Non-woven fabric LOGO defect detection method based on deep learning

Abstract

The invention belongs to the technical field of non-woven fabric defect detection, and particularly discloses a non-woven fabric LOGO defect detection method based on deep learning, which comprises the following steps: acquiring an image by using a camera CCD (Charge Coupled Device), carrying out gray stretching, logarithmic transformation and graying processing on the image, enhancing the boundary of a LOGO image, and detecting the LOGO defect by inputting a standard sample page and a gray image of a to-be-detected LOGO image. The method comprises the following steps: acquiring most significant features, performing geometric transformation by using a feature matching algorithm, realizing two-dimensional registration, performing preprocessing and data labeling on a registered LOGO image, dividing the LOGO image into a training data set and a test data set according to a preset proportion, constructing a Faster RCNN network model based on ResNet50 and FPN networks, and setting training parameters and a loss function. And inputting the training data set and the test data set into the Faster RCNN network model for training and testing to obtain a tested Faster RCNN network model for detecting the LOGO defects of the non-woven fabric, and the method realizes accurate detection of the LOGO defects of the non-woven fabric by using a deep learning technology through image processing and feature matching.

Description

Technical Field

[0001] The present invention relates to the technical field of non-woven fabric defect detection, and particularly to a method for detecting non-woven fabric LOGO defects based on deep learning. Background Art

[0002] As an important industrial material, non-woven fabrics have a wide range of applications in many fields such as medical and health, clothing, construction, packaging, and automobiles. However, during the production process of non-woven fabrics, due to the influence of environmental and human factors, various defects are likely to occur, such as dirt spots, wrinkles, missing yarns, breaks, and incomplete or misaligned LOGO prints. These defects not only affect the quality of non-woven fabrics but also reduce the production efficiency and economic benefits of enterprises.

[0003] Traditional non-woven fabric defect detection methods mainly rely on manual visual inspection or automated detection methods based on traditional image processing. However, manual visual inspection has problems such as low efficiency and high false detection rate, while traditional image processing methods are limited by the complexity and generalization ability of algorithms and are difficult to meet the requirements of large-scale and high-efficiency defect detection.

[0004] In recent years, with the rapid development of deep learning technology, remarkable achievements have been made in the fields of image recognition, classification, detection, etc. Deep learning technology can automatically extract high-level features in images and achieve accurate recognition and classification of complex targets by constructing deep neural network models. Therefore, applying deep learning technology to non-woven fabric LOGO defect detection is expected to solve the problems existing in traditional detection methods and improve the accuracy and efficiency of detection. Summary of the Invention

[0005] In view of the above-mentioned disadvantages of the prior art, the purpose of the present invention is to provide a method for detecting non-woven fabric LOGO defects based on deep learning, which is used to solve the problems of low detection efficiency, high false detection rate, and difficulty in adapting to large-scale and diverse defect detections in the prior art.

[0006] To achieve the above purpose and other related purposes, the present invention provides a method for detecting non-woven fabric LOGO defects based on deep learning, including: Using a CCD camera to collect images of non-woven fabric LOGO, stretching the collected images to a predetermined range in grayscale, performing logarithmic transformation on each pixel value of the RGB color channels, converting the transformed images into grayscale images, and enhancing the boundaries in the LOGO images by detecting the differences in adjacent pixel brightness values; Inputting the grayscale images of the standard samples and the grayscale images of the LOGO to be detected, obtaining the most significant features of the two, and performing geometric transformation on the LOGO images to be detected through a feature matching algorithm to register the standard samples and the LOGO images to be detected in two dimensions; Preprocess the registered LOGO image to obtain LOGO image data; perform annotation processing on the LOGO image data, and divide the annotated LOGO image data into a training data set and a test data set according to a preset ratio; Based on the ResNet50 and FPN networks, construct a Faster RCNN network model, and set the training parameters of the Faster RCNN network model and construct the loss function of the Faster RCNN network model; Input the training data set and the test data set into the Faster RCNN network model in sequence for training and testing, and save the network model parameters after testing to obtain the Faster RCNN network model after testing, which is used to detect non-woven fabric LOGO defects.

