A nonwoven fabric defect detection and classification method based on deep learning

Abstract

The application discloses a non-woven fabric defect detection and classification method based on deep learning, which comprises the following steps: collecting non-woven fabric continuous surface images to generate homologous texture frame sequences; constructing fiber micro-element topological connection state sequences; identifying irreversible connection changes of fiber micro-elements, and splitting target texture event sets; performing pulse coding on the target texture event sets to generate texture pulse sequences; constructing a topological causal liquid state machine model according to texture event markers and direction guide markers; inputting the texture pulse sequences into the model to generate liquid state paths; and performing trajectory merging and category mapping on the liquid state paths to generate non-woven fabric defect detection and classification results. The application improves the detection stability of low-contrast defects and boundary fuzzy defects.

Description

Technical Field

[0001] This invention relates to the field of textile material quality testing technology, and in particular to a method for detecting and classifying defects in nonwoven fabrics based on deep learning. Background Technology

[0002] In the field of textile material quality inspection technology, existing inspection methods for nonwoven fabrics, which are prone to surface defects such as fiber breakage, wrinkles, fiber agglomeration, uneven thickness, stains, and abnormal pores during production, mostly rely on manual visual inspection, ordinary machine vision inspection, or convolutional neural network classification inspection. These methods acquire images of the nonwoven fabric surface and extract grayscale, texture, edge, or depth features to complete defect identification. While these methods can detect obvious damage and high-contrast defects, they lack stability in identifying small defects, low-contrast defects, and defects with blurred boundaries in continuous production scenarios.

[0003] Existing deep learning detection methods typically treat nonwoven fabric images as static image patches, lacking modeling of state changes in the same texture region during continuous acquisition, making it difficult to distinguish between normal fiber texture fluctuations and real defect disturbances. Nonwoven fabric surface fibers are characterized by interlacing, adhesion, and uneven pore distribution; ordinary image features easily misclassify normal textures as defects and tend to overlook the continuous changes in the connection state of fiber micro-elements. While existing liquid state machine models can handle temporal inputs, conventional reservoir structures do not incorporate fiber micro-element topological connections, direction guidance markers, and causal relationships within the reservoir area. This makes it difficult to effectively separate defect disturbance paths from normal texture state paths, and defect classification results are easily affected by differences in defect morphology and scale.

[0004] Therefore, how to provide a deep learning-based method for nonwoven fabric defect detection and classification is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for detecting and classifying defects in nonwoven fabrics based on deep learning. This invention utilizes isotopic texture frame sequences, fiber micro-element topological connection states, and a topological causal liquid state machine model to achieve the localization and classification of defects in nonwoven fabrics. It has the advantages of stable weak defect identification, high classification accuracy, and high online detection efficiency.

[0006] A method for detecting and classifying defects in nonwoven fabrics based on deep learning according to an embodiment of the present invention includes the following steps:

[0007] Acquire images of the continuous surface of the nonwoven fabric, perform image correction processing, and generate a sequence of isomorphic texture frames.

[0008] Based on the co-position texture frame sequence, construct the fiber element connection state and generate the fiber element topological connection state sequence;

[0009] Irreversible connection changes of fiber micro-elements are identified based on the fiber micro-element topological connection state sequence, defect-type event start markers are generated, and target texture event sets are generated according to the defect-type event start markers.

[0010] The target texture event set is pulse-encoded to generate a texture pulse sequence, and texture event markers and direction guidance markers are configured.

[0011] A topological causal liquid state machine model is constructed based on texture event tags and direction guidance tags. The partitioned liquid reservoirs are configured according to the texture event tags, and the connection weights of the reserve neurons are configured according to the direction guidance tags to establish the causal connection relationship between the reserve areas in the partitioned liquid reservoirs.

[0012] Based on the topological causal liquid state machine model, the texture pulse sequence is read, and the response of the reserve neuron associated with the position label is marked according to the acquisition order. The state evolution is carried out through the connection weight of the reserve neuron and the causal connection relationship to generate the liquid state path.

[0013] The liquid state path is merged to form a set of defect trajectory segments. The set of defect trajectory segments is then subjected to category mapping to generate nonwoven fabric defect detection and classification results.

[0014] Optionally, the generation of the co-location texture frame sequence includes the following steps:

[0015] Frame serialization processing is performed on the continuous surface image of nonwoven fabric to generate continuous surface image frames;

[0016] Perform illumination homogenization, noise suppression, and motion displacement correction on continuous surface image frames to generate a corrected image frame sequence;

[0017] The texture region correspondence between adjacent corrected image frames is established based on the nonwoven fabric running direction and acquisition sequence.

[0018] The corrected image frame sequence is arranged in the same position according to the correspondence of texture regions to generate a corresponding texture frame sequence.

[0019] Optionally, the generation of the fiber micro-element topological connection state sequence includes the following steps:

[0020] Read the sequence of co-position texture frames, select texture blocks with the same corresponding markers, perform fiber texture backbone extraction on the corrected image content corresponding to the texture blocks, and generate a set of fiber texture backbones.

[0021] Fiber micro-elements are divided according to the direction and intersection of the fiber texture backbone, and a set of fiber micro-elements with position markers and collection order markers is generated.

[0022] Based on the fiber element set, the connection endpoints, connection directions, and adjacent objects between adjacent fiber elements are recorded to generate the fiber element connection state;

[0023] Read the texture blocks under the same corresponding marker, arrange the fiber element connection states according to the acquisition order, and generate a fiber element topology connection state sequence.

[0024] Optionally, the generation of the target texture event set includes the following steps:

[0025] Read the fiber micro-element topology connection state sequence, and compare the connection endpoints, connection directions and adjacent objects marked at the same position according to the acquisition order to generate connection state comparison results;

[0026] Based on the connection status comparison results, identify the disappearance of connection endpoints, changes in connection direction, and changes in adjacent objects, and generate fiber micro-element connection changes;

[0027] Using the acquisition sequence mark corresponding to the fiber micro-element connection change as the current acquisition sequence mark, read the connection endpoint, connection direction and adjacent object located after the current acquisition sequence mark under the same position mark, and record the fiber micro-element connection change that has not been restored to the previous connection state as the fiber micro-element irreversible connection change;

[0028] Based on the irreversible connection changes of fiber micro-elements, defect-type event start markers are generated. The fiber micro-element topological connection state sequence is divided into event segments according to the defect-type event start markers to generate a target texture event set.

