A cloth surface transparent adhesive tape target detection method and system

Abstract

The present application relates to the technical field of computer vision and textile engineering, and particularly relates to a cloth surface transparent tape target detection method and system. The method comprises the following steps: obtaining a cloth surface transparent tape picture data set, preprocessing the data set; constructing a cloth surface transparent tape detection model, training the detection model by using the preprocessed data set; and completing cloth surface transparent tape target detection by using the trained detection model. The present application can effectively solve the problem of recognition difficulty caused by the similar background texture and high reflectivity of the transparent tape, and has important practical significance for improving the enterprise automatic sorting efficiency.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of computer vision and textile engineering, specifically to a method and system for detecting targets on the surface of transparent adhesive tape. Background Technology

[0002] In the textile manufacturing industry, fabric quality grade is a key factor determining its market value. A textile company uses transparent tape for winding and marking fabric rolls during the roll packaging process: tape is wrapped around both ends of the roll for secure winding, and the middle section is marked as superior grade, while the rest is marked as inferior grade. To further optimize the sorting process, the company introduced an automated sorting device with an integrated vision inspection system to sort the fabrics during transport according to their grade. However, during the sorting of some yarn-dyed fabrics, the transparent tape's translucency, similar to the background fabric color, resulted in low image contrast; simultaneously, its reflective properties further interfered with visual recognition, leading to a high false positive rate. Therefore, developing a target detection algorithm that can effectively overcome the low contrast and reflective interference of transparent tape is of significant practical importance for improving the efficiency of automated sorting in enterprises. Summary of the Invention

[0003] To address the shortcomings of existing methods and the needs of practical applications, and in order to solve the aforementioned problems, this invention provides a method for detecting transparent adhesive tape targets on fabric surfaces, comprising the following steps: This invention involves acquiring a dataset of images of transparent tape on fabric surfaces, preprocessing the dataset, building a detection model for transparent tape on fabric surfaces, and training the detection model using the preprocessed dataset. The trained detection model is then used to detect the transparent tape on the fabric surface. This invention first acquires images of transparent tape on fabric surfaces under various materials, lighting conditions, and application methods, constructs a dataset, and preprocesses it (including normalization and geometric / lighting / noise enhancement). Then, a first feature extraction network with specific sub-modules is designed, and the encoder is optimized (adding an anti-reflection attention module and replacing RepC3 with DGST). A combined loss function is then used to construct and train a detection model. Finally, the trained model is used for detection, effectively solving the problems of difficult feature extraction and poor generalization caused by the reflectivity of transparent tape.

[0004] Optionally, the preprocessing of the dataset includes data augmentation operations.

[0005] Optionally, the construction of the transparent adhesive tape detection model on the fabric surface includes the following steps: This invention designs a first feature extraction network; incorporates an anti-reflective attention fusion module into the cross-scale feature fusion structure of an efficient hybrid encoder; and sets a bounding box loss function. By designing a first feature extraction network, this invention captures subtle features of transparent tape using specific sub-modules and an ELGCA layer, adapting to various fabric textures; adds an anti-reflective attention fusion module to suppress tape reflection interference and prevent useful features from being overwhelmed; and sets a bounding box loss function to balance regression accuracy and robustness, improving positioning accuracy. These three elements work synergistically to significantly improve detection accuracy and generalization ability, providing key technical support for industrial fabric tape detection.

[0006] Optionally, the first feature extraction network first completes preliminary downsampling feature extraction through a first submodule; Then, the feature extraction stage proceeds to four levels, each consisting of a first sub-module and a second sub-module of downsampling in sequence.

[0007] Optionally, the first submodule consists of a 3×3 convolutional layer, a batch normalization layer, and a Silu activation function.

[0008] Optionally, the second submodule is used to perform preliminary convolution and normalization processing on the input features through the first submodule; Subsequently, the feature map is divided into two paths: one serves as a shortcut to preserve local details, and the other is input into a deep feature extraction channel consisting of N ELGCA bottleneck layers. Finally, the deep features processed by the ELGCA network are concatenated with the shortcut connection branches in the channel dimension, and feature fusion is performed through the first submodule at the end.

