Lace fabric image recognition method based on DualPipe algorithm architecture
By using the DualPipe algorithm architecture to extract low-level and high-level features of lace fabric in parallel, the problem of high computational resource and data requirements in lace fabric image recognition is solved, achieving efficient and accurate recognition results, and is suitable for low-power devices.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-19
- Publication Date
- 2026-03-24
AI Technical Summary
Existing technologies for lace fabric image recognition require large amounts of computational resources and training data, and the recognition results are poor, especially when running on low-power devices where efficiency is low.
The DualPipe algorithm architecture is adopted, which uses a lightweight convolutional neural network model and a deep convolutional neural network model in parallel to extract low-level and high-level features of lace fabric, and achieves recognition through feature fusion and classification layers, including ReLU activation function and Dropout layer to improve model performance.
While reducing computing resources and training time, it ensures high accuracy and efficiency in lace fabric image recognition, making it suitable for low-power devices.
Smart Images
Figure CN121725328A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision and deep learning, specifically to a method for recognizing lace fabric images based on the DualPipe algorithm architecture. Background Technology
[0002] Lace fabric, as a delicate and complex textile, is widely used in clothing, home décor, wedding dress design, and other fields, featuring intricate and varied patterns. Currently, most textile companies still rely on manual identification methods. Traditional manual classification methods are inefficient, error-prone, and unable to meet the demands of large-scale production. The continuous development of computer vision technology has made it possible to efficiently and accurately identify lace fabric using computers.
[0003] Image recognition is one of the core tasks in the field of computer vision. Its goal is to extract meaningful information from images and perform classification or detection. The development of image recognition technology can be divided into three stages: the manual recognition stage, the traditional machine learning stage, and the deep learning stage.
[0004] In the traditional machine learning phase, methods rely on manually designed feature extraction algorithms such as SIFT and HOG. These algorithms are typically designed for specific tasks and lack generality. After feature extraction, shallow machine learning models are used for classification, such as SVM, Random Forest, and KNN classifiers. While these methods offer high interpretability and require fewer computational resources, their generalization ability is limited when dealing with complex image data, such as lace fabric. In large-scale image classification, traditional methods perform far worse than deep learning methods.
[0005] Deep learning, through its hierarchical feature extraction mechanism, can progressively construct higher-level, more abstract information from shallow local information in images. This multi-layered, non-linear combination enables deep learning models to capture complex patterns and structures in images, resulting in stronger robustness and higher accuracy in recognition tasks, especially in images with complex structures such as lace fabric images. Compared to traditional hand-designed features (such as SIFT and HOG), deep learning models can automatically learn feature representations more suitable for specific tasks without relying on manually designed rules. In various fields of image recognition (such as image classification and object detection) and even in various fields of image processing (such as image generation, image inpainting, and style transfer), deep learning models have demonstrated significantly superior performance compared to traditional methods. However, deep learning models are typically large in scale, with a huge number of parameters, requiring a large amount of data for training. This leads to high demands on computing resources, especially when deployed on devices with only low-power CPUs, often resulting in low operating efficiency. Therefore, how to effectively optimize the structure of deep learning models to reduce their computational and storage requirements while maintaining performance as much as possible has become a crucial research problem. Summary of the Invention
[0006] Based on the above background, it is necessary to provide a lace fabric image recognition method based on the DualPipe algorithm architecture to address the problems of large computational resource requirements, large training data requirements, and poor image recognition performance of complex lace fabrics in existing technologies.
[0007] The steps of this method are as follows: S1: Obtain images of various types of lace fabrics with different patterns and weaving methods, and preprocess the images of the various types of lace fabrics.
[0008] S2: This invention is based on the DualPipe algorithm architecture, applying two feature extraction models to extract features in parallel. The model setup steps are as follows: S21: A convolutional neural network model with a simple network structure and few parameters, which removes the classification layer, is used to extract low-level features of the lace fabric image (such as the color, edge shape, and simple mesh structure of the lace fabric). S22: A convolutional neural network model with a complex network structure and a large number of parameters is used to extract high-level features of lace fabric images (such as complex floral patterns composed of multiple simple patterns, repeating geometric shapes, symmetry, and the global structure of lace fabric) by removing the classification layer.
[0009] S3: Perform feature fusion: fuse the collected low-level features and high-level features.