[0029] In the target texture event set, retain the acquisition sequence marker, position marker, and irreversible connection changes of fiber micro-elements corresponding to the starting marker of defect-type events.

[0030] Optionally, the generation of the texture pulse sequence includes the following steps:

[0031] Read the acquisition sequence marker, position marker, and irreversible connection changes of fiber micro-elements from the target texture event set, arrange the target texture events according to the acquisition sequence marker, and generate a pulse coding sequence;

[0032] Configure texture event markers based on the connection endpoints, connection directions, and changes in adjacent objects during irreversible connection changes of fiber micro-elements; configure direction guidance markers based on the connection directions corresponding to irreversible connection changes of fiber micro-elements.

[0033] According to the pulse coding order, the acquisition sequence marker, position marker, texture event marker, and direction guide marker are written into the pulse coding unit to generate a texture pulse sequence.

[0034] Optionally, the construction of the topological causal liquid state machine model includes the following steps:

[0035] Read the texture event markers and direction guidance markers, and establish the input correspondence between the texture pulse sequence and the topological causal liquid state machine model;

[0036] Configure partitioned liquid reservoirs according to texture event markers, and each reservoir area within a partitioned liquid reservoir receives texture pulses corresponding to the texture event markers.

[0037] The connection weights of the reserve neurons are configured based on the directional guidance markers, carrying the directional transmission relationship corresponding to the texture pulses;

[0038] Based on the texture pulses received by the storage areas in the partitioned liquid storage pool and the connection weights of the storage neurons, a causal connection relationship between the storage areas is established, and a topological causal liquid state machine model is generated.

[0039] Optionally, the configuration of the partitioned liquid storage tank includes the following steps:

[0040] Read the texture event markers in the texture pulse sequence and identify disconnection markers, orientation change markers, and adjacent replacement markers;

[0041] In the topological causal liquid state machine model, a disconnected storage area, a directional change storage area, and an adjacent replacement storage area are configured to form a partitioned liquid storage pool;

[0042] Write texture pulses carrying disconnection class markers to the disconnection reserve area, write texture pulses carrying direction change class markers to the direction change reserve area, and write texture pulses carrying adjacent replacement class markers to the adjacent replacement reserve area.

[0043] When the same texture pulse carries two or more texture event markers, the same texture pulse is written into the corresponding storage area respectively, and each corresponding storage area retains the same acquisition sequence marker and the same location marker.

[0044] Optionally, configuring the connection weights of the reserve neurons includes the following steps:

[0045] Read the direction guide markers in the texture pulse sequence to determine the connection direction changes corresponding to the texture pulses;

[0046] Within the partitioned liquid reservoir, locate the reserve neuron connection edges corresponding to the texture pulse positions. Define the reserve neuron connection edges that continue the change in connection direction as directional connection edges, and define the reserve neuron connection edges that deviate from the change in connection direction as directional deviation connection edges.

[0047] Perform directional response enhancement processing on the directional connecting edges and deviation response suppression processing on the directional deviating connecting edges to generate reserve neuron connection weights;

[0048] Based on the connection weights of the reserve neurons, the transmission of texture pulses in the liquid state within the partitioned liquid reservoir is controlled, generating a liquid state path with directional constraints.

[0049] Optionally, the generation of the liquid state path includes the following steps:

[0050] The texture pulse sequence is marked and read according to the acquisition order, and the read texture pulse is submitted to the input of the topological causal liquid state machine model;

[0051] Based on location markers, the associated reserve neurons in the partitioned liquid reserve pool are triggered to generate the initial liquid state;

[0052] The initial liquid state is transmitted along the connection edges of the reserve neurons according to the connection weights of the reserve neurons, and is linked between the reserve regions according to the causal connection relationship to generate the liquid state of the reserve region;

[0053] The liquid state of the storage area is collected and marked according to the location and collection order, and a liquid state path is generated.

[0054] Optionally, the generation of the nonwoven fabric defect detection and classification results includes the following steps:

[0055] The liquid state path is divided into continuous trajectories according to location markers and acquisition sequence markers to form candidate trajectory segments;

[0056] Based on the reserve region and reserve neuron response in the candidate trajectory segment, a merging process is performed on the candidate trajectory segment to generate a set of defective trajectory segments;

[0057] Establish a category mapping relationship based on the reserve region to which the defect trajectory fragments belong in the set of defective trajectories, and determine the category response state based on the response of the reserve neurons;

[0058] The defect location is marked based on the location markers in the defect trajectory fragment set, the defect range is determined based on the acquisition sequence markers, and the nonwoven fabric defect detection and classification results are output by combining the category mapping relationship and the category response status.

[0059] The beneficial effects of this invention are:

[0060] (1) The present invention tracks the continuous changes of the same texture region by using the co-position texture frame sequence and the fiber micro-element topology connection state sequence, thereby reducing false detections caused by normal texture fluctuations and improving the stability of small defect detection.

[0061] (2) The present invention generates defect-type event initiation markers through irreversible connection changes of fiber micro-elements. Low-contrast defects and boundary blur defects are identified in the form of structural anomalies, reducing the false negative rate.

[0062] (3) This invention forms a liquid state path through a topological causal liquid state machine model, performs trajectory merging and category mapping, and improves the classification accuracy of defects such as holes, wrinkles and fiber agglomeration. Attached Figure Description

[0063] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0064] Figure 1 This is an overall flowchart of a deep learning-based method for detecting and classifying defects in nonwoven fabrics proposed in this invention.