[0009] Optionally, the cross-scale feature fusion structure of the efficient hybrid encoder further includes replacing the RepC3 module in the cross-scale feature fusion structure with a DGST module; the DGST module realizes parameter compression and cross-group feature interaction through a dynamic group selection module, and dynamically models the local and global context relationship in conjunction with a convolutional feedforward network; the dynamic group selection module first compresses the number of channels of the input feature map to half of the original number through a 1×1 convolution; then the compressed features are simultaneously fed into two parallel branches: one branch directly serves as an identity mapping, preserving the integrity of the input information; the other branch sequentially performs lightweight spatial feature extraction through a 3×3 group convolution, followed by a channel rearrangement operation; finally, the outputs of the two branches are fused by channel concatenation.

[0010] By compressing the channels to half using 1×1 convolution, the computational load is significantly reduced, achieving a lightweight model. In the dual-branch design, the identity mapping preserves the integrity of the input information and avoids feature loss. The 3×3 grouped convolution combined with channel rearrangement can enhance spatial feature extraction on the basis of lightweight design, adapting to the subtle features of transparent tape. Finally, channel splicing and fusion improve the quality of cross-scale features, which not only meets the needs of real-time industrial detection but also helps to improve the detection accuracy of transparent tape.

[0011] Optionally, the anti-reflection attention fusion module generates three-dimensional attention weights through an energy function. Generating three-dimensional attention weights through an energy function assesses the importance of neurons, highlighting the role of key neurons, thereby assigning higher priority to information-rich neurons in visual processing, which differs from traditional one-dimensional or two-dimensional channel or spatial attention mechanisms. Optionally, the bounding box loss function is set to satisfy the following formula: in, Represents the bounding box loss function. Indicates the weighting coefficient. Represents the SmoothL1 loss function. This represents the FocalEIoU loss function. The bounding box loss function, based on a weighted combination of SmoothL1 and FocalEIoU, solves the problem of the traditional L1 loss function being non-differentiable at zero and overcomes the deficiency of the GIoU loss function in distinguishing the relative positions of targets, thereby improving the model's localization accuracy.

[0012] Secondly, to efficiently execute the method for detecting transparent adhesive tape on a fabric surface provided by this invention, this invention also provides a system for detecting transparent adhesive tape on a fabric surface, including a processor, an input device, an output device, and a memory. The processor, input device, output device, and memory are interconnected. The memory stores a computer program containing program instructions. The processor is configured to call the program instructions to execute the method for detecting transparent adhesive tape on a fabric surface as described in the first aspect of this invention. The system for detecting transparent adhesive tape on a fabric surface provided by this invention has a compact structure and stable performance, and can stably execute the method for detecting transparent adhesive tape on a fabric surface provided by this invention, further improving the overall applicability and practical application capability of this invention. Attached Figure Description

[0013] Figure 1 A flowchart of a method for detecting transparent adhesive tape targets on a fabric surface provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of a first feature extraction network provided in an embodiment of the present invention; Figure 3 This is a framework diagram of a transparent adhesive tape target detection system for fabric surfaces provided in an embodiment of the present invention. Detailed Implementation

[0014] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0015] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0016] Please see Figure 1 To address the above problems, this invention provides a method for detecting transparent adhesive tape targets on fabric surfaces, such as... Figure 1 As shown, in one embodiment, the method includes the following steps: S01. Obtain a dataset of images of transparent tape on the fabric surface and preprocess the dataset.

[0017] In this embodiment, based on the actual production scenario of a textile enterprise, an A3200CG000 industrial camera, equipped with a PFM-TX55238W-A-1A1 white strip industrial light source and an MHO628SP industrial lens, was used to collect dataset images.

[0018] Specifically, fabric samples of different materials (cotton, linen, silk, chemical fiber), colors (white, light gray, beige, etc., which are easily confused with transparent tape), and textures (plain weave, twill weave, jacquard) were collected. Under different lighting conditions (natural light, workshop LED light), transparent tape was attached to the fabric surface in different ways (complete coverage, edge folding, partial adhesion) to build a dataset containing 5,000-10,000 images. The bounding boxes of the transparent tape were labeled using tools.