[0007] Optionally, perform logarithmic transformation on each pixel value of the RGB color channel, and the transformation formula is as follows: In the formula, represents the value of each pixel point of each color channel.

[0008] Optionally, geometric transformation is performed on the LOGO image to be detected through a feature matching algorithm, so that the standard sample and the LOGO image to be detected are registered in two dimensions, specifically including: By converting the coordinate values of the LOGO image from the RGB color space to the grayscale color space, if the value of each pixel of the grayscale image of the standard sample is , then the value of each pixel of the grayscale image of the LOGO image to be detected is , and the feature matching algorithm is as follows: In the formula, is a function containing multiple geometric transformation parameters.

[0009] Optionally, the preprocessing of the registered LOGO image includes: Select the maximum width and height of the pictures in each mini-batch to generate a new blank tensor of the corresponding size, and then map each picture in the mini-batch to this blank tensor with its upper left corner as the (0,0) coordinate through the broadcast mechanism, while retaining the original information in the picture and not changing the aspect ratio of the picture; Perform downsampling on the original LOGO image to reduce its resolution.

[0010] Optionally, the annotation processing of the LOGO image data includes: Use the LabelImg annotation tool to annotate the defect positions and defect types of the non-woven fabric images. The non-woven fabric LOGO defects include point defects, line defects, surface defects, scratching defects, and breakage defects, and obtain image data containing the defect positions and types of the LOGO defect images.

[0011] Optionally, based on the ResNet50 and FPN networks, construct a Faster RCNN network model, including: Select the residual network ResNet50 as the backbone network of the Faster RCNN object detection model to extract the features of the non-woven fabric LOGO images; Represent the feature activation outputs of the last residual block in each stage of ResNet50 as C2, C3, C4, C5. Input the C2, C3, C4, C5 into the FPN part. After adjusting the number of channels through a 1×1 convolution, through upsampling and feature fusion, and then through a 3×3 convolution to obtain new feature maps P2, P3, P4, P5; in addition, P5 passes through a 1×1 max pooling layer to obtain the feature map P6, and the convolution kernel size of the max pooling layer is 1×1 max pooling; After the C5 layer of ResNet50, introduce an involution convolution to replace the original 3×3 convolution kernel.

[0012] Optionally, the involution convolution is defined as: In the formula, represents the local area of the input feature map at position , represents applying an involution convolution operation to , is a non-linear activation function; and represent two linear transformations, where: , used to reduce the number of input channels to , where is the reduction ratio; , used to map the reduced feature to the output space, where is the spatial size of the involution kernel, represents how many channels share one involution kernel.

[0013] Optionally, the training parameters at least include: the number of input channels, the number of output channels of the fully connected layer, the size of the ROI feature layer, the Classification parameter, the model learning rate, the number of extraction boxes, the number of data samples grabbed in one training, and the total number of iterations.

[0014] Optionally, the loss function of the Faster RCNN network model is optimized using the RPN network, and the construction method includes: For the anchor boxes generated by the RPN network, calculate its classification loss. The anchor boxes are divided into two categories: foreground and background, and the logarithmic loss function is used for calculation. The logarithmic loss function is as follows: In the formula, is the probability that the anchor box predicted by the model is foreground, is the true label of the anchor box; is the index; For the anchor boxes classified as foreground, calculate their bounding box regression loss, and use function as the regression loss function. The regression loss function is as follows: In the formula, is the parameterized coordinates of the bounding box predicted by the model, and represent the coordinates of the center of the bounding box, represents the width of the bounding box, represents the height of the bounding box, is the parameterized coordinates of the true bounding box, is the index; is function, and its formula is as follows: In the formula, is the difference between the parameters of the bounding box predicted by the model and the parameters of the true bounding box, is a hyperparameter used to control the smoothness of the function; Combine the classification loss and the regression loss to form the total loss function. The formula is as follows: In the formula, is the classification loss part, is the normalization factor of the classification loss, is the regression loss part, is the normalization factor of the regression loss, is the weight coefficient for balancing the classification loss and the regression loss.

