background.

[0008] Furthermore, in step S1, the down sample is placed under standard atmospheric conditions for more than 24 hours to equilibrate. A preset weight of down sample is weighed and evenly sprinkled into the sample tray. The vibration device at the bottom of the sample tray is activated to allow the down sample to spread out naturally and to stand until the fluffy state of the down sample stabilizes.

[0009] Furthermore, in step S2, a high-resolution industrial camera is used to acquire 2D RGB images of the down sample; a TOF depth camera is used to acquire 3D point cloud images of the down sample; Gaussian filtering is used to remove depth noise from the 3D point cloud images, and pass-through filtering is used to remove background point clouds from the 3D point cloud images.

[0010] Furthermore, in step S31, the core morphological features include: Cluster area, which is the pixel area occupied by a single down cluster, reflects the size of the cluster; Cluster perimeter is the pixel length of the down cluster outline, reflecting the edge complexity of the cluster; Roundness, with a value ranging from 0 to 1, indicates that the down tuft is closer to a sphere and has better expansion. The calculation formula is: C=4πA / L 2 Where C represents circularity and A represents cluster area in pixels. 2 L represents the cluster perimeter, measured in pixels. The smaller the convex hull area ratio, the more fully the fibrous branches extend, resulting in a stronger sense of fluffiness; the calculation formula is: R=A / A h Where R is the convex hull area ratio; A h The pixel area of ​​the convex hull of the down feather cluster, in pixels. 2 ; The aspect ratio is the ratio of the length of the principal axis to the length of the minor axis of the smallest bounding rectangle of the down tuft, reflecting the tuft's directionality of extension. The distribution parameters of the above five core morphological features in the entire image are statistically analyzed, including mean, variance, and kurtosis, to form a 15-dimensional contour morphological feature set F. contour .

[0011] Furthermore, in step S32, the core texture features include: Gray-level co-occurrence matrix features convert 2D RGB images into grayscale images, construct gray-level co-occurrence matrices, calculate their contrast, energy, homogeneity, and entropy, and obtain 4-dimensional feature vectors that reflect the coarseness, regularity, and complexity of textures. Local binary pattern features are obtained by using the LBP operator to calculate the local texture pattern histogram and obtain a 16-dimensional feature vector characterizing the micro-texture structure of the down surface. Gradient magnitude statistics: The Sobel operator is used to calculate the horizontal and vertical gradients of the 2D RGB image, and the average value and variance of the gradient magnitude are calculated to obtain a 2D feature vector indicating the undulations of the down surface. Principal component analysis was used to reduce the dimensionality of the 22-dimensional feature vectors, and the top 8 feature vectors with the largest cumulative variance contribution were retained to form an 8-dimensional texture complexity feature set F. texture .

[0012] Furthermore, in step S33, the core pore features include: Porosity is the ratio of effective pore volume to the total volume of down feathers; The average pore diameter is the average equivalent sphere diameter of all effective pores, in mm. The uniformity of pore distribution ranges from 0 to 1, with values ​​closer to 0 indicating a more uniform pore distribution. It is represented by the coefficient of variation of the pore diameter and is calculated using the formula: U = std(D) / D avg Where U is the uniformity of pore distribution; std(D) is the standard deviation of pore diameter in mm; The three-dimensional volume ratio is the ratio of the three-dimensional volume of down feathers to the pixel area of ​​a 2D RGB image, in mm. The above four core pore features form a four-dimensional pore distribution feature set F. pore .

[0013] Furthermore, in step S4, the core cross features include: The dilatation-complexity feature is the product of average roundness and texture entropy, representing an ideal fluffy state with spherical dilatation and complex filament structure; The volume-porosity synergistic feature is the product of the three-dimensional volume ratio and porosity, reflecting the synergistic effect of macroscopic volume and microscopic pores. Texture-pore uniformity is the product of the variance of the gradient amplitude and the complement of pore distribution uniformity, which characterizes the consistency between surface undulation and pore distribution. The complement of pore distribution uniformity is 1-U.