[0065] Figure 2 This is a flowchart of the texture pulse sequence generation process in this invention;

[0066] Figure 3 This is a schematic diagram of the topological causal liquid state machine model in this invention. Detailed Implementation

[0067] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0068] refer to Figures 1-3 A method for detecting and classifying defects in nonwoven fabrics based on deep learning, comprising the following steps:

[0069] Acquire images of the continuous surface of the nonwoven fabric, perform image correction processing, and generate a sequence of isomorphic texture frames.

[0070] Based on the co-position texture frame sequence, construct the fiber element connection state and generate the fiber element topological connection state sequence;

[0071] Irreversible connection changes of fiber micro-elements are identified based on the fiber micro-element topological connection state sequence, defect-type event start markers are generated, and target texture event sets are generated according to the defect-type event start markers.

[0072] The target texture event set is pulse-encoded to generate a texture pulse sequence, and texture event markers and direction guidance markers are configured.

[0073] A topological causal liquid state machine model is constructed based on texture event tags and direction guidance tags. The partitioned liquid reservoirs are configured according to the texture event tags, and the connection weights of the reserve neurons are configured according to the direction guidance tags to establish the causal connection relationship between the reserve areas in the partitioned liquid reservoirs.

[0074] Based on the topological causal liquid state machine model, the texture pulse sequence is read, and the response of the reserve neuron associated with the position label is marked according to the acquisition order. The state evolution is carried out through the connection weight of the reserve neuron and the causal connection relationship to generate the liquid state path.

[0075] The liquid state path is merged to form a set of defect trajectory segments. The set of defect trajectory segments is then subjected to category mapping to generate nonwoven fabric defect detection and classification results.

[0076] In this embodiment, the generation of the co-location texture frame sequence includes the following steps:

[0077] Frame serialization processing is performed on the continuous surface image of nonwoven fabric to generate continuous surface image frames;

[0078] Perform illumination homogenization, noise suppression, and motion displacement correction on continuous surface image frames to generate a corrected image frame sequence;

[0079] The acquisition time, production line conveying direction, and intra-frame texture block coordinates of each frame are read. The areas with continuous fiber direction and traceable local texture structure in adjacent frames are used as displacement references to form inter-frame displacement records. The position of the next frame image is then adjusted according to the inter-frame displacement records.

[0080] The texture region correspondence between adjacent corrected image frames is established based on the nonwoven fabric running direction and acquisition sequence.

[0081] The corrected image frames are divided into texture blocks. The fiber direction, texture boundary and local texture structure of each texture block are extracted. Texture blocks with continuous structure are searched in adjacent frames according to the acquisition order to establish the correspondence between texture regions of consecutive frames.

[0082] The corrected image frame sequence is arranged in the same position according to the correspondence of texture regions to generate a corresponding texture frame sequence.

[0083] The system reads the positions of texture blocks in the preceding and following frames, as well as the acquisition order, from the texture region correspondence. It then writes the texture blocks corresponding to the positions in the preceding and following frames into the same corresponding marker. Texture blocks in the corrected image frame sequence are grouped according to these corresponding markers. Texture blocks under the same corresponding marker are arranged in the acquisition order, and the arrangement result saves the texture block positions, corresponding markers, and corrected image content, generating a corresponding texture frame sequence.

[0084] In this embodiment, the generation of the fiber micro-element topological connection state sequence includes the following steps:

[0085] Read the sequence of co-position texture frames, select texture blocks with the same corresponding markers, perform fiber texture backbone extraction on the corrected image content corresponding to the texture blocks, and generate a set of fiber texture backbones.

[0086] When processing the same texture frame sequence, the texture blocks are first grouped according to the corresponding markers, with each group corresponding to the same fabric position. Then, the corrected image content of the texture blocks within each group is read to identify continuous grayscale banded textures, forming candidate fiber extension regions. Centerline tracing is performed on these candidate fiber extension regions, with the tracing direction unfolding along the local fiber orientation. When a break occurs in the centerline, it is connected according to the orientation relationship between the two ends of the break. Isolated texture points and short, fragmented texture segments do not participate in subsequent processing; the retained centerlines form the main set of fiber textures.

[0087] Fiber micro-elements are divided according to the direction and intersection of the fiber texture backbone, and a set of fiber micro-elements with position markers and collection order markers is generated.

[0088] In the set of fiber texture trunks, the continuous extension of the center line is recorded as the direction of continuity, and the positions where the center lines contact, cross, and connect at the ends are recorded as the intersection relationship. The start and end positions of each fiber texture trunk are saved synchronously.

[0089] The locations of directional changes in the fiber texture backbone are determined based on the continuity of the direction, the intersection locations between different fiber texture backbones are determined based on the intersection relationship, and the end locations are determined by combining the start and end locations of the fiber texture backbone. The intersection locations, end locations, and directional change locations together serve as the dividing points of the fiber texture backbone. On the same fiber texture backbone, the dividing points are arranged according to the backbone's extension direction, and the backbone segments between adjacent dividing points form fiber micro-elements. When the dividing point is located at an intersection location, each backbone branch participating in the intersection generates a corresponding fiber micro-element.

[0090] Each fiber element is configured with a position marker and an acquisition sequence marker. The position marker is formed by the texture block position and the fiber element's position within the texture block, while the acquisition sequence marker is formed by the acquisition sequence of the image frames to which the fiber element belongs.

[0091] Based on the fiber element set, the connection endpoints, connection directions, and adjacent objects between adjacent fiber elements are recorded to generate the fiber element connection state;

[0092] After completing the fiber element division, other connected fiber elements are searched at the ends of each fiber element. The location where the connection occurs is recorded as the connection endpoint, the direction of extension at the connection is recorded as the connection direction, and the objects connected to the current fiber element are recorded as adjacent objects.

[0093] Read the texture blocks under the same corresponding marker, arrange the fiber element connection states according to the acquisition order, and generate a fiber element topology connection state sequence.

[0094] Fiber micro-element connection records with the same corresponding marker are entered into the same processing queue, and the records in the queue are arranged in the order of acquisition. Each record retains the acquisition order marker, position marker, connection endpoint, connection direction, and adjacent objects, thereby forming a comparable record of connection state changes.