[0015] Optionally, the training method of the Faster-RCNN network model is as follows: Use the training dataset and, according to the set training parameters, perform iterative training on the model. In each iteration, calculate the total loss function and update the model parameters according to the gradient descent method; During the training process, use the test dataset to evaluate the performance of the model and adjust the training parameters according to the evaluation results.

[0016] As described above, a non-woven fabric LOGO defect detection method based on deep learning proposed by the present invention has the following beneficial effects: The present invention realizes high-precision recognition of five main non-woven fabric LOGO defect types, namely point defects, line defects, surface defects, scratching defects, and breakage defects. Through the training of the deep learning model, this method can automatically extract and learn the features of these defects, thereby realizing accurate classification and positioning of the defects. This not only improves the accuracy of detection but also reduces the misdetection and missed detection in manual visual inspection.

[0017] The improved Faster RCNN network model of the present invention combines the advantages of the Region Proposal Network (RPN) and the Convolutional Neural Network (CNN), and can efficiently and accurately detect defects in non-woven fabric LOGOs; candidate regions are generated through the RPN network, greatly reducing the redundant calculation of sliding windows in traditional object detection methods and improving the detection speed. Secondly, the FasterRCNN network model has a strong feature extraction ability. Using ResNet50 as the backbone network can extract rich image features, which helps to identify various complex defect types. In addition, the addition of the FPN network significantly improves the effect of non-woven fabric LOGO defect detection. Using the feature pyramid structure, it realizes the fusion and utilization of multi-scale features, enhances the sensitivity and detection ability of the model to defects of different scales, and thus improves the accuracy and robustness of detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It shows a schematic flow chart of a non-woven fabric LOGO defect detection method based on deep learning according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0019] The following specific examples illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation manners, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention.

[0020] It should be noted that the illustrations provided in this embodiment only schematically illustrate the basic concept of the present invention. Therefore, only the components related to the present invention are shown in the drawings, rather than being drawn according to the number, shape, and size of the components in actual implementation. The types, quantities, and proportions of the components in actual implementation can be arbitrarily changed, and the component layout type may also be more complex. The structures, proportions, sizes, etc. shown in the drawings of this specification are only used to cooperate with the content disclosed in the specification for those familiar with this technology to understand and read, and are not used to limit the conditions under which the present invention can be implemented. Therefore, they do not have a substantial technical meaning. Any modification of the structure, change in the proportional relationship, or adjustment of the size, without affecting the effects that the present invention can produce and the purposes that can be achieved, should still fall within the scope that can be covered by the technical content disclosed in the present invention. At the same time, the terms such as "upper", "lower", "left", "right", "middle", and "one" cited in this specification are only for the convenience of clear narration and are not used to limit the scope of implementation of the present invention. The change or adjustment of their relative relationships, without substantial change in the technical content, should also be regarded as the scope of implementation of the present invention.

[0021] The present invention provides a non-woven fabric LOGO defect detection method based on deep learning. The technical solution of the present invention will be described in detail below in combination with specific embodiments and drawings.