[0014] Furthermore, in step S5, the optimization regression model uses grid search combined with 5-fold cross-validation to optimize hyperparameters, with the optimization objective being to minimize the mean absolute percentage error of the test set c; an early stopping strategy is adopted to avoid model overfitting; the number of training iterations is no less than 1000 rounds, and the loss changes of the training set c and the validation set c are monitored in real time during the training process.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention utilizes a feature system constructed by integrating the morphological, textural, and porosity characteristics of down, which is highly correlated with the physical essence of loft. Extracted features such as roundness and porosity can explain the causes of loft at a microscopic level, providing technical support for down quality optimization. Its strong microscopic interpretability surpasses traditional methods that only output final values, achieving industrial-grade accuracy. It realizes a fully automated process from sample preparation to data acquisition and intelligent prediction, ensuring stable and reliable test results. This significantly shortens the testing time for a single sample, improves testing efficiency, and meets the needs of large-scale industrial quality inspection.

[0016] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0017] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 A flowchart for an intelligent detection method for down loft. Figure 2 This is a schematic diagram of the data acquisition device. Figure 3 This is a flowchart for predicting the loft of the down sample to be tested. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] Please refer to Figures 1-3 The present invention provides an intelligent detection method for down loft, comprising the following steps: S1, Pretreatment and Standardization of Down Samples: Down samples were placed under standard atmospheric conditions (temperature 20±2℃, relative humidity 65±4%) for at least 24 hours for constant temperature and humidity pretreatment to restore the down fibers to their natural fluffy state, eliminate structural deformation caused by compression during transportation or storage, and ensure that the initial state of the samples was consistent. Approximately 0.5±0.001g of the pretreated down sample was weighed using an electronic balance with an accuracy of 0.001g. Fixing the sample weight is the core prerequisite for establishing the correlation of "unit mass loft" and ensures that the image features of different samples are comparable. Select a standard-sized circular sample dish (10-20cm inner diameter, with a matte black background on the bottom to avoid glare interference), and evenly sprinkle the weighed down sample into the sample dish; activate the vibration device at the bottom of the sample dish and vibrate for 30-90 seconds to allow the down sample to spread naturally, avoiding local accumulation or sparseness caused by manual spreading; after vibration, let it stand for 2-4 minutes until the down sample is fluffy and stable before data acquisition.

[0021] S2, Multimodal Data Acquisition: A high-resolution industrial camera with a resolution of 2592×1944 pixels (equivalent to 1200 DPI) and a lens focal length of 50mm was used and mounted 20-50cm directly above the sample tray. A ring LED fill light device was configured to enhance the outline and texture details of the down feathers. The acquisition parameters were set as follows: ISO 100, shutter speed 1 / 500s, image format RGB, and 2D RGB images were acquired. A TOF depth camera is used, which is synchronously triggered to acquire data with a high-resolution industrial camera; the depth camera has a resolution of 1024×768 pixels, a measurement distance of 0.3-3m, a depth accuracy of ±1mm, and a point cloud density of 1000 points / cm². 2 The data is collected in PLY point cloud file format, and 3D point cloud images are collected. A hardware synchronization trigger module is used to ensure that the acquisition time difference between 2D RGB images and 3D point cloud images is ≤10ms, thus ensuring the spatiotemporal consistency of the data. The acquired 3D point cloud images are filtered by Gaussian to remove depth noise, and a pass-through filter is used to remove the background point cloud at the bottom and edge of the sample tray, while retaining the effective point cloud area of ​​the down sample. The 2D RGB images were divided into a down cluster instance segmentation dataset and a down background coarse segmentation dataset.