[0019] Furthermore, the dataset is preprocessed to normalize the pixel values ​​to [0,1].

[0020] Furthermore, data augmentation operations are performed, including but not limited to: Geometric transformations: random rotation (0-360°), horizontal / vertical flip (probability 0.5), scaling (0.8-1.2x); Lighting adjustment: brightness (±20%), contrast (±15%), color temperature (±100K) variations; Noise additions: Gaussian noise (variance 0.01-0.03) and salt-and-pepper noise (density 0.001-0.005) to improve the model's generalization ability.

[0021] Furthermore, the dataset is divided into a training set, a validation set, and a test set in a ratio of 7:2:1.

[0022] S02. Construct a detection model for transparent adhesive tape on the fabric surface, and train the detection model using the preprocessed dataset.

[0023] In this embodiment, the construction of the transparent adhesive tape detection model on the fabric surface includes the following steps: First, design the first feature extraction network.

[0024] Please see Figure 2 The first feature extraction network first completes preliminary downsampling feature extraction through a first submodule; then it enters a four-level feature extraction stage, each level consisting of a downsampling first submodule and a second submodule sequentially. The output of the second submodule is denoted as S3, the output of the third submodule as S4, and the output of the fourth submodule as S5.

[0025] The first submodule consists of a 3×3 convolutional layer, a batch normalization layer, and a Silu activation function. The convolutional layer is used to extract local spatial features, the batch normalization layer accelerates network convergence and improves training stability through normalization, and the Silu activation function introduces non-linear features to enhance the network's expressive power.

[0026] The second submodule first performs preliminary convolution and normalization processing on the input features through the first submodule; Subsequently, the feature map is divided into two paths: one serves as a shortcut to preserve local details, and the other is input into a deep feature extraction channel consisting of N ELGCA bottleneck layers. Finally, the deep features processed by the ELGCA network are concatenated with the shortcut connection branches in the channel dimension, and feature fusion is performed through the first submodule at the end.

[0027] The ELGCA network employs local and global parallel branches: the local branch uses 3×3 depthwise separable convolutions to extract fine-grained contextual information, while the global branch simultaneously applies average pooling and max pooling to converge two different types of global contextual information and inputs this information into the PTAttention unit to capture long-range dependencies. The outputs of the two branches are fused along the channel dimension to form a feature representation that balances lightweightness and expressiveness.

[0028] Second, an anti-reflection attention fusion module is added to the cross-scale feature fusion structure of the efficient hybrid encoder.

[0029] In this embodiment, the high-efficiency hybrid encoder consists of two network structures: an AIFI module and a cross-scale feature fusion structure.

[0030] Specifically, the AIFI module operates as follows: Input feature S5 first enters a multi-head self-attention mechanism to achieve global context modeling. Then, the self-attention output is added to the original input through a first residual connection, followed by layer normalization to enhance training stability while preserving the original semantic information. Next, the features undergo non-linear mapping and feature enhancement via a feedforward neural network, and then the output is added to the normalized result of the previous layer through a second residual connection and normalized again to obtain output F5.

[0031] The working principle of the cross-scale feature fusion structure is as follows: F5 is sequentially passed through a 1×1 convolution module and an upsampling module to obtain output U1; U1 and the features of S4 processed by the 1×1 convolution module are fused for the first time in the anti-reflection attention fusion module to obtain output R1; R1 is sequentially passed through a DGST module, a 1×1 convolution module, and an upsampling module to obtain output U2; U2 and the features of S3 processed by the 1×1 convolution module are fused for the second time in the anti-reflection attention fusion module to obtain output R2; R2 is sequentially passed through a DGST module and a downsampling module. The output D1 is obtained. The features of D1 and R1 after being processed by the DGST module and the 1×1 convolution module are fused for the third time in the anti-reflection attention fusion module to obtain the output R3. R3 is then processed by the DGST module and the downsampling module to obtain the output D2. The features of D2 and F5 after being processed by the 1×1 convolution module are fused for the fourth time in the anti-reflection attention fusion module to obtain the output R4. Finally, R4, R3 and R2 are processed by the DGST module and then fused for the last time in the anti-reflection attention fusion module to obtain the output R5, which is then input to the detection head.