[0095] In this embodiment, the generation of the target texture event set includes the following steps:

[0096] Read the fiber micro-element topology connection state sequence, and compare the connection endpoints, connection directions and adjacent objects marked at the same position according to the acquisition order to generate connection state comparison results;

[0097] Under the same location marking, the fiber micro-element connection status is marked and arranged according to the acquisition order. The connection status that comes first in the adjacent arrangement is determined as the previous connection status, and the connection status that comes later is determined as the current connection status. Then, the connection endpoints, connection directions, and adjacent objects are compared item by item, and the differences in the fields are written into the connection status comparison results.

[0098] Based on the connection status comparison results, identify the disappearance of connection endpoints, changes in connection direction, and changes in adjacent objects, and generate fiber micro-element connection changes;

[0099] When the current connection state lacks a connection endpoint from the preceding connection state, the connection endpoint disappears; when the connection direction changes in the current connection state, the connection direction changes; when the adjacent objects change in the current connection state, the adjacent objects change. These changes result in changes in fiber element connections.

[0100] Using the acquisition sequence mark corresponding to the fiber micro-element connection change as the current acquisition sequence mark, read the connection endpoint, connection direction and adjacent object located after the current acquisition sequence mark under the same position mark, and record the fiber micro-element connection change that has not been restored to the previous connection state as the fiber micro-element irreversible connection change;

[0101] When a fiber element connection change occurs, the corresponding acquisition sequence marker is determined as the current acquisition sequence marker. Under the same position marker, the connection state following the current acquisition sequence marker is read; if the read connection endpoints, connection directions, and adjacent objects do not return to the corresponding content in the previous connection state, an irreversible connection change of the fiber element is formed.

[0102] Based on the irreversible connection changes of fiber micro-elements, defect-type event start markers are generated. The fiber micro-element topological connection state sequence is divided into event segments according to the defect-type event start markers to generate a target texture event set.

[0103] The acquisition sequence markers and position markers corresponding to irreversible connection changes in fiber micro-elements are written into the defect-type event start marker. Starting from the acquisition sequence marker in the defect-type event start marker, the fiber micro-element topological connection state sequence located after the start point is read under the same position marker to form the target texture event set.

[0104] In the target texture event set, retain the acquisition sequence marker, position marker, and irreversible connection changes of fiber micro-elements corresponding to the starting marker of defect-type events.

[0105] The target texture event set stores the acquisition sequence marker, position marker, and irreversible connection changes of fiber micro-elements corresponding to the starting marker of defect-type events. Pulsed encoding can read the event start point, event position, and irreversible connection change content.

[0106] In this embodiment, the generation of the texture pulse sequence includes the following steps:

[0107] Read the acquisition sequence marker, position marker, and irreversible connection changes of fiber micro-elements from the target texture event set, arrange the target texture events according to the acquisition sequence marker, and generate a pulse coding sequence;

[0108] Read the acquisition sequence markers from the target texture event set, and arrange the target texture events according to the acquisition sequence markers; retain the corresponding position markers during arrangement to form the writing order relationship of the texture pulse sequence.

[0109] Configure texture event markers based on the connection endpoints, connection directions, and changes in adjacent objects during irreversible connection changes of fiber micro-elements; configure direction guidance markers based on the connection directions corresponding to irreversible connection changes of fiber micro-elements.

[0110] Texture event tags are written according to the type of irreversible connection change in the fiber element. The disappearance of a connection endpoint corresponds to a connection break tag, a change in connection direction corresponds to a direction change tag, and a change in adjacent objects corresponds to an adjacent replacement tag. When the irreversible connection change of the same fiber element includes multiple changes, multiple category tags are retained simultaneously. Direction guidance tags record the direction of change from the previous connection direction to the current connection direction, used for subsequent directional configuration of reserve neuron connection weights.

[0111] According to the pulse coding order, the acquisition sequence marker, position marker, texture event marker, and direction guide marker are written into the pulse coding unit to generate a texture pulse sequence.

[0112] The individual pulse records in the texture pulse sequence include acquisition sequence markers, position markers, texture event markers, and direction guidance markers. The acquisition sequence marker is used to determine the order in which pulses are written, the position marker is used to determine the position where pulses are written, and the texture event marker and direction guidance marker are used for reading by the partitioned liquid storage tank.

[0113] In this embodiment, the construction of the topological causal liquid state machine model includes the following steps:

[0114] Read the texture event markers and direction guidance markers, and establish the input correspondence between the texture pulse sequence and the topological causal liquid state machine model;

[0115] The input correspondence is used to establish the selection relationship between texture event markers and reserve areas, as well as the configuration relationship between direction guidance markers and reserve neuron connection weights.

[0116] Configure partitioned liquid reservoirs according to texture event markers, and each reservoir area within a partitioned liquid reservoir receives texture pulses corresponding to the texture event markers.

[0117] The storage areas within the partitioned liquid storage tank are set according to the texture event marker category. When a texture pulse enters the partitioned liquid storage tank, the texture event marker is read first, and then the pulse enters the corresponding storage area according to the marker category. When the same texture pulse carries multiple texture event markers, each pulse enters the corresponding storage area, and the acquisition sequence marker and location marker are retained after entering the storage area.

[0118] The connection weights of the reserve neurons are configured based on the directional guidance markers, carrying the directional transmission relationship corresponding to the texture pulses;

[0119] Reserve neurons act as computational nodes within a partitioned liquid reservoir, receiving texture pulses and generating liquid states. Each reserve neuron is associated with a corresponding reservoir region, a location marker, and a connection edge between reserve neurons. Directional guidance markers are used to establish directional transmission relationships within the reservoir region, allowing liquid states to propagate between reserve neurons along the corresponding directions.

[0120] Based on the texture pulses received by the storage areas in the partitioned liquid storage pool and the connection weights of the storage neurons, a causal connection relationship between the storage areas is established, and a topological causal liquid state machine model is generated.