[0022] S3, Down Feather Region Segmentation and Multi-Dimensional Feature Extraction: S31, Contour morphology feature extraction based on instance segmentation: The down clump instance segmentation dataset (containing no fewer than 450 labeled images, covering down samples of different loft levels) was divided into training set a, validation set a, and test set a in an 8:1:1 ratio, where test set a consists of unlabeled images. The deep learning-based instance segmentation model was trained using training set a and validation set a to learn how to identify independent down clumps. The training iterations were no fewer than 300 rounds, with a batch size of 8 and an initial learning rate of 1e. -4 The cross-entropy loss function is used for optimization. The trained instance segmentation model is used to infer the 2D RGB images of test set a, outputting the instance mask and bounding box for each independent down cluster, achieving accurate separation of the "individual down cluster" from the background and other clusters; five core morphological features of each down cluster instance mask are calculated: Cluster area A is the pixel area occupied by a single down cluster, in pixels. 2 This reflects the size of the cluster; Cluster perimeter L is the pixel length of the down cluster outline, in pixels, reflecting the edge complexity of the cluster; Circularity C, ranging from 0 to 1, indicates that the down tuft is closer to a sphere and has better expansion. The calculation formula is: C = 4πA / L 2 ; The smaller the convex hull area ratio R, the more fully the filament branches extend, resulting in a stronger sense of fluffiness; the calculation formula is: R = A / A h (A) h The pixel area of ​​the convex hull of the down feather cluster, in pixels. 2 ); The aspect ratio AR is the ratio of the principal axis length to the minor axis length of the smallest bounding rectangle of the down tuft, reflecting the tuft's directionality of extension. The distribution parameters of the above five core morphological features in the entire image are statistically analyzed, including mean, variance, and kurtosis, to form a 15-dimensional contour morphological feature set F. contour .

[0023] S32, Texture complexity feature extraction based on local feature description: The coarse segmentation dataset of down background (containing no fewer than 450 labeled images, covering down samples of different loft levels) was divided into training set b, validation set b, and test set b in an 8:1:1 ratio, where test set b consists of unlabeled images. The deep learning-based semantic segmentation model was trained using the training set b and validation set b to learn how to distinguish down from the background. The training iterations were no fewer than 300 rounds, with a batch size of 8 and an initial learning rate of 1e. -4 The cross-entropy loss function is used for optimization. The trained semantic segmentation model is used to coarsely segment the 2DRGB images in test set b, obtaining a down region mask that distinguishes only the down feathers from the background. This mask is used to limit the effective area for texture analysis and avoid background interference. Three types of core texture features are extracted from the down region mask: The Gray-Level Co-occurrence Matrix (GLCM) feature converts a 2D RGB image into a grayscale image, constructs the GLCM, and calculates its contrast, energy, homogeneity, and entropy to obtain a 4-dimensional feature vector that reflects the coarseness, regularity, and complexity of the texture. Local binary patterns (LBP) features are obtained by using the LBP operator to calculate the local texture pattern histogram and obtain a 16-dimensional feature vector characterizing the micro-texture structure of the down surface. Gradient magnitude statistics: The Sobel operator is used to calculate the horizontal and vertical gradients of the 2D RGB image, and the average value (Gavg) and variance (Gvar) of the gradient magnitude are calculated to obtain a 2D feature vector that indicates the undulation of the down surface. The larger the variance of the gradient magnitude, the more intense the undulation of the down surface, and the richer the shadow and layering of the corresponding fluffy structure. Principal Component Analysis (PCA) was used to reduce the dimensionality of the 22-dimensional feature vectors mentioned above, retaining the top 8 feature vectors with the largest cumulative variance contribution rate, forming an 8-dimensional texture complexity feature set F. texture .

[0024] S33, Pore distribution feature extraction based on 3D point cloud: The preprocessed 3D point cloud image was converted into a voxel mesh to construct a 3D voxel model. The gray value of each voxel represents the down fiber density at that location. Based on the voxel gray values, a double thresholding method was used to segment the porous voxel regions (voxels with gray values ​​below T1 were identified as pores). A watershed algorithm (label control) was used to separate connected pore regions to avoid counting errors caused by pore adhesion, and voxels with a volume smaller than 0.01 mm were discarded. 3 The tiny pores (considered as measurement errors) are preserved, retaining the effective pore area; Calculate the characteristics of the four core pores: Porosity P is the ratio (%) of effective pore volume to total down volume. Average pore diameter (D) avg ), which is the average equivalent sphere diameter of all effective pores, in mm; The porosity uniformity U ranges from 0 to 1, with values ​​closer to 0 indicating a more uniform porosity distribution. It is represented by the coefficient of variation of the pore diameter and is calculated using the formula: U = std(D) / Davg (std(D) is the standard deviation of pore diameter, in mm); 3D volume ratio (R) 3D ), which is the ratio of the three-dimensional volume of down feathers to the pixel area of ​​a 2D RGB image, in mm, and directly relates the mapping relationship between the two-dimensional image and the three-dimensional volume; The above four core pore features form a four-dimensional pore distribution feature set F. pore .

