Ultra-light UNet segmentation method and system for jacquard fabric texture

Abstract

The invention relates to the technical field of image segmentation, and discloses a jacquard fabric texture-oriented ultra-light UNet segmentation method and system, and the method comprises the steps: receiving a to-be-segmented jacquard fabric texture image through an encoder path, and carrying out the feature extraction through employing a plurality of stages of encoders which are connected in sequence; the SCAB module performs optimization based on channel attention and space attention on the feature map output by each level of encoder to obtain optimized features; the decoder path receives the feature pattern output by the last-stage encoder and the optimization features of the corresponding levels, multiple levels of decoders which are connected in sequence are adopted for feature transformation and up-sampling, and finally a segmentation result is output; the encoder and the decoder both adopt an LGB module, a convolutional layer or a PVMT module to perform feature extraction, and the LGB module comprises bottleneck convolution, a gating mechanism and channel attention; and the PVMT module comprises a plurality of Mamba models. According to the method, the characteristic extraction capability is maintained while the parameter quantity is reduced, and deep mixed characteristic representation of complex jacquard textures is realized.

Description

Technical Field

[0001] This invention relates to the field of image segmentation technology, and in particular to an ultra-lightweight UNet segmentation method and system for Jacquard fabric texture. Background Technology

[0002] Jacquard warp knitting technology is a core technology in the modern textile industry for producing high-end lace and complex patterned fabrics. In the digital design and production process of jacquard fabrics, the accurate segmentation and recognition of jacquard fabric texture images (i.e., jacquard maps) is a key link connecting pattern design and machine instructions.

[0003] With the development of computer vision technology, deep learning-based semantic segmentation methods have gradually replaced traditional manual drawing or simple thresholding algorithms, becoming the mainstream approach to improve the efficiency of Jacquard image processing. Among existing deep learning segmentation architectures, UNet and its variants are widely used due to the effectiveness of their encoder-decoder structure. Early convolutional neural network (CNN)-based methods, while excellent in extracting local texture features, were inherently limited by the local receptive field of the convolutional kernel, making it difficult to effectively understand the complex large-scale texture structures in Jacquard images. To address the lack of global contextual information, researchers introduced Transformer-based segmentation architectures. Although self-attention mechanisms can effectively capture global features, their computational complexity increases quadratically with image resolution, resulting in a huge number of model parameters, slow training and inference speeds, and extremely high memory consumption.

[0004] This makes such models difficult to deploy on industrial-grade control computers or embedded devices with limited computing resources. Recently, selective state-space models (SSMs), represented by Mamba, have attracted attention due to their linear computational complexity and excellent sequence modeling capabilities. However, directly applying standard Mamba modules to visual segmentation tasks still faces challenges: on the one hand, the number of parameters in standard Mamba models is still too large for industrial applications that pursue "ultra-lightweight" design; on the other hand, a single SSM structure is often less stable than convolutional networks when processing high-frequency details in images.

[0005] In summary, current Jacquard segmentation technology mainly faces the following technical bottlenecks: how to break the contradiction between the locality of CNN and the high computational cost of Transformer under extremely limited computing resources, while taking into account the restoration of local edge details and the aggregation of global semantic information, and achieve high-precision and high-efficiency automated segmentation, is a key technical problem that urgently needs to be solved in the current textile CAD field. Summary of the Invention

[0006] The purpose of this invention is to solve the problems in the prior art.

[0007] The technical solution adopted by this invention to solve its technical problem is: to provide an ultra-lightweight UNet segmentation method for Jacquard fabric texture, comprising the following steps:

[0008] The encoder path receives the Jacquard fabric texture image to be segmented, uses a series of interconnected multi-level encoders for feature extraction and downsampling, and outputs a multi-level encoder feature map.

[0009] The SCAB module optimizes the feature map output by each encoder level based on channel attention and spatial attention to obtain the optimized features for the corresponding level.

[0010] The decoder path receives the feature map output by the last encoder and performs feature transformation and upsampling using a series of sequentially connected decoders. Each decoder simultaneously receives the optimized features of the corresponding level; the last decoder outputs the segmentation result.

[0011] The encoder uses an LGB module, a convolutional layer, or a PVMT module for feature extraction, and the decoder uses an LGB module, a convolutional layer, or a PVMT module for feature transformation. The LGB module uses bottleneck convolution, gating mechanism, and channel attention for feature extraction. The PVMT module uses several Mamba models for feature extraction.

[0012] Preferably, the encoder path receives the jacquard fabric texture image to be segmented, and uses a multi-level encoder connected in sequence for feature extraction and downsampling, outputting a multi-level encoder feature map, represented as:

[0013] ;

[0014] in, This indicates a downsampling operation. In this case, M represents the operation of the LGB module; In this case, M represents the convolution operation; In this case, M represents the operation of the PVMT module; and Let these represent the input and output of the i-th level encoder, respectively. hour, The image represents the Jacquard fabric texture to be segmented; the set of all encoder features is used as the encoder feature map set. .