[0121] Liquid responses are transmitted between the various reservoirs through causal connections. A liquid state path formed in one reservoir can influence the liquid response of another reservoir. The affected object is determined by both location markers and texture event markers. The topological causal liquid state machine model stores the input correspondence, partitioned liquid reservoirs, reservoir neuron connection weights, and causal connections to support subsequent liquid state evolution.

[0122] In this embodiment, the configuration of the partitioned liquid storage tank includes the following steps:

[0123] Read the texture event markers in the texture pulse sequence and identify disconnection markers, orientation change markers, and adjacent replacement markers;

[0124] Before the texture pulse enters the partitioned liquid reservoir, the texture event markers are read. The reading results are distinguished by disconnection markers, direction change markers, and adjacent replacement markers, so that the subsequent reservoir can obtain clear input categories.

[0125] In the topological causal liquid state machine model, a disconnected storage area, a directional change storage area, and an adjacent replacement storage area are configured to form a partitioned liquid storage pool;

[0126] The disconnection reserve region corresponds to the liquid response caused by the disappearance of the fiber element connection endpoints; the direction change reserve region corresponds to the liquid response caused by the change in the connection direction of the fiber element; and the adjacent replacement reserve region corresponds to the liquid response caused by the change in the adjacent objects of the fiber element. These three reserve regions respectively store the liquid evolution states caused by different connection changes, preventing different defect disturbances from mixing within the same reserve region.

[0127] Write texture pulses carrying disconnection class markers to the disconnection reserve area, write texture pulses carrying direction change class markers to the direction change reserve area, and write texture pulses carrying adjacent replacement class markers to the adjacent replacement reserve area.

[0128] After the texture pulse is written to the corresponding storage area, the acquisition sequence marker, position marker, texture event marker, and direction guidance marker are saved along with it to participate in the evolution of the liquid state in the storage area, thus avoiding the loss of event source and direction information after writing.

[0129] When the same texture pulse carries two or more texture event markers, the same texture pulse is written into the corresponding storage area respectively, and each corresponding storage area retains the same acquisition sequence marker and the same location marker.

[0130] When a single texture pulse contains multiple texture event markers, they are each moved to their respective storage areas. Each storage area retains markers with the same acquisition order, the same location, and the same direction guidance, allowing subsequent causal connections to identify the source of the same texture event.

[0131] In this embodiment, the configuration of the reserve neuron connection weights includes the following steps:

[0132] Read the direction guide markers in the texture pulse sequence to determine the connection direction changes corresponding to the texture pulses;

[0133] The direction guide marker stores the previous connection direction and the current connection direction. During processing, both direction contents are read, and the pointing relationship from the previous connection direction to the current connection direction is determined as the connection direction change.

[0134] Within the partitioned liquid reservoir, locate the reserve neuron connection edges corresponding to the texture pulse positions. Define the reserve neuron connection edges that continue the change in connection direction as directional connection edges, and define the reserve neuron connection edges that deviate from the change in connection direction as directional deviation connection edges.

[0135] Within the corresponding reserve area, reserve neurons are located based on the position markers of texture pulses, and the connection edges between these reserve neurons and adjacent reserve neurons are read. When the extension direction of the connection edge is consistent with the change in the connection direction, it is assigned to the direction of the connecting edge; otherwise, it is assigned to the direction of the deviating connection edge.

[0136] Perform directional response enhancement processing on the directional connecting edges and deviation response suppression processing on the directional deviating connecting edges to generate reserve neuron connection weights;

[0137] Directional receiving edges are configured as enhancing transmission states, while directional deviating edges are configured as inhibiting transmission states. These two types of transmission states together form the reserve neuron connection weights. The reserve neuron connection weights act on the reserve neuron connection edges, controlling the transmission response when the liquid state is transmitted from the current reserve neuron to adjacent reserve neurons.

[0138] Based on the connection weights of the reserve neurons, the transmission of texture pulses in the liquid state within the partitioned liquid reservoir is controlled, generating a liquid state path with directional constraints.

[0139] The liquid state is a state representation formed by the responses of reserve neurons, the propagation of connection edges, and the state updates within the reserve region after a texture pulse is written into the partitioned liquid reservoir. After the texture pulse is written into the partitioned liquid reservoir, it first triggers the response of the reserve neuron corresponding to the position marker, and then propagates to adjacent reserve neurons according to the connection weights of the reserve neurons. The liquid state extends preferentially along the direction of the connection edge, and is suppressed along the direction of the connection edge, forming a liquid state path constrained by the direction-guided marker.

[0140] In this embodiment, the generation of the liquid state path includes the following steps:

[0141] The texture pulse sequence is marked and read according to the acquisition order, and the read texture pulse is submitted to the input of the topological causal liquid state machine model;

[0142] The input terminal receives texture pulses read according to the acquisition sequence marker, and reads the acquisition sequence marker, position marker, texture event marker, and direction guidance marker. The acquisition sequence marker is used to determine the order in which texture pulses enter the topological causal liquid state machine model.

[0143] Based on location markers, the associated reserve neurons in the partitioned liquid reserve pool are triggered to generate the initial liquid state;

[0144] Based on the location markers, the corresponding reserve neurons are located within the partitioned liquid reservoir. These reserve neurons are identified as associated reserve neurons. The associated reserve neurons read the texture event markers and generate connection break, direction change, and adjacency replacement responses according to the connection break, direction change, and adjacency replacement markers, respectively. The associated reserve neurons continue reading the direction guidance markers, incorporating the change from the previous connection direction to the current connection direction into the corresponding response, generating an associated reserve neuron response. The associated reserve neuron response retains the acquisition sequence marker and location marker, serving as the starting point for the transmission of the initial liquid state to the connection edges of the reserve neurons.