[0022] Refer to Figure 1 As shown, a non-woven fabric LOGO defect detection method based on deep learning of the present invention includes the following steps: S10: Use a CCD camera to collect images of the non-woven fabric LOGO, stretch the collected images to a predetermined range in grayscale, perform logarithmic transformation on each pixel value of the RGB color channels, convert the transformed images into grayscale images, and enhance the boundaries in the LOGO images by detecting the differences in the brightness values of adjacent pixels; S20: Input the grayscale images of the standard sample and the LOGO grayscale image to be detected, obtain the most significant features of the two, and perform geometric transformation on the LOGO image to be detected through a feature matching algorithm so that the standard sample and the LOGO image to be detected are registered in two dimensions; S30: Preprocess the registered LOGO images to obtain LOGO image data; perform annotation processing on the LOGO image data, and divide the annotated LOGO image data into a training data set and a test data set according to a preset ratio; S40: Based on the ResNet50 and FPN networks, construct a Faster RCNN network model, and set the training parameters of the Faster RCNN network model and construct the loss function of the Faster RCNN network model; S50; Input the training dataset and the test dataset into the Faster RCNN network model in sequence for training and testing, and save the network model parameters after testing to obtain the Faster RCNN network model after testing. This Faster RCNN network model after testing is used to detect non-woven fabric LOGO defects.

[0023] In summary, the non-woven fabric LOGO defect detection method based on deep learning has the advantages of high efficiency, accuracy, automation, etc., and is expected to become one of the mainstream technologies in the future non-woven fabric quality detection field.

[0024] In the embodiments of the present invention, a CCD camera is used to collect images of the LOGO on the non-woven fabric. In order to enhance the image contrast, the collected images are stretched to a predetermined range in grayscale, making the bright and dark parts in the images more distinct, and a logarithmic transformation is performed on each pixel value of the RGB color channels. The transformation formula is as follows: In the formula, represents the value of each pixel point in each color channel.

[0025] This logarithmic transformation can compress the pixel values in the high-brightness area and expand the pixel values in the low-brightness area, which helps to improve the visual effect of the image and makes the subsequent processing more stable. After the logarithmic transformation, the image is converted into a grayscale image, and the boundary in the LOGO image is enhanced by detecting the difference in adjacent pixel brightness values, laying a foundation for subsequent feature extraction and defect detection.

[0026] In order to register the standard sample and the LOGO image to be detected in two dimensions, first, both need to be converted into grayscale images to obtain their most significant features. Then, a geometric transformation is performed on the LOGO image to be detected through a feature matching algorithm, so that the standard sample and the LOGO image to be detected are registered in two dimensions.

[0027] Specifically, by converting the coordinate values of the LOGO image from the RGB color space to the grayscale color space, if the value of each pixel in the grayscale image of the standard sample is then the value of each pixel in the grayscale image of the LOGO image to be detected is The feature matching algorithm is as follows: In the formula, is a function containing multiple geometric transformation parameters, such as translation, rotation, scaling, etc. By solving this function, the LOGO image to be detected can be accurately registered with the standard sample in two dimensions, providing an accurate alignment basis for subsequent defect detection.

[0028] Specifically, the most significant features of the grayscale image of the standard sample and the grayscale image of the logo to be detected can be obtained through SIFT, SURF or ORB feature extraction algorithms; and the most significant features are registered through a function. More specifically, if the most significant features recognized are two circular contours, the standard sample and the logo image to be detected are registered through the line significant feature formed by the two centers of the circles. The specific method is to identify the position information of the line significant feature of the standard sample, that is, the coordinate values of the two centers of the circles, and then calculate the slope and length of the line. The position information of the line significant feature of the logo image to be detected is identified in the same way as the slope and length of the line, and the translation of the logo image to be detected is realized through the translation transformation formula.

[0029] The registered logo image is preprocessed, including selecting the maximum width and height of the images in each mini-batch, generating a new blank tensor of the corresponding size, and then mapping each image in the mini-batch to the blank tensor with its upper left corner as the (0,0) coordinate through the broadcasting mechanism, while retaining the original information in the image and not changing the aspect ratio of the image. In addition, the original logo image is downsampled to reduce its resolution and calculation amount.

[0030] Next, the preprocessed logo image data is labeled. The LabelImg labeling tool is used to label the defect positions and defect types of the non-woven fabric images. The defect types include point defects, line defects, surface defects, scratching defects and breakage defects. After labeling, image data including the defect positions and types of the logo defect images is obtained. The labeled logo image data is divided into a training data set and a test data set according to a preset ratio for subsequent network model training and testing.