[0015] Preferably, the decoder path receives the feature map output by the last encoder, performs feature transformation and upsampling using sequentially connected multi-stage decoders, and finally outputs the segmentation result, including the following steps:

[0016] The multi-level decoder sequentially performs feature transformation and upsampling, outputting feature maps as follows:

[0017] ;

[0018] in, This represents the optimized features calculated by the encoder features of the i-th level in the SCAB module. This represents using bilinear interpolation for upsampling. In this case, M represents the operation of the LGB module; In this case, M represents the convolution operation; In this case, M represents the operation of the PVMT module; and Let represent the input and output of the i-th level decoder, respectively. hour, This represents the feature map output by the last stage encoder.

[0019] The decoded features output by the first-level decoder are mapped to a probability range of 0 to 1 using a sigmoid activation function, generating the final Jacquard fabric texture image segmentation result. , represented as:

[0020] ;

[0021] in, This represents the decoding characteristics output by the first-level decoder. This represents the Sigmoid activation function.

[0022] Preferably, the operation of the LGB module includes the following steps:

[0023] Utilizing two stacked bottleneck convolution layers on the input feature map Feature extraction is performed to obtain deep features. ;

[0024] The gated path generates gate weights through an independent branch of a bottleneck convolution layer, which are applied to the input feature map. Weighted features are obtained by weighting. , This represents a separate branch used to learn the gating weights;

[0025] Deep features Weighted features The feature map is obtained through element-wise multiplication. , represented as:

[0026] ;

[0027] in, It is an activation function; Represents element-wise multiplication operations;

[0028] The squeezing and excitation modules utilize global average pooling. Obtain channel-level statistics , represented as:

[0029] ;

[0030] Through the fully connected layer The dependencies between channels are learned, and finally channel weight vectors are generated. , represented as:

[0031] ;

[0032] The final output attention-enhanced feature map is represented as follows:

[0033] ;

[0034] in, It is a residual connection. This represents the feature map output by the LGB module.

[0035] Preferably, the operation of the PVMT module includes the following steps:

[0036] The input feature map is flattened into a sequence and then normalized through layers to obtain normalized features;

[0037] The channels of the normalized features are divided into several groups, and each group is fed into several Mamba models to extract features. The outputs of the several Mamba models are then concatenated to obtain a local feature map, represented as follows:

[0038] ;

[0039] ;

[0040] in, This represents the normalized feature of the k-th channel, which serves as the input to the k-th Mamba model. K represents the total number of Mamba models in the PVMT module. This represents the feature output by the k-th Mamba model; Represents a selective state-space model; For splicing operations; Represents a local feature map;

[0041] Normalized features are fed into the Transformer's multi-head self-attention module to aggregate global information and compute global features. , represented as:

[0042] ;

[0043] ;

[0044] in, This represents the standard scaled dot product attention mechanism; Indicating multi-head attention mechanism Dimensions representing each attention head; It captures the information correlation between any two locations in an image; query ,key ,value All are normalized features Linear projection;

[0045] Introducing the balance operator By fusing local and global representations, hybrid features are obtained. Then, features are mapped through layer normalization and linear projection layers. This completes the deep feature abstraction, represented as:

[0046] ;

[0047] ;

[0048] in, This represents element-wise addition. Representative layer normalization operation, This represents the output feature map of the PVMT module.

[0049] Preferably, the workflow of the SCAB module includes the following steps:

[0050] Based on feature map set The i-th layer encoder feature generation channel attention map Spatial attention map ;

[0051] Channel attention map Spatial attention map Spatial geometric enhancement and channel semantic weighting are performed to obtain spatial intermediate features. and channel intermediate features , respectively represented as:

[0052] ;

[0053] ;

[0054] in, Represents element-wise multiplication operations; This represents the output characteristics of the i-th layer encoder;

[0055] Spatial intermediate features and channel intermediate features Adding them together yields the corresponding set of feature maps. Optimized feature map of the i-th feature map in .

[0056] This invention also provides an ultra-lightweight UNet segmentation system for Jacquard fabric textures, comprising:

[0057] The encoding module receives the Jacquard fabric texture image to be segmented, uses a multi-level encoder connected in sequence to perform feature extraction and downsampling, and outputs a multi-level encoder feature map.

[0058] The optimization module, SCAB module, optimizes the feature map output by each level encoder based on channel attention and spatial attention to obtain the optimized features of the corresponding level.