[0032] Furthermore, the DGST module achieves parameter compression and cross-group feature interaction through a dynamic group selection module, and combines a convolutional feedforward network to dynamically model local and global context relationships, thereby enhancing the flexibility of feature expression and capturing spatial correlations between features.

[0033] The dynamic grouping selection module first compresses the number of channels in the input feature map to half of the original number using a 1×1 convolution; The compressed features are then fed into two parallel branches simultaneously: one branch acts directly as an identity mapping, preserving the integrity of the input information; the other branch sequentially performs lightweight spatial feature extraction through a 3×3 grouped convolution, followed by a channel rearrangement operation, thereby breaking the information isolation caused by the grouped convolution.

[0034] Finally, the outputs of the two branches are fused by channel concatenation, combining the untransformed original information with the nonlinearly refined features to achieve a balance between feature reuse and efficient feature extraction.

[0035] Furthermore, the convolutional feedforward network first feeds the input features into two parallel paths: one branch acts as the core feature enhancer, first projecting and transforming the channel dimension through a 1×1 convolution, and then directly feeding the output into a second 1×1 convolution for a second feature transformation; the other branch directly passes the original input features without any transformation. Finally, the outputs of these two branches are element-wise added through residual connections, fusing the non-linearly enhanced high-order features with the original features that retain complete details.

[0036] The detection head first applies a minimum uncertainty query selection mechanism to select a fixed number of image features with the highest uncertainty from the features output by the efficient hybrid encoder, in order to initialize the target query of the decoder. In this embodiment, the detection head gradually adjusts the query through an iterative optimization process to more accurately predict the target bounding box and class confidence, thereby achieving the target detection task.

[0037] In this embodiment, the anti-reflection attention fusion module generates three-dimensional attention weights through an energy function.

[0038] Specifically, the anti-reflection attention fusion module is based on the SimAM attention mechanism, which focuses on enhancing the edge features of transparent areas, effectively alleviating the detection difficulties caused by the similarity between the transparent tape and the background color and strong reflective interference.

[0039] The SimAM attention mechanism is based on neuroscience findings. It generates three-dimensional attention weights through an energy function to assess the importance of neurons and highlight the role of key neurons. This gives higher priority to information-rich neurons in visual processing, which is different from traditional one-dimensional or two-dimensional channel or spatial attention mechanisms.

[0040] The energy function is calculated using the following formula: In the formula and Let represent the mean and variance of all elements in the input feature, respectively. For the target neuron, These are hyperparameters. As can be understood, they represent energy values. The lower the level, the more neurons The more distinct a neuron is from its surrounding neurons, the greater its importance in visual processing. Therefore, the three-dimensional weights of each neuron can be determined through... express.

[0041] Furthermore, the procedure for evaluating neuron importance using three-dimensional weights is as follows: In the anti-reflection attention fusion module, multi-level feature maps are concatenated and integrated through channels before being input into the SimAM attention mechanism to generate a three-dimensional attention weight matrix. This weight is applied to the original feature maps through element-wise multiplication, strengthening key features and suppressing noise interference. The weighted features are then added to the original features from another branch to enhance the feature representation, making the features more discriminative and representative. Subsequently, the enhanced features are concatenated along the channel dimension and fed into the next layer of the network for further processing.

[0042] Third, set the bounding box loss function.

[0043] In this embodiment, the bounding box loss function is set to satisfy the following formula: in, Represents the bounding box loss function. Indicates the weighting coefficient. Represents the SmoothL1 loss function. This represents the FocalEIoU loss function.

[0044] The bounding box loss function based on the weighted combination of SmoothL1 and FocalEIoU not only solves the problem of the traditional L1 loss function being non-differentiable at the zero point, but also overcomes the deficiency of the GIoU loss function in distinguishing the relative position of the target, thereby improving the localization accuracy of the model.