[0025] S4, Multi-dimensional Feature Fusion and Digital Quantization Representation: The contour shape feature set F contour (15-dimensional) Texture complexity feature set F texture (8-dimensional) Pore distribution feature set F pore (4-dimensional) concatenation is performed to form a 27-dimensional basic feature vector; Three core cross features are constructed based on physical meaning to enhance the correlation between features and fluffiness, including: Inflation-Complexity Feature (C avg ×E texture ), which is the product of average roundness and texture entropy, characterizing the ideal fluffy state with spherical expansion and complex filament structure; Volume-pore synergistic features (R 3D ×P), which is the product of the three-dimensional volume ratio and porosity, reflecting the synergistic effect of macroscopic volume and microscopic pores; Texture-pore uniformity characteristics (G var ×(1 U) is the product of the variance of the gradient magnitude and the complement of the pore distribution uniformity, which represents the consistency between surface undulation and pore distribution. The pore distribution uniformity U is an inverse index, and the "complement" is to convert it into a positive index (the larger the value, the more uniform the pore distribution), so that it can be multiplied with other features and more intuitively represent the ideal state of the consistency between surface undulation and pore distribution. The complement of the pore distribution uniformity is 1-U. Integrating 27 basic features with 3 core cross features, a 30-dimensional comprehensive feature vector F is formed. final This vector is the "digital fingerprint" of down loft, providing input for subsequent model training.

[0026] As a preferred technical solution, the inflation-complexity feature MSC is defined as tanh(1.5×(FD-1)×P)×(1-U 0.6 )×lg(1.1+R 3D), where FD represents the fractal dimension, which is used to characterize the complexity of the down tuft outline structure. The higher the value, the more complex the branches of the down tuft and the better the loft. The fractal dimension is usually between 1.0 and 2.0. FD-1 is used to normalize the fractal dimension and compress the complexity of the down filament outline to the standard range. P represents porosity, which is the ratio of the effective pore volume to the total volume of the down, reflecting the internal air storage capacity.

[0027] Multiplying the normalized fractal dimension by the porosity enables microscopic dual-feature coupling, ensuring that only complex branching and abundant porosity (characteristics of high-quality down) result in a high product. Then, multiplying by a sensitivity adjustment coefficient of 1.5 moderately amplifies the coupling difference between FD and P, increasing the distinction between superior and ordinary down without causing numerical distortion. Finally, a hyperbolic tangent saturation function is used for range clamping, confining the overall result to the (0,1) range, suppressing unreasonable combinations and avoiding high scores from ineffective combinations of high fractal complexity and low porosity or low fractal complexity and high porosity. tanh(1.5×(FD-1)×P), as a microstructure quality factor, characterizes the quality of the down's microstructure, assessing only the combined quality of fractal complexity and porosity.

[0028] U represents the uniformity of pore distribution, characterizing the dispersion of pore diameter; the closer the value is to 0, the more uniform the distribution. A power of 0.6 serves as a non-linear smoothing coefficient, used to weaken extreme abrupt changes in U, smoothing the transition and avoiding drastic jumps in MSC caused by small fluctuations in uniformity, thus improving robustness. The smaller U is, the more uniform the distribution becomes (1-U...). 0.6 The larger the value of (1-U), the more uniform the porosity, and the higher the gain of this term. Ultimately, this term (1-U) 0.6 As a pore uniformity and regularity term, it can be used to characterize the regularity of pore spatial arrangement and quantify the efficient utilization capability of pore distribution.

[0029] R 3D This represents the normalized 3D volume ratio, i.e., the ratio of the normalized 3D volume of down to the pixel area of ​​the 2D image, relating the mapping relationship between the 2D projection and the true volume. A constant of 1.1 serves as the minimum guaranteed offset, and a non-linear mapping is performed through a logarithmic function, exhibiting high sensitivity in small volume ranges and saturation in large volume ranges. This matches the down loft pattern; small differences at low expansion can distinguish grades, while the loft tends to stabilize at high expansion, preventing ultra-large volume samples from monopolizing feature scores. Finally, lg(1.1+R) 3D The term ) is used as a macroscopic volume bulk gain term, which can be used to characterize the macroscopic expansion volume gain degree and describe the cross-scale mapping ability of 2D projection to 3D real bulk volume.