[0059] The decoding module receives the feature map output by the last encoder and performs feature transformation and upsampling using a series of sequentially connected decoders. Each decoder simultaneously receives the optimized features of the corresponding level. The last decoder outputs the segmentation result.

[0060] The encoder uses an LGB module, a convolutional layer, or a PVMT module for feature extraction, and the decoder uses an LGB module, a convolutional layer, or a PVMT module for feature transformation. The LGB module uses bottleneck convolution, gating mechanism, and channel attention for feature extraction. The PVMT module uses several Mamba models for feature extraction.

[0061] The present invention also provides an electronic device, comprising:

[0062] One or more processors;

[0063] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to perform any of the methods described above.

[0064] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the program, when executed by a processor, implements any of the methods described above.

[0065] The present invention also provides a computer program product, comprising a computer program, characterized in that, when the computer program is executed by a processor, it implements any of the methods described above.

[0066] The present invention has the following beneficial effects:

[0067] (1) This invention innovatively proposes a lightweight gated bottleneck block (LGB-Block) module that integrates bottleneck convolution, gating mechanism and channel attention, which ensures that the model feature extraction capability is not compromised while effectively compressing the number of model parameters. This invention introduces the LGB-Block (LightweightGated Bottleneck Block) module to partially replace the encoder convolutional layer, which reduces the number of parameters while maintaining the model feature extraction capability.

[0068] (2) This invention innovatively proposes a dual-path feature extraction module strategy that uses "channel grouping Mamba" and "global Transformer" in parallel. This strategy greatly reduces the number of parameters while addressing the challenges of preserving edge details and semantic coherence of Jacquard images, thus achieving deep hybrid feature representation of complex textures in Jacquard images.

[0069] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. Attached Figure Description

[0070] Figure 1 This is a flowchart illustrating the steps of an ultra-lightweight UNet segmentation method for Jacquard fabric texture according to an embodiment of the present invention.

[0071] Figure 2 This is a flowchart illustrating an ultra-lightweight UNet segmentation method for Jacquard fabric texture according to an embodiment of the present invention.

[0072] Figure 3 This is a schematic diagram of the structure of the LGB module according to an embodiment of the present invention;

[0073] Figure 4 This is a subjective comparison diagram of the segmentation effect of the embodiments of the present invention with other methods;

[0074] Figure 5 This is a schematic diagram of the structure of an ultra-lightweight UNet segmentation system for Jacquard fabric texture according to an embodiment of the present invention;

[0075] Figure 6 A schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0076] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0077] like Figure 1 and Figure 2 As shown in the figure, an ultra-lightweight UNet segmentation method for Jacquard fabric texture according to an embodiment of the present invention includes the following steps:

[0078] S101, the encoder path receives the Jacquard fabric texture image to be segmented, uses a multi-level encoder connected in sequence to perform feature extraction and downsampling, and outputs a multi-level encoder feature map;

[0079] S102, the SCAB module optimizes the feature map output by each encoder level based on channel attention and spatial attention to obtain the optimized features of the corresponding level.

[0080] S103, the decoder path receives the feature map output by the last encoder, and uses a series of sequentially connected decoders to perform feature transformation and upsampling. Each decoder simultaneously receives the optimized features of the corresponding level; the last decoder outputs the segmentation result.

[0081] The encoder uses an LGB module, a convolutional layer, or a PVMT module for feature extraction, and the decoder uses an LGB module, a convolutional layer, or a PVMT module for feature transformation. The LGB module uses bottleneck convolution, gating mechanism, and channel attention for feature extraction. The PVMT module uses several Mamba models for feature extraction.

[0082] Specifically, in step S101, the received Jacquard fabric texture image to be processed is set as... , and Indicates image size, The input channel number is used. Given that the target region to be segmented by Jacquard mapping typically occupies about 50% of the entire image space, and even more in some scenarios, the first and second layers use a lightweight gated bottleneck block (LGB-Block) that integrates bottleneck convolution, gating mechanisms, and channel attention to extract features from the image under processing. Each level captures features at a specific scale. This process can efficiently locate and distinguish the target region from the background region while maintaining low computational complexity, thus laying a solid prior foundation for the refined segmentation of subsequent models.

[0083] BottConv, as the fundamental operational unit of LGB-Block, enables ultra-lightweight feature mapping. BottConv utilizes depthwise separable convolution to decompose standard convolutions, significantly reducing the computational cost of multiply-accumulate operations. Compared to standard convolutional Conv, BottConv achieves an exponential improvement in computational efficiency for input features. The computational complexity of BottConv can be simplified as follows:

[0084] ;

[0085] in, This indicates the number of channels in the intermediate layer of BottConv. This represents the number of output channels in the BottConv layer.