[0145] The initial liquid state is transmitted along the connection edges of the reserve neurons according to the connection weights of the reserve neurons, and is linked between the reserve regions according to the causal connection relationship to generate the liquid state of the reserve region;

[0146] The initial liquid state originates from the associated reserve neuron and propagates along the connection edges of the associated reserve neuron. When the connection edge is in the direction of connection, enhancement propagation is performed; when the connection edge deviates from the direction, inhibition propagation is performed, forming the connection edge propagation result. The source and target reserve regions recorded in the causal connections are read. The reserve region that generates the connection edge propagation result is determined as the source reserve region, and the reserve region pointed to by the causal connections is determined as the target reserve region. Under the same location marking and acquisition sequence marking, the connection edge propagation result of the source reserve region is sent to the corresponding reserve neuron in the target reserve region, and the corresponding reserve neuron in the target reserve region updates its response accordingly. The connection edge propagation result of the source reserve region and the updated reserve neuron response of the target reserve region together form the liquid state of the reserve region.

[0147] The liquid state of the storage area is collected and marked according to the location and collection order, and a liquid state path is generated.

[0148] When generating a liquid state path, the liquid states in the storage area are first collected according to the location markers, and then the liquid states in the storage area are arranged according to the collection order under the same location markers. The arranged state sequence forms the liquid state path.

[0149] In this embodiment, the generation of nonwoven fabric defect detection and classification results includes the following steps:

[0150] The liquid state path is divided into continuous trajectories according to location markers and acquisition sequence markers to form candidate trajectory segments;

[0151] Under the same location marker, the liquid state of the storage area is arranged according to the collection order. When adjacent collection order markers all contain liquid state in the storage area, the relevant states are assigned to the same candidate trajectory segment; when the collection order is interrupted, the current candidate trajectory segment ends.

[0152] Based on the reserve region and reserve neuron response in the candidate trajectory segment, a merging process is performed on the candidate trajectory segment to generate a set of defective trajectory segments;

[0153] Candidate trajectory segments with the same location label are entered into a merging process. If there is a causal connection between the reserve regions corresponding to the candidate trajectory segments, and the responses of the reserve neurons point to the same connection change process, then the related candidate trajectory segments are merged into defective trajectory segments.

[0154] Establish a category mapping relationship based on the reserve region to which the defect trajectory fragments belong in the set of defective trajectories, and determine the category response state based on the response of the reserve neurons;

[0155] The category mapping is determined based on the reserve region involved in the defect trajectory segment. A disconnected reserve region corresponds to a fiber breakage category mapping, a direction-change reserve region corresponds to a wrinkle category mapping, and an adjacent replacement reserve region corresponds to a fiber clustering category mapping. When multiple reserve regions coexist, a composite category mapping is formed based on causal connections. The response of a reserve neuron that continuously traverses the defect trajectory segment and corresponds to the category mapping is determined as the category response state.

[0156] The defect location is marked based on the location markers in the defect trajectory fragment set, the defect range is determined based on the acquisition sequence markers, and the nonwoven fabric defect detection and classification results are output by combining the category mapping relationship and the category response status.

[0157] The location markers in the defect trajectory segment are mapped to the defect locations. The texture block coverage areas corresponding to the start and end acquisition sequence markers in the defect trajectory segment are determined as the defect ranges. The defect location, defect range, category mapping relationship, and category response status together form the nonwoven fabric defect detection and classification results.

[0158] Example 1: To verify the feasibility of this invention in practice, it was applied to a continuous production quality inspection scenario for nonwoven fabrics. In this scenario, the surface of the nonwoven fabric exhibits fiber interlacing, uneven local pores, variations in texture depth, and slight wrinkle transition areas. Some defects appear similar to normal texture fluctuations in a single-frame image, especially light-colored stains, minor breaks, slight fiber agglomerations, and areas of uneven thickness with unclear boundaries. Using manual visual inspection or ordinary image block classification methods can easily lead to misjudgments and missed detections. To closely resemble a continuous production environment, this example selects continuously acquired images of the nonwoven fabric surface as the inspection object. The images simultaneously include areas of normal texture, low-contrast defect areas, small defect areas, and defect areas with large morphological variations.

[0159] In practical applications, continuously acquired images of the nonwoven fabric surface are first corrected to obtain a sequence of co-located texture frames, allowing the tracking of changes in the same texture region during continuous acquisition. Then, the fiber texture backbone is extracted from the co-located texture frame sequence, and fiber micro-elements are divided according to the direction, intersection, and end positions of the fiber backbone. The connection endpoints, connection directions, and adjacent objects of each fiber micro-element are recorded. For normal texture regions, the connection state between fiber micro-elements remains stable during continuous acquisition; even slight grayscale fluctuations do not lead to continuous irreversible connection changes. For real defect regions, fiber micro-elements are prone to disappearance of connection endpoints, changes in connection direction, or changes in adjacent objects, and these changes cannot be restored to the previous connection state in subsequent acquisition sequences, thus forming a defect-type event initiation marker.

[0160] After splitting the target texture event set, the target texture events are encoded into texture pulse sequences. Each texture pulse carries a texture event marker and a direction guidance marker. The disconnection marker reflects fiber breakage, hole edges, and missing connections; the direction change marker reflects wrinkles, local stretching, and abrupt changes in orientation; and the adjacent replacement marker reflects fiber aggregation, interlacing anomalies, and local structural reorganization. The direction guidance marker records the change in connection direction from the previous state to the current state. The subsequent reserve neuron connection weights can control the propagation of the liquid state along the defect perturbation direction, preventing the defect response from being smoothed out by normal texture during normal diffusion.

[0161] The topological causal liquid state machine model configures partitioned liquid reservoirs during use. Disconnection reservoirs, orientation change reservoirs, and adjacent replacement reservoirs handle different types of texture event responses. When a texture impulse enters the model, it triggers the corresponding reservoir neuron response based on its location marker, forming an initial liquid state. This initial liquid state is propagated along the connection edges of the reservoir neurons; directional connection edges perform reinforcement propagation, while directional deviation edges perform inhibition propagation. State propagation occurs between different reservoirs based on causal connections. Disconnection, orientation change, and adjacent replacement responses under the same location marker and acquisition sequence marker can form merging liquid state paths. Finally, the liquid state paths are merged to form a set of defect trajectory fragments, and the nonwoven fabric defect detection and classification results are output through category mapping relationships and category response states.