[0045] Specifically, the formulas for the SmoothL1 and FocalEIoU loss functions are as follows: in , and These represent the true value, the predicted value, and the sample size, respectively. , and These represent overlap loss, center distance loss, and aspect ratio loss, respectively. and , and , and These represent the center coordinates, width, and height of the predicted bounding box and the ground truth bounding box, respectively. Represents the square of the Euclidean distance. and It is a constant used for normalization. Parameters used to control the degree of outlier suppression.

[0046] This invention uses a first feature extraction network as the backbone network, which improves feature extraction capabilities while effectively reducing the number of model parameters. In the feature fusion stage, an anti-reflection attention fusion module is introduced to enhance the multi-scale feature fusion effect, thereby solving the recognition difficulties caused by the similarity between transparent tape and background texture and high reflectivity. The DGST module is used to replace the original RepC3 module to avoid the high computational cost and real-time bottleneck caused by the multi-branch structure. At the same time, the bounding box regression loss function is replaced with a weighted combination of SmoothL1 and FocalEIoU to further optimize the model convergence process and significantly improve detection accuracy.

[0047] Furthermore, the input image size was set to 640×640, the number of training rounds was 150, the batch size was 16, and the initial learning rate was 0.0001; the optimizer was AdamW, the momentum parameter was 0.9, and the weight decay coefficient was 0.0001. The detection model was trained using the preprocessed dataset. S03. Using the trained detection model, complete the detection of transparent tape targets on the fabric surface.

[0048] After processing the fabric image to be detected according to the standardization method described above, input it into the trained detection model. The model outputs the predicted bounding box and confidence score of the transparent tape. Set the confidence score threshold (default 0.5) and filter out the prediction results with confidence scores ≥ the threshold, which are the final detection results.

[0049] To verify the improved detection accuracy of the transparent adhesive tape detection model on fabric surfaces proposed in this embodiment, the mean accuracy (mAP, mAP@0.5), number of model parameters (params), floating-point operations (FLOPs), and detection speed (FPS) were used as key indicators to evaluate the detection results. Table 1 shows the model performance comparison. Under consistent experimental conditions and parameters, the experimental results show that compared with the RT-DETR model, the number of parameters was reduced by 43.8%, the floating-point operations were reduced by 39.6%, mAP@0.5 was improved by 2.2%, and the detection speed was improved by 13.9%. While reducing the number of model parameters and the floating-point operations, the detection accuracy and speed were significantly improved, fully demonstrating the effectiveness of the introduced module and its ability to better meet the needs of enterprises for real-time detection of transparent adhesive tape on fabric surfaces.

[0050] RT-DETR 20.1 58.6 107.2 96.3 This invention 11.3 35.4 122.1 98.4 Table 1 Performance Comparison Table Please see Figure 3 In an embodiment, to efficiently execute the method for detecting transparent adhesive tape on a fabric surface provided by this invention, this invention also provides a system for detecting transparent adhesive tape on a fabric surface, comprising: an input device, an output device, a processor, and a memory, wherein the input device, output device, processor, and memory are interconnected, and the memory contains program instructions for the steps of the method for detecting transparent adhesive tape on a fabric surface. The system for detecting transparent adhesive tape on a fabric surface provided by this invention has a compact structure and stable performance, and can stably execute the method for detecting transparent adhesive tape on a fabric surface provided by this invention, further enhancing the overall applicability and practical application capability of this invention.

[0051] In this embodiment, the processor may be a central processing unit, but it can also be other general-purpose processors, digital signal processors, application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (OPGs), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor. Input devices can be used to acquire data. Output devices can be used to output the results obtained by storing program instructions contained in a computer program in the memory provided by this invention. The memory may include read-only memory and random access memory (RAM), and provides instructions and data to the processor. A portion of the memory may also include non-volatile random access memory (RAM).