[0086] At the same time, LGB-Block introduces a gating mechanism. Modules to dynamically adjust feature maps The output enables adaptive selection of effective features. This mechanism allows the network to learn which local features are more valuable for the current segmentation task and amplify their weights while suppressing the propagation of background or noise features. In other words, it helps the network to initially locate the target region and background region of the Jacquard map.

[0087] The feature path extracts deep features through two stacked bottleneck convolutions (BottConv). The gated path generates gated weights through a single branch of BottConv. Both weights and weights are then processed using element-wise multiplication to suppress redundancy and obtain the feature map. :

[0088] ;

[0089] in, It is a feature map projected by two BottConv projections. and This represents the operation of two stacked bottleneck convolutions; It is a separate branch used to learn gating weights; It is an activation function; This represents element-wise multiplication operations.

[0090] To further enhance the module's ability to locate the Jacquard target and background regions, LGB-Block embeds a Squeeze-and-Excitation Block (SE Block), which utilizes global average pooling. Obtain channel-level statistics Then through two fully connected layers The dependencies between channels are learned, and finally channel weight vectors are generated. :

[0091] ;

[0092] ;

[0093] in, and These represent two fully connected layers respectively. The bias term; It is ReLU6; For the Sigmoid function; , The weights are 1×1 convolution weights.

[0094] In summary, LGB-Block output feature maps with attention enhancement It can be represented as:

[0095] ;

[0096] in, Represents the SE module; This represents the identity mapping, i.e., a 1×1 convolutional projection; It uses residual connections to ensure the stability of the gradient flow in the module.

[0097] Each encoder stage downsamples the extracted features to obtain the encoder output features.

[0098] To enhance the first The texture and edge information of the output feature map of the multi-level encoder is further enhanced by a convolutional layer to obtain a multi-level feature map. .

[0099] Specifically, for multi-level feature maps Spatiotemporal feature modeling is performed to capture the local fine-grained dependencies and global long-distance dependencies of features at different levels, resulting in feature sets adapted to different levels.

[0100] To address the challenge of maintaining both edge detail preservation and semantic coherence in Jaka images due to their complex microscopic textures and macroscopic structural layouts, which makes it difficult for existing lightweight models to achieve this balance with low computational power, this paper innovatively proposes a dual-path feature extraction module (PVMT-Module) that combines "channel grouping Mamba" and "global Transformer". This module leverages the efficient linear complexity of channel grouping Mamba to accurately capture local high-frequency texture dependencies, solving the boundary definition problem in complex Jaka image textures. Simultaneously, it combines the multi-head attention mechanism of the Transformer to establish global long-range associations, ensuring the semantic consistency of Jaka images. Through this complementary fusion of local acuity and global receptive field, a deep hybrid feature representation of the complex textures of Jaka images is achieved.

[0101] The multi-level feature maps will sequentially enter the encoder at levels 4 to 6, with each level employing a parallel strategy of "channel grouping Mamba" and "global Transformer". In each level, the feature maps are first processed... Flatten into a sequence And through layer normalization to This prepares for subsequent spatiotemporal feature modeling; The process will proceed in parallel through two paths: one is through a channel-grouping Mamba model to capture local features, and the other is through a Transformer model with multi-head attention to extract global features. Normalized features Represented as:

[0102] ;

[0103] in, Representative layer normalization operation; for The flattened sequence has a length of .

[0104] Specifically, one path employs a channel grouping strategy to address the issue that the standard Mamba model's parameter count and computational overhead are insufficient for lightweight design requirements. The parameter complexity of the Mamba module is primarily determined by the number of input channels. Decision. In this path, put... Number of input channels The input channels are divided into four parts, each with its own features captured locally using the Mamba model before being concatenated to restore the total number of input channels. This enables parallel Mamba implementation. That is, it... Along the channel dimension Divided into 4 subspaces ,in , This is the batch size. For each subspace Using a selective state-space model Perform characteristic evolution. First, express the state equation of the continuous system as:

[0105] ;

[0106] ;

[0107] in, Represents a hidden state; Represents the state transition matrix; Represents the input control matrix; This is for outputting the projection matrix.

[0108] Then, using the zero-order hold principle and time scale parameters... Discretize the above equation into a recursive form. The state transition matrix after discretization is... Input control matrix and each subspace Hidden state That is:

[0109] ;

[0110] ;

[0111] ;

[0112] in, Represent the identity matrix, dimension, and transition matrix. same.

[0113] Then through learnable residual shrinkage After that, you can get the first Output of each subspace block and the local features ultimately captured The specific calculation is as follows:

[0114] ;

[0115] ;

[0116] in, Represents a selective state-space model; This is for splicing operations.