[0162] To verify the feasibility of this invention in practice, it was applied to a continuous production quality inspection scenario for nonwoven fabrics. This scenario selected continuous image samples containing normal textures, holes, stains, wrinkles, fiber agglomerations, uneven thickness, and texture breaks as the inspection objects. The method of this invention was compared with manual visual inspection, conventional image block classification methods, conventional convolutional detection methods, and other methods. Evaluation metrics were selected that are statistically applicable to each method, including defect detection rate, defect missed detection rate, false detection rate in normal areas, low-contrast defect detection rate, blurred boundary defect detection rate, and defect classification accuracy. Specific results are shown in Table 1.

[0163] Table 1. Comparison of General Effects of Defect Detection and Classification in Nonwoven Fabrics

[0164] Defect detection rate 83.42% 87.40% 91.85% 95.62% Defect missed detection rate 16.58% 12.60% 8.15% 4.38% False detection rate in normal areas 1.35% 1.14% 0.84% 0.46% Low contrast defect detection rate 76.31% 80.48% 86.12% 92.36% Detection rate of defects with blurred boundaries 71.94% 78.22% 84.05% 90.47% Defect classification accuracy 80.09% 84.26% 88.94% 92.86%

[0165] As shown in Table 1, the method of this invention outperforms manual visual inspection, conventional image block classification methods, and conventional convolutional detection methods in all general detection metrics. The overall defect detection rate reaches 95.62%, a further improvement compared to the 91.85% of conventional convolutional detection methods. This indicates that the invention does not rely solely on image grayscale or static texture features for judgment, but rather identifies structural changes in defects during continuous acquisition through isotopic texture frame sequences, irreversible connections of fiber micro-elements, and liquid state paths, thereby improving the overall detection capability.

[0166] Regarding the control of missed detections, the defect missed detection rate of this invention is 4.38%, lower than the 8.15% of conventional convolutional detection methods. This result indicates that for defects that are not easily highlighted in a single frame image, such as small fractures, light-colored stains, and local thickness unevenness, this invention can capture them through the continuous changes in the connection state of fiber micro-elements, avoiding weak defects being masked by normal fiber texture fluctuations. In terms of the false detection rate in normal regions, this invention has a rate of 0.46%, lower than other methods, indicating that the topological causal liquid state machine model can distinguish between normal texture changes and real defect disturbances, reducing the misjudgment of normal fiber interlacing and porosity changes as defects.

[0167] For the low-contrast defects and blurred boundary defects that this invention focuses on addressing, the invention achieves detection rates of 92.36% and 90.47%, respectively, both higher than conventional convolution detection methods. This indicates that the invention defines the defect origin as an irreversible change in the connection of fiber micro-elements, and forms a state path through directional guidance markers and partitioned liquid reservoirs, thus preserving structural anomaly information in low-contrast defects and boundary transition regions.

[0168] The defect classification accuracy reached 92.86%, indicating that the liquid state response corresponding to different storage areas can provide a more stable basis for defect category judgment, improving the classification consistency of defects with large morphological differences such as holes, wrinkles, and fiber agglomerations. Overall, the data in Table 1 supports the technical effectiveness of this invention in terms of weak defect detection, false detection suppression, and defect classification stability.

[0169] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for detecting and classifying defects in nonwoven fabrics based on deep learning, characterized in that, Includes the following steps: Acquire images of the continuous surface of the nonwoven fabric, perform image correction processing, and generate a sequence of isomorphic texture frames. Based on the co-position texture frame sequence, construct the fiber element connection state and generate the fiber element topological connection state sequence; Irreversible connection changes of fiber micro-elements are identified based on the fiber micro-element topological connection state sequence, defect-type event start markers are generated, and target texture event sets are generated according to the defect-type event start markers. The target texture event set is pulse-encoded to generate a texture pulse sequence, and texture event markers and direction guidance markers are configured. A topological causal liquid state machine model is constructed based on texture event tags and direction guidance tags. The partitioned liquid reservoirs are configured according to the texture event tags, and the connection weights of the reserve neurons are configured according to the direction guidance tags to establish the causal connection relationship between the reserve areas in the partitioned liquid reservoirs. Based on the topological causal liquid state machine model, the texture pulse sequence is read, and the response of the reserve neuron associated with the position label is marked according to the acquisition order. The state evolution is carried out through the connection weight of the reserve neuron and the causal connection relationship to generate the liquid state path. The liquid state path is merged to form a set of defect trajectory segments. The set of defect trajectory segments is then subjected to category mapping to generate nonwoven fabric defect detection and classification results.

2. The method for detecting and classifying defects in nonwoven fabrics based on deep learning according to claim 1, characterized in that, The generation of the co-location texture frame sequence includes the following steps: Frame serialization processing is performed on the continuous surface image of nonwoven fabric to generate continuous surface image frames; Perform illumination homogenization, noise suppression, and motion displacement correction on continuous surface image frames to generate a corrected image frame sequence; The texture region correspondence between adjacent corrected image frames is established based on the nonwoven fabric running direction and acquisition sequence. The corrected image frame sequence is arranged in the same position according to the correspondence of texture regions to generate a corresponding texture frame sequence.

3. The method for detecting and classifying defects in nonwoven fabrics based on deep learning according to claim 1, characterized in that, The generation of the fiber micro-element topological connection state sequence includes the following steps: Read the sequence of co-position texture frames, select texture blocks with the same corresponding markers, perform fiber texture backbone extraction on the corrected image content corresponding to the texture blocks, and generate a set of fiber texture backbones. Fiber micro-elements are divided according to the direction and intersection of the fiber texture backbone, and a set of fiber micro-elements with position markers and collection order markers is generated. Based on the fiber element set, the connection endpoints, connection directions, and adjacent objects between adjacent fiber elements are recorded to generate the fiber element connection state; Read the texture blocks under the same corresponding marker, arrange the fiber element connection states according to the acquisition order, and generate a fiber element topology connection state sequence.