[0031] Based on the ResNet50 and FPN networks, a Faster RCNN network model is constructed. The residual network ResNet50 is selected as the backbone network of the Faster RCNN object detection model to extract the features of the non-woven fabric logo image. ResNet50 has strong feature extraction capabilities and can effectively handle the logo defect detection task under complex backgrounds.

[0032] Specifically, the feature activation outputs of the last residual block in each stage of ResNet50 are denoted as C2, C3, C4, and C5. The feature maps of C2, C3, C4, and C5 are input into the FPN part. After adjusting the number of channels through a 1×1 convolution, FPN fuses the feature maps of different scales through upsampling and feature fusion, and then obtains new feature maps P2, P3, P4, and P5 through a 3×3 convolution. In addition, P5 passes through a 1×1 max pooling layer to obtain the feature map P6. These feature maps have different scales and resolutions and can capture LOGO defects of different scales.

[0033] After the C5 layer of ResNet50, dilated convolutions are introduced to replace the original 3×3 convolution kernels. Dilated convolutions can reduce the number of model parameters and computational volume while maintaining the performance of the model.

[0034] The dilated convolution is defined as: In the formula, represents the local region of the input feature map at position , represents applying the dilated convolution operation to , is a non-linear activation function; and represent two linear transformations, where: , which is used to reduce the number of input channels to , where is the reduction ratio; , which is used to map the reduced feature to the output space, where is the spatial size of the dilated convolution kernel, represents how many channels share one dilated convolution kernel. The use of dilated convolutions can make the model more efficient and compact.

[0035] The loss function of the Faster RCNN network model is optimized using the RPN network. The construction process of the loss function of the Faster RCNN network model is as follows: For the anchor boxes generated by the RPN network, calculate their classification loss and bounding box regression loss. The classification loss is calculated using the logarithmic loss function, and the formula is as follows: In the formula, is the probability that the anchor box predicted by the model is a foreground, is the true label of the anchor box (foreground is 1, background is 0); ​as the index. The logarithmic loss function can measure the difference between the probability distribution predicted by the model and the true label, and guide the model to optimize in the correct direction.

[0036] For the anchor boxes classified as foreground, calculate their bounding box regression loss, using function as the regression loss function. The regression loss function is as follows: In the formula, are the parameterized coordinates of the bounding box predicted by the model, and represent the coordinates of the center of the bounding box, represents the width of the bounding box, represents the height of the bounding box, are the parameterized coordinates of the true bounding box, is the index; is function, The function can balance the influence of large errors and small errors on the loss, making the model more stable during training. Its formula is as follows: In the formula, is the difference between the parameters of the bounding box predicted by the model and the parameters of the true bounding box, is a hyperparameter used to control the smoothness of the function; by adjusting the value, the model can pay more attention to the optimization of small errors or large errors during training.

[0037] Combine the classification loss and the regression loss to form the total loss function. The formula is as follows: In the formula, is the classification loss part, is the normalization factor of the classification loss (such as the number of anchor points selected by the RPN network), is the regression loss part, is the normalization factor of the regression loss (such as the number of foreground anchor points), is the weight coefficient for balancing the classification loss and the regression loss. By optimizing the total loss function, the model can achieve good performance in both classification and regression tasks.

[0038] Set the training parameters of the Faster RCNN network model. Use the training dataset to iteratively train the Faster RCNN network model according to the set training parameters. In each iteration, calculate the total loss function and update the model parameters according to the gradient descent method. The training parameters at least include the number of input channels, the number of output channels of the fully connected layer, the size of the ROI feature layer, the Classification parameter, the model learning rate, the number of extraction boxes, the number of data samples grabbed in one training, and the total number of iterations.