[0117] Specifically, another approach uses the Transformer's multi-head self-attention module to aggregate global information, thus compensating for SSM's potential shortcomings in capturing global context. The program will then proceed in parallel to another path, utilizing the standard scaled dot product attention mechanism to compute global features. Define the query. ,key ,value All Linear projection:

[0118] ;

[0119] ;

[0120] in, Dimensions representing each attention head; It captures the information correlation between any two locations in the image.

[0121] Then, by introducing a balancing operator By fusing local and global representations, hybrid features are obtained. Subsequently, features are mapped through layer normalization and linear projection layers. This completes the deep feature abstraction, represented as:

[0122] ;

[0123] ;

[0124] in, Representative layer normalization operation.

[0125] By abstracting features through multiple levels, a set of features for each encoder is finally obtained, which serves as the encoder output feature map. .

[0126] Specifically, in S102, the feature sets adapted to different levels are bridged to generate multi-scale optimized feature sets for multiple levels of the encoder.

[0127] To address the challenge of complex Jakarta background textures, often exhibiting high texture similarity to the target region to be segmented, which easily introduces background noise interference when directly utilizing the original features, this invention constructs a Spatial-Channel Attention Bridge (SCAB) between the encoder and decoder for multi-scale information optimization and attention filtering. This module, acting as a crucial hub for feature transmission, performs dual-dimensional filtering of input features through spatial geometric enhancement and channel semantic weighting, achieving deep optimization of multi-scale information. The SCAB module receives feature maps output from six levels in the encoder path. It covers everything from shallow, high-resolution texture details to deep, low-resolution semantic generalizations.

[0128] Spatial geometry enhancement extracts salient regions of features in the spatial dimension through a shared convolutional network in parallel. Utilizing a dual-path branch of average pooling and max pooling, it captures high-frequency edges and low-frequency background in space, generating a spatial attention map. This is actually a spatial filter that can identify and enhance the geometric boundaries of a target while suppressing background regions irrelevant to the task, thereby obtaining intermediate spatial features. ,Right now:

[0129] ;

[0130] in, This represents the output characteristics of the i-th layer encoder; Representing the Spatial attention map generated from the output feature maps of each level ; This represents element-wise multiplication operations.

[0131] Channel semantic weighting concatenates features from all scales after pooling, calculates cross-scale channel attention, adaptively learns the importance of feature channels at each level using fully connected layers, and finally generates a channel attention map using a sigmoid activation function. This attention map quantifies the contribution of each channel to the final segmentation task. Specifically, the model automatically suppresses channels containing background noise or redundant information while enhancing channels containing key features, thus obtaining intermediate features for each channel. The calculation process is as follows:

[0132] ;

[0133] in, Representing the Channel attention maps generated from the output feature maps of each level .

[0134] Ultimately, the intermediate features of the space and channel intermediate features Add them together to form a multi-scale optimized feature set. This fusion process achieves comprehensive feature recalibration: in the channel dimension, key feature channels are significantly amplified, redundant channels carrying background noise are effectively suppressed, solving the problem that effective information in Jacquard textures is easily buried; in the spatial dimension, the target area is highlighted, and its boundary information with the background is clearly enhanced.

[0135] This step delivers high-quality, "cleaned" features, enabling the decoder to obtain more accurate prior guidance when fusing with upsampled features. This greatly improves the model's segmentation accuracy and edge recovery capabilities in complex texture scenes such as Jacquard maps.

[0136] Specifically, in S103, the multi-scale optimized feature set is input into the decoder path. Through a deep-to-shallow cascading strategy, the features of the previous level are transformed and upsampled. Combined with the multi-scale feature set, the Jacquard segmentation image is reconstructed.

[0137] The multi-scale feature set optimized by the SCAB module The input path is sent to the decoder. Addressing the challenge of precisely segmenting target region edges in Jacquard maps where they are often intertwined with background lines, this invention designs a strictly symmetrical decoding structure with the encoder, employing a cascaded recovery strategy from deep to shallow. Each decoding level follows a specific sequence of "feature transformation first, then spatial upsampling." First, the decoder module performs a nonlinear transformation and integration of the previous level's features. Then, bilinear interpolation is used for upsampling to recover the spatial dimensions of the feature map.

[0138] The optimized multi-scale feature set is input into the decoder path, which is structurally strictly symmetrical with the encoder path. The decoding process employs a deep-to-shallow cascading strategy. The model first transforms the previous-level features through the decoder module, and then uses bilinear interpolation for upsampling to restore the spatial dimensions.