4. The method for detecting and classifying defects in nonwoven fabrics based on deep learning according to claim 1, characterized in that, The generation of the target texture event set includes the following steps: Read the fiber micro-element topology connection state sequence, and compare the connection endpoints, connection directions and adjacent objects marked at the same position according to the acquisition order to generate connection state comparison results; Based on the connection status comparison results, identify the disappearance of connection endpoints, changes in connection direction, and changes in adjacent objects, and generate fiber micro-element connection changes; Using the acquisition sequence mark corresponding to the fiber micro-element connection change as the current acquisition sequence mark, read the connection endpoint, connection direction and adjacent object located after the current acquisition sequence mark under the same position mark, and record the fiber micro-element connection change that has not been restored to the previous connection state as the fiber micro-element irreversible connection change; Based on the irreversible connection changes of fiber micro-elements, defect-type event start markers are generated. The fiber micro-element topological connection state sequence is divided into event segments according to the defect-type event start markers to generate a target texture event set. In the target texture event set, retain the acquisition sequence marker, position marker, and irreversible connection changes of fiber micro-elements corresponding to the starting marker of defect-type events.

5. The method for detecting and classifying defects in nonwoven fabrics based on deep learning according to claim 1, characterized in that, The generation of the texture pulse sequence includes the following steps: Read the acquisition sequence marker, position marker, and irreversible connection changes of fiber micro-elements from the target texture event set, arrange the target texture events according to the acquisition sequence marker, and generate a pulse coding sequence; Configure texture event markers based on the connection endpoints, connection directions, and changes in adjacent objects during irreversible connection changes of fiber micro-elements; configure direction guidance markers based on the connection directions corresponding to irreversible connection changes of fiber micro-elements. According to the pulse coding order, the acquisition sequence marker, position marker, texture event marker, and direction guide marker are written into the pulse coding unit to generate a texture pulse sequence.

6. The method for detecting and classifying defects in nonwoven fabrics based on deep learning according to claim 1, characterized in that, The construction of the topological causal liquid state machine model includes the following steps: Read the texture event markers and direction guidance markers, and establish the input correspondence between the texture pulse sequence and the topological causal liquid state machine model; Configure partitioned liquid reservoirs according to texture event markers, and each reservoir area within a partitioned liquid reservoir receives texture pulses corresponding to the texture event markers. The connection weights of the reserve neurons are configured based on the directional guidance markers, carrying the directional transmission relationship corresponding to the texture pulses; Based on the texture pulses received by the storage areas in the partitioned liquid storage pool and the connection weights of the storage neurons, a causal connection relationship between the storage areas is established, and a topological causal liquid state machine model is generated.

7. The method for detecting and classifying defects in nonwoven fabrics based on deep learning according to claim 6, characterized in that, The configuration of the partitioned liquid storage tank includes the following steps: Read the texture event markers in the texture pulse sequence and identify disconnection markers, orientation change markers, and adjacent replacement markers; In the topological causal liquid state machine model, a disconnected storage area, a directional change storage area, and an adjacent replacement storage area are configured to form a partitioned liquid storage pool; Write texture pulses carrying disconnection class markers to the disconnection reserve area, write texture pulses carrying direction change class markers to the direction change reserve area, and write texture pulses carrying adjacent replacement class markers to the adjacent replacement reserve area. When the same texture pulse carries two or more texture event markers, the same texture pulse is written into the corresponding storage area respectively, and each corresponding storage area retains the same acquisition sequence marker and the same location marker.

8. The method for detecting and classifying defects in nonwoven fabrics based on deep learning according to claim 6, characterized in that, The configuration of the reserve neuron connection weights includes the following steps: Read the direction guide markers in the texture pulse sequence to determine the connection direction changes corresponding to the texture pulses; Within the partitioned liquid reservoir, locate the reserve neuron connection edges corresponding to the texture pulse positions. Define the reserve neuron connection edges that continue the change in connection direction as directional connection edges, and define the reserve neuron connection edges that deviate from the change in connection direction as directional deviation connection edges. Perform directional response enhancement processing on the directional connecting edges and deviation response suppression processing on the directional deviating connecting edges to generate reserve neuron connection weights; Based on the connection weights of the reserve neurons, the transmission of texture pulses in the liquid state within the partitioned liquid reservoir is controlled, generating a liquid state path with directional constraints.

9. The method for detecting and classifying defects in nonwoven fabrics based on deep learning according to claim 1, characterized in that, The generation of the liquid state path includes the following steps: The texture pulse sequence is marked and read according to the acquisition order, and the read texture pulse is submitted to the input of the topological causal liquid state machine model; Based on location markers, the associated reserve neurons in the partitioned liquid reserve pool are triggered to generate the initial liquid state; The initial liquid state is transmitted along the connection edges of the reserve neurons according to the connection weights of the reserve neurons, and is linked between the reserve regions according to the causal connection relationship to generate the liquid state of the reserve region; The liquid state of the storage area is collected and marked according to the location and collection order, and a liquid state path is generated.

10. The method for detecting and classifying defects in nonwoven fabrics based on deep learning according to claim 1, characterized in that, The generation of the nonwoven fabric defect detection and classification results includes the following steps: The liquid state path is divided into continuous trajectories according to location markers and acquisition sequence markers to form candidate trajectory segments; Based on the reserve region and reserve neuron response in the candidate trajectory segment, a merging process is performed on the candidate trajectory segment to generate a set of defective trajectory segments; Establish a category mapping relationship based on the reserve region to which the defect trajectory fragments belong in the set of defective trajectories, and determine the category response state based on the response of the reserve neurons; The defect location is marked based on the location markers in the defect trajectory fragment set, the defect range is determined based on the acquisition sequence markers, and the nonwoven fabric defect detection and classification results are output by combining the category mapping relationship and the category response status.

Images