[0052] In one possible implementation, the memory may include a stored program area and a stored data area. The stored program area may store the operating system and applications required for at least one function; the stored data area may store data created during use. Furthermore, the memory may include read-only memory and random access memory, and provides instructions and data to the processor. The memory stores the operating system and operating instructions, executable modules, or data structures, or subsets thereof, or extended sets thereof. The operating instructions may include various operation instructions for implementing various operations. The operating system may include various system programs for implementing various basic tasks and handling hardware-based tasks.

[0053] The embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method for detecting transparent adhesive tape targets on fabric surfaces.

[0054] The storage medium can include various media that can store program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0055] In summary, this invention first acquires images of transparent tape made of fabric under various materials, lighting conditions, and application methods, constructs a dataset, and preprocesses it (including standardization and geometric / lighting / noise enhancement). Then, it designs a first feature extraction network with specific sub-modules, optimizes the encoder (adding an anti-reflection attention module and replacing RepC3 with DGST), sets a combined loss function to construct a detection model, and trains it. Finally, it uses the trained model for detection, effectively solving the problems of difficult feature extraction and poor generalization caused by the reflectivity of transparent tape.

[0056] 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. Such 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, and they should all be covered within the scope of the present invention.

Claims

1. A method for detecting transparent adhesive tape targets on the surface of fabric, characterized in that, Includes the following steps: Obtain a dataset of images of transparent tape on the surface of fabric, and preprocess the dataset; A detection model for transparent adhesive tape on fabric surfaces was constructed, and the detection model was trained using a preprocessed dataset. The trained detection model is used to detect transparent tape targets on the fabric surface. The preprocessing of the dataset includes data augmentation operations, which include geometric transformation, lighting adjustment, and noise addition. The construction of the transparent adhesive tape detection model on the fabric surface includes the following steps: Design the first feature extraction network; An anti-reflection attention fusion module is added to the cross-scale feature fusion structure of the efficient hybrid encoder; Set the bounding box loss function; The anti-reflection attention fusion module generates three-dimensional attention weights through an energy function; The energy function The calculation formula is as follows: ， In the formula, and Let represent the mean and variance of all elements in the input feature, respectively. For the target neuron, For hyperparameters; The first feature extraction network first completes preliminary downsampling feature extraction through a first sub-module; Then, the feature extraction stage proceeds to four levels, each of which consists of a downsampled first sub-module and a second sub-module in sequence. The second submodule is used to perform preliminary convolution and normalization processing on the input features through the first submodule; Subsequently, the feature map is divided into two paths: one serves as a shortcut to preserve local details, and the other is input into a deep feature extraction channel consisting of N ELGCA bottleneck layers. Finally, the deep features processed by the ELGCA network are concatenated with the shortcut connection branches in the channel dimension, and feature fusion is performed through the first submodule at the end. The cross-scale feature fusion structure of the high-efficiency hybrid encoder also includes replacing the RepC3 module in the cross-scale feature fusion structure with the DGST module; The DGST module achieves parameter compression and cross-group feature interaction through a dynamic group selection module, and combines a convolutional feedforward network to dynamically model the relationship between local and global contexts. The dynamic grouping selection module first compresses the number of channels in the input feature map to half of the original number using a 1×1 convolution; The compressed features are then fed into two parallel branches simultaneously: one branch directly serves as an identity mapping, preserving the integrity of the input information; the other branch sequentially performs lightweight spatial feature extraction through a 3×3 grouped convolution, followed by a channel rearrangement operation. Finally, the outputs of the two branches are merged by channel concatenation; The bounding box loss function is set to satisfy the following formula: in, Represents the bounding box loss function. Indicates the weighting coefficient. Represents the SmoothL1 loss function. This represents the FocalEIoU loss function.

2. The method for detecting transparent adhesive tape targets on fabric surfaces according to claim 1, characterized in that, The first submodule consists of a 3×3 convolutional layer, a batch normalization layer, and a Silu activation function.

3. A target detection system for transparent adhesive tape on fabric surface, characterized in that, The transparent adhesive tape target detection system for fabric surface includes: an input device, an output device, a processor, and a memory, wherein the input device, output device, processor, and memory are interconnected, and the memory includes program instructions for executing the transparent adhesive tape target detection method for fabric surface according to any one of claims 1-2.

Images