[0139] The upsampled features are not directly output, but are added pixel-by-pixel residuals to the same-scale optimized feature set from the SCAB module. Addressing the issue of complex jacquard fabric textures and the tendency to lose high-frequency details in deep networks, this invention abandons the traditional channel stitching method and instead employs an additive fusion strategy. This additive fusion strategy, while maintaining a concise channel count, forces the decoder to learn the differences between deep features and shallow details, thereby effectively mitigating the gradient vanishing problem and enhancing feature robustness. Specifically, the decoded features... Represented as:

[0140] ;

[0141] in, This represents the optimized features calculated by the SCAB module based on the encoder features of the i-th level; This represents an upsampling operation using bilinear interpolation; M represents the feature extraction operation, which varies in different levels of the decoder, as follows:

[0142] , ;

[0143] , ;

[0144] , ;

[0145] in, This indicates the operation of the PVMT module; This indicates the operation of the LGB module; Indicates the convolution operation; and These represent the input and output of the i-th level decoder, respectively. hour, This represents the feature map output by the last stage encoder.

[0146] The decoded features output by the first-level decoder are mapped to a probability range of 0 to 1 using a sigmoid activation function, generating the final Jacquard fabric texture image segmentation result. , represented as:

[0147] ;

[0148] in, This represents the decoding characteristics output by the first-level decoder. This represents the Sigmoid activation function.

[0149] In this embodiment, an ultra-lightweight UNet segmentation method and system for Jacquard fabric texture image segmentation is constructed. Its core idea is to utilize LGB-Block, which integrates bottleneck convolution, gating mechanisms, and channel attention, to achieve efficient feature extraction with low computational complexity. Furthermore, a strategy of parallel "channel grouping Mamba" and "global Transformer" is employed to establish a deep feature representation that considers both local fine-grained dependencies and global long-range dependencies.

[0150] Specifically, the method of this invention first utilizes the LGB-Block module to perform multi-level shallow feature extraction on the input Jacquard texture image, and achieves adaptive selection of effective features and noise suppression through gating mechanism and SE Block; then, the PVMT-Module is proposed, which utilizes the linear complexity advantage of channel grouping Mamba in conjunction with the global receptive field of Transformer to aggregate deep information and generate a hybrid feature set adapted to different resolution levels; at the encoder-decoder connection, the Spatial-Channel Attention Bridge (SCAB) is introduced to recalibrate the encoded features to "remove falsehoods and retain truth", realizing spatial geometric enhancement and channel semantic weighting; finally, through the cascaded decoder path from deep to shallow, the optimized features and upsampled features are fused pixel by pixel residual, thereby significantly improving the segmentation accuracy and edge robustness of the model in complex Jacquard texture scenes.

[0151] Specifically, the verification experiments of this embodiment were conducted on an NVIDIA RTX 4090 GPU server, with the training process built based on the PyTorch framework. The optimizer used was Adamw, and the main parameter settings included: loss function BceDiceLoss, input image size of 256×256, input channels of 3, batch size of 64, number of iterations of 400, and specific parameters such as average Intersection over Union (IoU), accuracy, sensitivity, and specificity output every 20 iterations. The method of this invention was comprehensively compared with representative algorithms of current state-of-the-art lightweight and high-performance segmentation networks. Multiple comparative experiments were conducted on the Jacquard fabric texture dataset, verifying the comprehensive advantages of the proposed method in terms of computational complexity and segmentation accuracy. Detailed information is shown in Tables 1 and 2. In this experiment, the profile function in the thop library is used to evaluate the model complexity of the method of this invention and other algorithms; Yolov11x-seg represents the version with the best Yolov11 segmentation performance, and Yolov11n-seg represents the version with the smallest Yolov11 segmentation parameters.

[0152] Table 1 - Comparison of segmentation accuracy metrics with other state-of-the-art methods on the Jacquard fabric texture dataset:

[0153]

[0154] Table 2 - Comparison of computational complexity metrics with other state-of-the-art methods on the Jacquard fabric texture dataset:

[0155]

[0156] As shown in Tables 1 and 2, our proposed method demonstrates superior overall performance on the Jacquard fabric texture dataset. In terms of segmentation accuracy, our method comprehensively outperforms the comparison models, achieving an IoU of 0.9794 and a Dice coefficient as high as 0.9896. Compared to the Yolov11x-seg algorithm, our method improves IoU and Dice by 0.76% and 0.39%, respectively, and also enhances specificity by approximately 1 percentage point (reaching 0.9828), demonstrating the algorithm's superiority in handling complex textures and edge details.

[0157] In terms of computational complexity, this method demonstrates significant lightweight advantages. The number of parameters is only 0.0359M, a staggering 99.94% reduction compared to YOLOv11x-seg's 62.0044M. Simultaneously, the computational cost (MACs) is only 107.889M, far lower than the YOLO series and even about 37.2% lower than similar ultra-lightweight networks like UltraLight-VM-UNet. Although the inference speed (71.71 FPS) is slightly lower than UltraLight-VM-UNet, this method achieves a significant improvement in accuracy while drastically reducing computational cost, realizing an optimal balance between resource consumption and segmentation accuracy, and possesses great potential for deployment on edge devices.

