CN-121982377-A - Intelligent labeling and classifying method for clothing patterns based on deep learning

CN121982377ACN 121982377 ACN121982377 ACN 121982377ACN-121982377-A

Abstract

The invention relates to the technical field of clothing pattern standards and classification, and discloses a clothing pattern intelligent labeling and classifying method based on deep learning, which comprises the steps of image data acquisition and processing, namely, receiving clothing image data, and carrying out standardized preprocessing on the received clothing image data to form an initial clothing image data set; the method comprises the steps of performing fold region positioning, analyzing a surface coincidence index based on each image data in an initial clothing image data set, judging whether folds exist or not based on the surface coincidence index, positioning a region judged to exist the folds, generating a fold mask based on positioning coordinates of the fold region, extracting deep learning semantic features to form feature vectors after fold interference is shielded, and outputting and visualizing the results. The method avoids the influence of pattern distortion caused by wrinkles on feature extraction, and solves the problem that the traditional feature extraction method is easy to be interfered by the wrinkles to lose efficacy.

Inventors

WU MINGFANG
CHE YONGSHUN
LIN YILUN
CHEN HONGCAI
YAO WENSHENG

Assignees

泉州市数字云谷信息产业发展有限公司
福建省数字福建云计算运营有限公司

Dates

Publication Date: 20260505
Application Date: 20251230

Claims (10)

1. The intelligent labeling and classifying method for the patterns of the clothing based on deep learning is characterized by comprising the following steps of: Firstly, image data acquisition processing, namely receiving clothing image data, and carrying out standardized preprocessing on the received clothing image data to form an initial clothing image data set; Step two, positioning a fold region, namely analyzing a surface coincidence index based on each image data in an initial clothing image data set, judging whether folds exist or not based on the surface coincidence index, and positioning a region judged to exist the folds; generating a fold mask based on the positioning coordinates of the fold region, and extracting deep learning semantic features to form feature vectors after the fold interference is shielded; And step four, outputting and visualizing the result, namely outputting the formed feature vector.
2. The intelligent labeling and classifying method for patterns of clothing based on deep learning according to claim 1, wherein the step one supports multi-source input of receiving clothing image data, specifically comprises: The method comprises the steps of acquiring local files, calling a camera of computer equipment to acquire images in real time, downloading network images through a URL, and calling an API to acquire remote image data.
3. The intelligent labeling and classifying method for patterns of clothing based on deep learning according to claim 2, wherein in the first step, the standardized preprocessing includes: The method comprises the steps of firstly unifying file formats of images with different sources, secondly standardizing image resolution, then performing color space conversion, simultaneously eliminating noise generated by camera acquisition and noise brought by network image compression through a denoising algorithm, and finally completing normalization processing of pixel values.
4. The intelligent labeling and classifying method for patterns of clothing based on deep learning according to claim 2, wherein the step two is to set any clothing image in the initial clothing image data set as The size is Taking any pixel point Centering on, build the size as Is a local neighborhood window of (2) ; Is the neighborhood radius; Is the side length of the window.
5. The intelligent labeling and classifying method for patterns of clothing based on deep learning according to claim 4, wherein the step two of judging whether wrinkles exist based on the surface coincidence index comprises: S1.1, opposite pixels Gray gradient Direction component And Direction component Calculation with Sobel operator: ; ; Wherein, the Is convolution operation; For initial clothing image Wherein the coordinates are Gray values of the pixels of (a); Pixel dot Is the gradient vector of (a) The gradient amplitude is ; S1.2, vs. neighborhood Window Computing neighborhood average gradient vectors Directional similarity of single pixel gradient vector to average vector Surface coincidence index The method comprises the following steps: ; Wherein, the The total number of pixels in the neighborhood; for local neighborhood windows In, the coordinates are Gradient vectors of pixels of (a); ; Wherein, the Is constant and takes value Dot product ; ; S1.3, image pairs Is defined by the number of pixels of the pixel array Traversing S1.1-S1.2 to generate a surface coincidence index diagram consistent with the original image in size 。
6. The intelligent labeling and classifying method for patterns of clothing based on deep learning according to claim 5, wherein in S1.3, a surface coincidence index threshold is set When the surface coincidence index Less than the surface coincidence index threshold Judging the neighborhood where the pixel is located as a fold area; When the surface coincidence index Greater than or equal to the surface coincidence index threshold And judging the neighborhood where the pixel is positioned as a wrinkle-free area.
7. The intelligent labeling and classifying method for patterns of clothing based on deep learning according to claim 6, wherein the step two of locating the region where the wrinkles are determined to exist comprises: S2.1 according to the surface coincidence index threshold Mapping the surface coincidence index Conversion to a binary image ; ; S2.2, vs. binary image Performing an expansion etching operation to eliminate isolated noise points and complement broken fold regions : ; Wherein, the Is that A structural element; s2.3 extraction Calculating the external rectangular coordinates of each connected domain The positioning result of the fold area is obtained.
8. The intelligent labeling and classifying method for patterns of clothing based on deep learning according to claim 7, wherein in the third step, the patterns are masked by wrinkles Expressed as: ; Wherein, the Is the coordinates A mask value of 0 indicates that the corresponding pixel is a wrinkle interference region, shielding is needed during feature extraction, and a mask value of 1 indicates a non-wrinkle region, and the complete pattern features are reserved.
9. The intelligent labeling and classifying method for patterns of clothing based on deep learning according to claim 8, wherein in the third step, the deep learning feature extraction includes: S3.1, imaging clothing Inputting ResNet50 0 pre-training model, extracting feature map at avgpool layer Dimension is , The number of channels; s3.2, mask pattern Downsampling to AND using bilinear interpolation Same size Obtaining a downsampled mask ; S3.3, mapping the characteristic diagram And downsampling mask Multiplication element by element, shielding the characteristics of the fold area, obtaining a disturbance-free characteristic diagram : ; Wherein, the For element-by-element multiplication; s3.4, will Flattened into a one-dimensional vector Dimension is , I.e. the feature vector.
10. The intelligent labeling and classifying method for patterns of clothing based on deep learning according to claim 9, wherein the fourth step comprises constructing a standardized output template and outputting based on the feature vector.