[0039] Specifically, in this embodiment: Number of input channels: For RGB images, the number of input channels is 3; Number of output channels of the fully connected layer: In non-woven fabric LOGO defect detection, multiple defect types (such as point defects, line defects, etc.) need to be identified. Therefore, the number of output channels should match the number of defect types, and the fully connected layer should have 5 output nodes; Size of the ROI feature layer: Set to 7 non-woven fabric defect detection 7; Classification parameter: Set according to the number of defect types. There are 5 defect types, so the classification parameter is 5; Model learning rate: The initial learning rate can be set to 0.001 or 0.0001 and adjusted according to the training process; Number of extraction boxes: Set to 3000; Number of data samples grabbed in one training: Set to 16; Total number of iterations: Set to 10000 times.

[0040] Input the training dataset and the test dataset into the Faster RCNN network model in sequence for training and testing. During the training process, use the test dataset to evaluate the performance of the model. According to the evaluation results (such as accuracy, recall rate, etc.), adjust the training parameters to optimize the model performance. Finally, obtain the tested Faster RCNN network model, save the parameters of the tested network model, and obtain the tested Faster RCNN network model. This tested Faster RCNN network model can effectively detect non-woven fabric LOGO defects.

[0041] Through the above specific implementation manner, the present invention provides a method for detecting non-woven fabric LOGO defects based on deep learning. This method can effectively extract the features of non-woven fabric LOGO images and accurately detect the defects on the LOGO. At the same time, by optimizing the network structure and the loss function, the detection accuracy and efficiency of the model are improved.

[0042] The above embodiments are only illustrative of the principles and effects of the present invention and are not intended to limit the present invention. Any person familiar with this technology can modify or change the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or changes made by those with ordinary knowledge in the technical field without departing from the spirit and technical idea disclosed by the present invention should still be covered by the claims of the present invention.

Claims

1. A non-woven fabric LOGO defect detection method based on deep learning, characterized in that: include: Use a CCD camera to capture images of the non-woven fabric LOGO, stretch the captured image to a predetermined range, perform logarithmic transformation on each pixel value of the RGB color channel, convert the transformed image into a grayscale image, and enhance the boundaries in the LOGO image by detecting the difference in brightness values ​​of adjacent pixels; Input the grayscale image of the standard sample and the grayscale image of the LOGO to be detected, obtain the most significant features of the two, and use the feature matching algorithm to perform geometric transformation on the LOGO image to be detected, so that the standard sample and the LOGO image to be detected are registered in two dimensions; Preprocess the registered LOGO image to obtain LOGO image data; Annotate the LOGO image data, and divide the annotated LOGO image data into a training data set and a test data set according to a preset ratio; Based on ResNet50 and FPN networks, a Faster RCNN network model is constructed, and the training parameters of the Faster RCNN network model are set and the loss function of the Faster RCNN network model is constructed; The training data set and the test data set are sequentially input into the Faster RCNN network model for training and testing, and the network model parameters after the test are saved to obtain the Faster RCNN network model after the test, and the Faster RCNN network model after the test is used to detect non-woven fabric LOGO defects.

2. According to a non-woven fabric LOGO defect detection method based on deep learning according to claim 1, it is characterized in that: The logarithmic transformation is performed on each pixel value of the RGB color channel, and the transformation formula is as follows: In the formula, Represents the value of each color channel of each pixel.

3. The non-woven fabric LOGO defect detection method based on deep learning according to claim 2, characterized in that: The geometric transformation of the LOGO image to be detected is performed by using a feature matching algorithm so that the standard sample and the LOGO image to be detected are registered in two dimensions, specifically including: By converting the coordinate value of the LOGO image from the RGB color space to the grayscale color space, if the value of each pixel of the grayscale image of the standard sample is , then the value of each pixel in the grayscale image of the LOGO image to be detected is , the feature matching algorithm is as follows: In the formula, is a function containing multiple geometric transformation parameters.

4. The non-woven fabric LOGO defect detection method based on deep learning according to claim 3 is characterized in that: The preprocessing of the registered LOGO image comprises: Select the maximum width and height of the images in each mini-batch, generate a new blank tensor of the corresponding size, and then map each image in the mini-batch to the blank tensor through a broadcast mechanism, with its upper left corner as the (0,0) coordinate, while retaining the original information in the image and not changing the aspect ratio of the image; The original LOGO image is downsampled to reduce its resolution.