[0030] Inflation-complexity feature tanh(1.5×(FD-1)×P)×(1-U) 0.6 )×lg(1.1+R 3DThe value combines three factors: microstructure quality, pore uniformity and regularity, and macroscopic volume loft gain. The larger the value, the more uniform the pore distribution and the more significant the macroscopic expansion of the down under the support of a complex microstructure, which matches the ideal loft state of spherical expansion and complex down filament structure.

[0031] S5, Multifactor Optimization Regression Model Training and Fluffiness Prediction: First, select no fewer than 500 down samples of different varieties (white duck down, grey duck down, goose down) and different fill power grades, covering the mainstream product types in the industry; then, according to the IDFB (International Down and Feather Bureau) standard method, use a fill power meter to determine the true fill power value P of each sample. true,i The unit is cm. 3 / g serves as a model label to ensure the accuracy and authority of the true values; Obtain the 30-dimensional comprehensive feature vector F for each down sample according to steps S1-S4. final,i , will F final,i With P true,i A fluffiness dataset is constructed in a one-to-one correspondence; the fluffiness dataset is divided into a training set c, a validation set c, and a test set c in an 8:1:1 ratio to train and validate the regression model. Then, the regression model was optimized, which has the advantages of fast training speed, low memory consumption, and strong adaptability to high-dimensional features. The hyperparameters were optimized by combining grid search with 5-fold cross-validation, with the optimization objective being to minimize the mean absolute percentage error (MAPE) of the test set. An early stopping strategy was adopted (training was stopped if the MAPE of the test set did not decrease after 50 consecutive iterations) to avoid overfitting of the model. The number of training iterations was set to no less than 1000. The loss changes of the training set and validation set were monitored in real time during the model training process. For the down sample to be tested, obtain its comprehensive feature vector F according to steps S1-S4. final,new Input it into the trained optimal regression model, and the output is the predicted fluffiness value P. pred .

[0032] The trained optimized regression model has a determination coefficient R² of no less than 0.98 on the test set and a mean absolute percentage error (MAPE) of less than 5%, meeting the requirements for industrial-grade detection accuracy. Feature importance analysis of the optimized regression model verifies that the cumulative importance of the top 5 key features is greater than 70%, demonstrating the strong correlation between the proposed features and down loft.

[0033] In the description of this specification, the terms "one embodiment," "some embodiments," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0034] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for intelligent detection of down loft, characterized in that, Includes the following steps: S1, Pretreatment and standardization of down samples; S2, Multimodal data acquisition; Simultaneously acquire 2D RGB images and 3D point cloud images of down samples, and divide the 2D RGB images into down cluster instance segmentation datasets and down background coarse segmentation datasets; S3, down region segmentation and multi-dimensional feature extraction; S31, a deep learning-based instance segmentation model is trained using a down clump instance segmentation dataset. The trained model obtains the instance mask for each independent down clump, calculates the core morphological features of each down clump instance mask, and statistically analyzes the distribution parameters of these core morphological features across the entire image, forming a contour morphological feature set F. contour ; S32, a deep learning-based semantic segmentation model is trained using a coarse segmentation dataset of down background. The trained semantic segmentation model is then used to obtain a down region mask to distinguish down from the background. Core texture features within the down region mask are extracted to form a texture complexity feature set F. texture ; S33, convert the 3D point cloud image into a voxel mesh, segment the pore voxel region based on the voxel gray value, separate connected pore regions, calculate the core pore features, and form a pore distribution feature set F. pore ; S4, the contour morphology feature set F contour Texture complexity feature set F texture Pore ​​distribution feature set F pore The features are concatenated to form a basic feature vector; then, core cross features are constructed, and the basic feature vector and core cross features are merged to form a comprehensive feature vector F. final ; S5. Select down samples of different varieties and different loft grades, and determine the true loft value P of each down sample. true,i Obtain the comprehensive feature vector F for each down sample according to S1-S4. final,i and the comprehensive feature vector F of the down sample to be tested final,new , will F final,i With P true,i A one-to-one correspondence was used to construct a fluffiness dataset, train and validate the regression model, obtain the optimal regression model, and combine the comprehensive feature vector F. final,new Input the optimal regression model and output the predicted fluffiness value P. pred ; In step S4, the core cross features include the inflation-complexity feature, which is defined as tanh(1.5×(FD-1)×P)×(1-U) 0.6 )×lg(1.1+R 3D ), FD represents fractal dimension, P represents porosity, U represents pore distribution uniformity, R 3D Indicates the three-dimensional volume ratio; The core pore characteristics include porosity, pore distribution uniformity, and three-dimensional volume ratio.