Description

Intelligent labeling and classifying method for clothing patterns based on deep learning Technical Field The invention relates to the technical field of garment pattern standards and classification, in particular to a garment pattern intelligent labeling and classifying method based on deep learning. Background The labeling and classifying of the clothing patterns is a process of automatically identifying pattern elements in clothing images through artificial intelligence technology and labeling the pattern elements with structured labels or classifying the pattern elements. The method has the core effects of improving the digital management efficiency and the intelligent application level of the clothing industry, accelerating the commodity putting-on process, reducing the manual marking cost, helping consumers to quickly find the style of the heart instrument through a precise pattern feature extraction, optimizing and searching and recommending system, and on the other hand, providing data support for design trend analysis, copyright protection, supply chain collaboration and personalized customization, promoting the industry to change from traditional experience driving to data driving, guiding a production plan through pattern popularity prediction, or realizing cross-class matching recommendation by utilizing style classification. In intelligent labeling and classifying of clothing patterns, natural wrinkles often cause pattern deformation, so that the patterns are distorted, and the traditional feature extraction method is easy to fail or the model is easy to label by mistake. Disclosure of Invention (One) solving the technical problems Aiming at the defects of the prior art, the invention provides the intelligent labeling and classifying method for the patterns of the clothing based on deep learning, which has the advantages of avoiding the influence of pattern distortion caused by wrinkles on the feature extraction and solving the problem that the traditional feature extraction method is easy to be interfered by the wrinkles to lose efficacy. (II) technical scheme In order to achieve the purpose, the invention provides the technical scheme that the intelligent labeling and classifying method for the clothing patterns based on deep learning comprises the following steps: Firstly, image data acquisition processing, namely receiving clothing image data, and carrying out standardized preprocessing on the received clothing image data to form an initial clothing image data set; Step two, positioning a fold region, namely analyzing a surface coincidence index based on each image data in an initial clothing image data set, judging whether folds exist or not based on the surface coincidence index, and positioning a region judged to exist the folds; generating a fold mask based on the positioning coordinates of the fold region, and extracting deep learning semantic features to form feature vectors after the fold interference is shielded; And step four, outputting and visualizing the result, namely outputting the formed feature vector. Preferably, the step one supports the multi-source input to receive the clothing image data, specifically includes: The method comprises the steps of acquiring local files, calling a camera of computer equipment to acquire images in real time, downloading network images through a URL, and calling an API to acquire remote image data. Preferably, in the first step, the standardized pretreatment includes: The method comprises the steps of firstly unifying file formats of images with different sources, secondly standardizing image resolution, then performing color space conversion, simultaneously eliminating noise generated by camera acquisition and noise brought by network image compression through a denoising algorithm, and finally completing normalization processing of pixel values. Preferably, the step two sets any clothing image in the initial clothing image data set asThe size isTaking any pixel pointCentering on, build the size asIs a local neighborhood window of (2);Is the neighborhood radius; Is the side length of the window. Preferably, the step two, judging whether the wrinkles exist based on the surface coincidence index includes: S1.1, opposite pixels Gray gradientDirection componentAndDirection componentCalculation with Sobel operator: ; ; Wherein, the Is convolution operation; For initial clothing image Wherein the coordinates areGray values of the pixels of (a); Pixel dot Is the gradient vector of (a)The gradient amplitude is; S1.2, vs. neighborhood WindowComputing neighborhood average gradient vectorsDirectional similarity of single pixel gradient vector to average vectorSurface coincidence indexThe method comprises the following steps: ; Wherein, the The total number of pixels in the neighborhood; for local neighborhood windows In, the coordinates areGradient vectors of pixels of (a); ; Wherein, the Is constant and takes valueDot product; ; S1.3, image pairsIs