[0010] S4: The obtained fusion features are fed into the classification layer to achieve the recognition and classification of lace fabric images: S41: Set up the first linear layer to compress high-dimensional features, reduce computational complexity, and make the fused features more representative. The configuration steps for the first linear layer are as follows: S411: Introducing the ReLU activation function to perform a non-linear mapping on the result of the linear transformation, enabling the model to fit more complex functions and improve the model's expressive power; S412: Apply the Dropout layer to randomly discard a portion of the input tensor to prevent the model from overfitting during training. During training, Dropout will randomly discard the output of neurons with a certain probability, but during testing, it will not discard any neurons and will use all neurons for inference. S42: Map the features to the category space, output the predicted score for each category, classify according to these predicted scores, and select the category with the highest score as the prediction result. Beneficial effects
[0011] This invention employs a DualPipe algorithm architecture, utilizing a lightweight (simple network structure, few parameters) convolutional neural network model and a deep (complex network structure, large parameters) convolutional neural network model to extract features in parallel. This algorithm architecture can simultaneously run the lightweight model to extract simple features and the deep model to extract complex features, significantly improving the model's operating efficiency. Finally, the features extracted from the two paths are fused and classified. This parallel feature extraction method of the lightweight and deep models not only saves computational resources and model training time to a certain extent but also ensures high accuracy in lace fabric image recognition. The model training operation in this method is simple; a pre-trained convolutional neural network model can be used, eliminating the need to manually configure model weights. Only the classification layers of both models need to be removed, and the images used for training need to be resized and pre-processed. Attached Figure Description
[0012] Figure 1 This is a schematic diagram illustrating the general flow of a lace fabric image recognition method based on the DualPipe algorithm architecture of the present invention.
[0013] Figure 2 This is a detailed flowchart illustrating a lace fabric image recognition method based on the DualPipe algorithm architecture in a specific embodiment of the present invention. Detailed Implementation
[0014] This invention provides a method for recognizing lace fabric images, the method comprising the following steps: S1: Obtain images of various lace fabrics with different patterns and weaving methods, and divide the images into different folders according to their categories; S2: Preprocess the images of the various types of lace fabrics, introduce the DualPipe algorithm architecture, use a lightweight model to extract the low-level features of the images of the various types of lace fabrics, and use a deep model to extract the high-level features of the images of the various types of lace fabrics. S3: Fuse the collected low-level and high-level features; S4: The fused features are passed into the classification layer and mapped to the final category space. A nonlinear activation function is introduced to achieve the recognition of lace fabric images.
[0015] To make the technical process of the present invention clearer, the technical method of the present invention will be further described below with reference to a specific embodiment and accompanying drawings: like Figure 1 The diagram shown is a flowchart illustrating a lace fabric identification method in one embodiment. It specifically includes the following steps: Step S1: Obtain images of various types of lace fabrics and preprocess the lace fabric images: In practice, various methods can be used to collect images of existing lace fabrics, such as digital cameras and smartphones. After collecting the original images, place them in different folders according to their categories, naming the folders after the image categories, and manually create training and testing sets.
[0016] In this embodiment, the preprocessing method includes the following steps: S11: Adjust image size uniformly: The classic models such as ResNet and MobileNet have an input size of 224*224, so the image size is uniformly adjusted to 224*224.
[0017] S12: Unify image format conversion: Converting an image from its original format to PyTorch's Tensor format involves using an input image whose pixel values typically range from [0, 255] (uint8 type). The `ToTensor` function in the PyTorch library normalizes these values to [0, 1] (float32 type). The image's channel order changes from (H, W, C) (height, width, channels) to (C, H, W), which is the standard input format for PyTorch models. For example, an RGB image will be converted to a Tensor with a shape of (3, 224, 224).
[0018] S13: Normalize each color channel of the image: Standardizing each color channel of the image to make the distribution of pixel values close to a standard normal distribution helps to accelerate model convergence and improve performance.
[0019] S2: A lightweight model is used to extract low-level features from the images of the various types of lace fabrics, and a deep model is used to extract high-level features from the images of the various types of lace fabrics. S21: Remove the classification layer of the MobileNetV3 model to ensure that the MobileNetV3 model is only used to extract low-level features and not to perform classification; the MobileNetV3 model with the classification layer removed has 2 standard convolutional layers and 13 depthwise separable convolutional layers. The feature extraction part of MobileNetV3, which removes the classification layer, consists of multiple convolutional blocks. Each convolutional block includes: depthwise separable convolution (reducing computation) and an attention mechanism (enhancing important features through the Squeeze-and-Excitation module). The output of the MobileNetV3 model with the classification layer removed is a high-dimensional feature vector (with shape (batch_size, 960)) containing an abstract representation of the input image, where batch_size represents the number of batches during model training and 960 represents the dimension of the vector.
[0020] S22: Remove the classification layer of the ResNet50 model to ensure that the ResNet50 model is only used to extract low-level features and not to perform classification; the ResNet50 model with the classification layer removed has 49 convolutional layers and 16 residual blocks.
[0021] The feature extraction part of ResNet50, which removes the classification layer, consists of multiple convolutional layers and residual blocks, specifically including: an initial convolutional layer (a 7x7 convolutional layer for extracting low-level features), a max pooling layer (for downsampling), residual blocks (four-stage residual blocks, each stage containing multiple residual units, each residual unit including multiple convolutional layers and a skip connection), and a global average pooling layer (converting feature maps into feature vectors).
[0022] The output of the ResNet50 model with the classification layer removed is a high-dimensional feature vector (with shape (batch_size, 2048)) containing an abstract representation of the input image, where batch_size represents the number of batches during model training and 2048 represents the dimension of the vector.