[0158] In addition to comparing segmentation performance and computational complexity metrics, this embodiment also compares subjective visualization results (such as...). Figure 4 As shown in the figure, the segmentation accuracy advantage of the proposed method is intuitively demonstrated: when dealing with regions with complex topological structures (such as small hollows inside the shoe upper, irregular edge contours, etc.), compared with the details loss (such as the adhesion of UltraLight-VM-UNet in hollows) or missed detection (such as the missed segmentation of solid regions by the Yolov11 series) phenomenon of the comparison method, the segmentation mask generated by this method is closest to the label, and can more clearly restore the geometric features inside the target, proving the advancement of the proposed method.

[0159] See Figure 5 As shown, this application provides an embodiment of an ultra-lightweight UNet segmentation system for Jacquard fabric textures, and... Figure 1 Corresponding to the illustrated method embodiments, the methods include:

[0160] Encoding module 501 receives the Jacquard fabric texture image to be segmented via encoder path, performs feature extraction and downsampling using sequentially connected multi-level encoders, and outputs multi-level encoder feature map;

[0161] The optimization module 502 and the SCAB module optimize the feature map output by each encoder level based on channel attention and spatial attention to obtain the optimized features of the corresponding level.

[0162] The decoding module 503 receives the feature map output by the last encoder and performs feature transformation and upsampling using a series of sequentially connected decoders. Each decoder simultaneously receives the optimized features of the corresponding level. The last decoder outputs the segmentation result.

[0163] The encoder uses an LGB module, a convolutional layer, or a PVMT module for feature extraction, and the decoder uses an LGB module, a convolutional layer, or a PVMT module for feature transformation. The LGB module uses bottleneck convolution, gating mechanism, and channel attention for feature extraction. The PVMT module uses several Mamba models for feature extraction.

[0164] See Figure 6 The diagram shows a hardware structure of an electronic device provided in an embodiment of the present invention, including a processor 601 and a memory 602; wherein the memory 602 is used to store computer execution instructions; and the processor 601 is used to execute the computer execution instructions stored in the memory to implement the various steps performed by the electronic device in the above embodiment. For details, please refer to the relevant descriptions in the foregoing method embodiments.

[0165] Alternatively, the memory 602 can be either standalone or integrated with the processor 601.

[0166] When the memory 602 is set up independently, the electronic device also includes a bus 603 for connecting the memory 602 and the processor 601.

[0167] This invention also provides a computer storage medium storing computer execution instructions, which, when executed by a processor, implement the method described above.

[0168] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0169] In the embodiments provided by this invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0170] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0171] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0172] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods of the various embodiments of this application.

[0173] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.

[0174] The memory may include high-speed RAM, and may also include non-volatile storage (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk or optical disc, etc.

[0175] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0176] The aforementioned storage medium can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium accessible to general-purpose or special-purpose computers.

[0177] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. Both the processor and the storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic device or host device.

[0178] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0179] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A lightweight UNet segmentation method for Jacquard fabric texture, characterized in that, Includes the following steps: The encoder path receives the Jacquard fabric texture image to be segmented, uses a series of interconnected multi-level encoders for feature extraction and downsampling, and outputs a multi-level encoder feature map. The SCAB module optimizes the feature map output by each encoder level based on channel attention and spatial attention to obtain the optimized features for the corresponding level. The decoder path receives the feature map output by the last encoder and performs feature transformation and upsampling using a series of sequentially connected decoders. Each decoder simultaneously receives the optimized features of the corresponding level; the last decoder outputs the segmentation result. The encoder uses an LGB module, a convolutional layer, or a PVMT module for feature extraction, and the decoder uses an LGB module, a convolutional layer, or a PVMT module for feature transformation. The LGB module uses bottleneck convolution, gating mechanism, and channel attention for feature extraction. The PVMT module uses several Mamba models for feature extraction.

2. The ultra-lightweight UNet segmentation method for Jacquard fabric texture according to claim 1, characterized in that, The encoder path receives the Jacquard fabric texture image to be segmented, and uses a series of sequentially connected multi-level encoders for feature extraction and downsampling, outputting a multi-level encoder feature map, represented as follows: ； in, This indicates a downsampling operation. In this case, M represents the operation of the LGB module; In this case, M represents the convolution operation; In this case, M represents the operation of the PVMT module; and Let these represent the input and output of the i-th level encoder, respectively. hour, The image represents the Jacquard fabric texture to be segmented; the set of all encoder features is used as the encoder feature map set. .