2. The intelligent detection method for down loft according to claim 1, characterized in that, In step S31, obtaining the instance mask of each independent down cluster includes: dividing the down cluster instance segmentation dataset into training set a, validation set a, and test set a; training the deep learning-based instance segmentation model using training set a and validation set a; and inferring the 2D RGB image of test set a using the trained instance segmentation model to output the instance mask of each independent down cluster.

3. The intelligent detection method for down loft according to claim 1, characterized in that, In step S32, obtaining the down region mask includes: dividing the down background coarse segmentation dataset into training set b, validation set b, and test set b; training the deep learning-based semantic segmentation model using training set b and validation set b; and performing coarse segmentation on the 2D RGB image in test set b using the trained semantic segmentation model to obtain the down region mask that distinguishes only the down from the background.

4. The intelligent detection method for down loft according to claim 1, characterized in that, Step S1 specifically includes: placing the down sample under standard atmospheric conditions for more than 24 hours to equilibrate, weighing the down sample of a preset weight, evenly spreading the down sample into the sample tray, starting the vibration device at the bottom of the sample tray to allow the down sample to spread naturally, and letting it stand until the down sample's fluffy state stabilizes.

5. The intelligent detection method for down loft according to claim 1, characterized in that, Step S2 specifically includes: acquiring 2D RGB images of the down sample using a high-resolution industrial camera; acquiring 3D point cloud images of the down sample using a TOF depth camera; removing depth noise from the 3D point cloud image using Gaussian filtering; and removing background point clouds from the 3D point cloud image using pass-through filtering.

6. The intelligent detection method for down loft according to claim 1, characterized in that, In step S31, the core morphological features include cluster area, cluster perimeter, circularity, convex hull area ratio, and aspect ratio; the mean, variance, and kurtosis of the five core morphological features are calculated for the entire image to form a 15-dimensional contour morphological feature set F. contour .

7. The intelligent detection method for down loft according to claim 6, characterized in that, In step S32, extracting the core texture feature kernel includes: Gray-level co-occurrence matrix features convert 2D RGB images into grayscale images, construct gray-level co-occurrence matrices, calculate their contrast, energy, homogeneity, and entropy, and obtain 4-dimensional feature vectors that reflect the coarseness, regularity, and complexity of textures. Local binary pattern features are obtained by using the LBP operator to calculate the local texture pattern histogram and obtain a 16-dimensional feature vector characterizing the micro-texture structure of the down surface. Gradient magnitude statistics: The Sobel operator is used to calculate the horizontal and vertical gradients of the 2D RGB image, and the average value and variance of the gradient magnitude are calculated to obtain a 2D feature vector indicating the undulations of the down surface. Forming a texture complexity feature set F texture This includes: using principal component analysis to reduce the dimensionality of the 22 eigenvectors (4-dimensional, 16-dimensional, and 2-dimensional eigenvectors), retaining the top 8 eigenvectors with the largest cumulative variance contribution, forming an 8-dimensional texture complexity feature set F. texture .

8. The intelligent detection method for down loft according to claim 7, characterized in that, In step S33, the core pore characteristics also include the average pore diameter; Among them, a four-dimensional pore distribution feature set F is formed based on four core pore characteristics, including porosity, average pore diameter, pore distribution uniformity, and three-dimensional volume ratio. pore .

9. The intelligent detection method for down loft according to claim 7, characterized in that, In step S4, the core cross features also include volume-pore synergy features and texture-pore uniformity features.

10. The intelligent detection method for down loft according to claim 1, characterized in that, In step S5, the optimal regression model uses grid search combined with 5-fold cross-validation to optimize the hyperparameters. The optimization objective is to minimize the mean absolute percentage error of the test set c, and an early stopping strategy is adopted.

Images