5. The non-woven fabric LOGO defect detection method based on deep learning according to claim 4 is characterized in that: The labeling process of the LOGO image data includes: The LabelImg annotation tool is used to annotate the defect positions and defect types of the non-woven fabric image. The non-woven fabric LOGO defects include point defects, line defects, surface defects, scratch defects and damage defects, and image data containing the defect positions and types of the LOGO defect image is obtained.

6. The non-woven fabric LOGO defect detection method based on deep learning according to claim 1, characterized in that: The Faster RCNN network model is constructed based on ResNet50 and FPN network, including: The residual network ResNet50 is selected as the backbone network of the Faster RCNN target detection model to extract the features of the non-woven fabric LOGO image; The feature activation output of the last residual block of each stage of ResNet50 is represented as C2, C3, C4, C5, and the C2, C3, C4, C5 are input into the FPN part. After adjusting the number of channels through 1×1 convolution, new feature maps P2, P3, P4, P5 are obtained through upsampling and feature fusion, and then through 3×3 convolution. In addition, P5 is passed through a 1×1 maximum pooling layer to obtain feature map P6, and the convolution kernel size of the maximum pooling layer is 1×1 maximum pooling; After the C5 layer of ResNet50, an inward convolution is introduced to replace the original 3×3 convolution kernel.

7. The method for detecting non-woven fabric LOGO defects based on deep learning according to claim 6, characterized in that: The inner convolution is defined as: In the formula, Indicates that the input feature map is at position local area, Express Apply the inner convolution operation, is a nonlinear activation function; and Represents two linear transformations, where: , used to convert the number of input channels Dimensionality reduction to ,in It is the price reduction ratio; , which is used to map the reduced-dimensional features to the output space, where is the spatial size of the involution core, Indicates how many channels share one convolution core.

8. The non-woven fabric LOGO defect detection method based on deep learning according to claim 1, characterized in that: The training parameters include at least: the number of input channels, the number of fully connected layer output channels, the size of the ROI feature layer, the Classification parameters, the model learning rate, the number of extraction boxes, the number of data samples captured in one training and the total number of iterations.

9. The non-woven fabric LOGO defect detection method based on deep learning according to claim 8, characterized in that: The loss function of the Faster RCNN network model is optimized using the RPN network, and the construction method includes: For the anchor box generated by the RPN network, its classification loss is calculated. The anchor box is divided into two categories: foreground and background. The logarithmic loss function is used for calculation. The logarithmic loss function is as follows: In the formula, is the probability that the anchor box predicted by the model is the foreground, is the true label of the anchor box; is the index; For the anchor box classified as foreground, calculate its bounding box regression loss using The function is used as the regression loss function, and the regression loss function is as follows: In the formula, are the parameterized coordinates of the bounding box predicted by the model, and represents the coordinates of the center of the bounding box, represents the width of the bounding box, represents the height of the bounding box, are the parameterized coordinates of the true bounding box, is the index; for Function, its formula is as follows: In the formula, is the difference between the bounding box parameters predicted by the model and the true bounding box parameters, is a hyperparameter used to control The smoothness of the function; Combining the classification loss and regression loss to form the total loss function, the formula is as follows: In the formula, is the classification loss part, is the normalization factor of the classification loss, is the regression loss part, is the normalization factor of the regression loss, is the weight coefficient for balancing classification loss and regression loss.

10. The method for detecting non-woven fabric LOGO defects based on deep learning according to claim 9, characterized in that: The Faster-RCNN network model training method is as follows: Using the training data set, the model is iteratively trained according to the set training parameters. In each iteration, the total loss function is calculated, and the model parameters are updated according to the gradient descent method. During the training process, the test data set is used to evaluate the performance of the model, and the training parameters are adjusted based on the evaluation results.

Images