[0023] S3: Along the channel dimension (i.e., the depth dimension of the feature map), the features extracted by the lightweight model and the deep model are merged according to the 1-dimensional channel, thereby generating a fused feature of two different levels of information. The specific operation method is as follows: The `torch.cat` function from the PyTorch library is used to concatenate the features extracted from ResNet50 (with its classification layer removed) and MobileNetV3 (with its classification layer removed) in one-dimensional channels. The shape of the concatenated feature vector is (batch_size, 2048 + 960), or (batch_size, 3008), where `batch_size` represents the number of batches used during model training, and 3008 represents the dimension of the fused feature vector.
[0024] S4: Input the obtained fused features into the classification layer of the model to complete the recognition and classification of lace fabric images: The classification layer setup includes the following steps: S41: First linear layer: Maps the concatenated high-dimensional features (3008 dimensions) to a medium-dimensional feature space (1024 dimensions) to further extract and compress features and reduce redundant information. The configuration of this linear layer includes the following steps: S411: Introducing the ReLU activation function to perform a non-linear mapping on the result of the linear transformation, enabling the model to fit more complex functions and thus improve the model's expressive power; S412: Apply a Dropout layer to randomly discard a portion of the input tensor, preventing the model from overfitting during training. Dropout randomly discards neuron outputs with a certain probability during training, but does not discard any neurons during testing, using all neurons for inference. S42: Set up the second linear layer as follows: The features are mapped to the category space, and the predicted score for each category is output. The categories are classified according to these predicted scores, and the category with the highest score is selected as the prediction result.
[0025] Those skilled in the art can understand the entire process in the above embodiments and implement any step of the process by scheduling relevant hardware through a computer program. The program can be stored in a computer-readable storage medium.
[0026] The above embodiments illustrate only one specific implementation of the present invention. Although the description is detailed, it should not be construed as limiting the scope of the present invention. Those skilled in the art can make various modifications and improvements without departing from the core spirit of the present invention, and these should all be considered within the scope of protection of the present invention. Therefore, the patent protection scope of the present invention should be determined by the contents of the claims.
Claims
1. A method for recognizing lace fabric images based on the DualPipe algorithm architecture, the method comprising the following steps: S1: Obtain images of various types of lace fabrics with different patterns and weaving methods, and preprocess the images of the various types of lace fabrics. S2: Introducing the DualPipe algorithm architecture, two feature extraction models are applied in parallel to extract the features of lace fabric images: a lightweight model is used to extract the low-level features of the multiple types of lace fabric images, and a deep model is used to extract the high-level features of the multiple types of lace fabric images. S3: The collected low-level features and high-level features are fused to obtain fused features; S4: The fused features are passed into the classification layer, the fused features are mapped to the final category space, and a nonlinear activation function is introduced to realize the recognition of lace fabric images.
2. The method according to claim 1, wherein, The preprocessing method described in step S1 includes: S11: Adjust image size uniformly; S12: Unify image format conversion; S13: Normalize each color channel of the image.
3. The method according to claim 1, wherein, In step S2, the DualPipe algorithm architecture includes the following features: Dual-path structure: The model has two paths that can perform operations simultaneously, extracting features from different angles; Feature fusion: The features extracted from the two paths are fused and then input into the subsequent classification layer for final decision-making.
4. The method according to claim 1, wherein, The fusion features in step S3 include: Low-level features: Simple information in the images of various types of lace fabrics, including the color, edge shape, and simple mesh structure of the lace fabric; Advanced features: The complex information in the images of the various types of lace fabrics, including complex floral patterns composed of multiple simple patterns, repeating geometric shapes, symmetry, and the global structure of the lace fabric.
5. The method according to claim 1, wherein, The feature extraction method described in step S2 is as follows: S21: Lightweight Model: A convolutional neural network model with fewer than 1B floating-point operations, removing the classification layer to ensure that the model is only used for feature extraction and not for classification. S22: Deep Model: A convolutional neural network model with more than 1B floating-point operations, with the classification layer removed to ensure that the model is only used for feature extraction and not for classification.
6. The method according to claim 1, wherein, The feature fusion method described in step S3 is as follows: Along the depth dimension of the feature map, the features extracted by the lightweight model and the deep model are merged according to the 1-dimensional channel, thereby generating a fused feature of two different levels of information.
7. The method according to claim 1, wherein, The classification layer mentioned in step S4 is: A fully connected layer containing two linear layers is responsible for mapping the concatenated features to the final classification space through a linear transformation, and applying an activation function to add non-linearity.
8. The method according to claim 7, wherein the step of setting the two linear layers in step S4 is as follows: S41: First linear layer: Compresses high-dimensional features to reduce computational complexity, while making the fused features more representative; S411: Introducing the ReLU activation function can perform a non-linear mapping on the result of the linear transformation, enabling the model to fit more complex functions, thereby improving the model's expressive power. S412: Apply Dropout layer: Randomly discard a portion of the input tensor to prevent the model from overfitting during training. During training, Dropout will randomly discard the output of neurons with a certain probability, while during testing, no neurons will be discarded, and all neurons will be used for inference. S42: The second linear layer maps features to the category space, outputs a predicted score for each category, classifies based on these predicted scores, and selects the category with the highest score as the prediction result.
9. A computer-readable storage medium for storing computer program instructions, characterized in that: The method as described in any one of claims 1-8 is implemented when the computer program instructions are executed by a processor.