3. The ultra-lightweight UNet segmentation method for Jacquard fabric texture according to claim 2, characterized in that, The decoder path receives the feature map output by the last encoder, performs feature transformation and upsampling using a series of sequentially connected decoders, and finally outputs the segmentation result, including the following steps: The multi-level decoder sequentially performs feature transformation and upsampling, outputting feature maps as follows: ； in, This represents the optimized features calculated by the encoder features of the i-th level in the SCAB module. This represents using bilinear interpolation for upsampling. In this case, M represents the operation of the LGB module; In this case, M represents the convolution operation; In this case, M represents the operation of the PVMT module; and Let represent the input and output of the i-th level decoder, respectively. hour, This represents the feature map output by the last stage encoder. The decoded features output by the first-level decoder are mapped to a probability range of 0 to 1 using a sigmoid activation function, generating the final Jacquard fabric texture image segmentation result. , represented as: ； in, This represents the decoding characteristics output by the first-level decoder. This represents the Sigmoid activation function.

4. The ultra-lightweight UNet segmentation method for Jacquard fabric texture according to claim 1, characterized in that, The operation of the LGB module includes the following steps: Utilizing two stacked bottleneck convolution layers on the input feature map Feature extraction is performed to obtain deep features. ; The gated path generates gate weights through an independent branch of a bottleneck convolution layer, which are applied to the input feature map. Weighted features are obtained by weighting. , This represents a separate branch used to learn the gating weights; Deep features Weighted features The feature map is obtained through element-wise multiplication. , represented as: ； in, It is an activation function; Represents element-wise multiplication operations; The squeezing and excitation modules utilize global average pooling. Obtain channel-level statistics , represented as: ； Through the fully connected layer The dependencies between channels are learned, and finally channel weight vectors are generated. , represented as: ； The final output attention-enhanced feature map is represented as follows: ； in, It is a residual connection. This represents the feature map output by the LGB module.

5. The ultra-lightweight UNet segmentation method for Jacquard fabric texture according to claim 1, characterized in that, The operation of the PVMT module includes the following steps: The input feature map is flattened into a sequence and then normalized through layers to obtain normalized features; The channels of the normalized features are divided into several groups, and each group is fed into several Mamba models to extract features. The outputs of the several Mamba models are then concatenated to obtain a local feature map, represented as follows: ； ； in, This represents the normalized feature of the k-th channel, which serves as the input to the k-th Mamba model. K represents the total number of Mamba models in the PVMT module. This represents the feature output by the k-th Mamba model; Represents a selective state-space model; For splicing operations; Represents a local feature map; Normalized features are fed into the Transformer's multi-head self-attention module to aggregate global information and compute global features. , represented as: ； ； in, This represents the standard scaled dot product attention mechanism; Indicating multi-head attention mechanism Dimensions representing each attention head; It captures the information correlation between any two locations in an image; query ,key ,value All are normalized features Linear projection; Introducing the balance operator By fusing local and global representations, hybrid features are obtained. Then, features are mapped through layer normalization and linear projection layers. This completes the deep feature abstraction, represented as: ； ； in, This represents element-wise addition. Representative layer normalization operation, This represents the output feature map of the PVMT module.

6. The ultra-lightweight UNet segmentation method for Jacquard fabric texture according to claim 1, characterized in that, The workflow of the SCAB module includes the following steps: Based on feature map set The i-th layer encoder feature generation channel attention map Spatial attention map ; Channel attention map Spatial attention map Spatial geometric enhancement and channel semantic weighting are performed to obtain spatial intermediate features. and channel intermediate features , respectively represented as: ； ； in, Represents element-wise multiplication operations; This represents the output characteristics of the i-th layer encoder; Spatial intermediate features and channel intermediate features Adding them together yields the corresponding set of feature maps. Optimized feature map of the i-th feature map in .

7. An ultra-lightweight UNet segmentation system for jacquard fabric textures, characterized in that, include: The encoding module receives the Jacquard fabric texture image to be segmented, uses a multi-level encoder connected in sequence to perform feature extraction and downsampling, and outputs a multi-level encoder feature map. The optimization module, SCAB module, optimizes the feature map output by each level encoder based on channel attention and spatial attention to obtain the optimized features of the corresponding level. The decoding module receives the feature map output by the last encoder and performs feature transformation and upsampling using a series of sequentially connected decoders. Each decoder simultaneously receives the optimized features of the corresponding level. The last decoder outputs the segmentation result. The encoder uses an LGB module, a convolutional layer, or a PVMT module for feature extraction, and the decoder uses an LGB module, a convolutional layer, or a PVMT module for feature transformation. The LGB module uses bottleneck convolution, gating mechanism, and channel attention for feature extraction. The PVMT module uses several Mamba models for feature extraction.

8. An electronic device, comprising: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to perform the method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-6.

Images