Lining paper image recognition method and device based on deep learning

Abstract

The embodiment of the invention discloses a rapid screening and distinguishing method and device for tobacco pesticide residues. The method comprises the following steps: acquiring a target image corresponding to lining paper; based on the target image, outputting a recognition result by the trained Lite-MTNet model; and adjusting the current conveying route of the lining paper based on the type of the lining paper, and generating warning information based on the defect area. In the embodiment, the target image can be processed according to the trained Lite-MTNet model, the model adopts a parameter sharing mode to extract low-level and high-level semantic features of the paper image, and combined optimization can be carried out on classification and segmentation tasks in combination with a dual-task combined loss function, so that a large amount of label data is not needed, and the classification and segmentation efficiency is improved. The method can quickly and continuously adapt to a new scene and a new defect type after deployment in a semi-supervised learning mode, thereby realizing high-performance identification and detection of the lining paper.

Description

Technical Field

[0001] The embodiments in this specification belong to the field of image recognition technology for inner lining paper, and in particular relate to an image recognition method and apparatus for inner lining paper based on deep learning. Background Technology

[0002] In modern cigarette packaging production lines, inner lining paper, as a key packaging material covering the appearance of cigarette products, comes in a wide variety of types. This includes differences in materials across different brands and specifications, as well as the widespread use of complex printed patterns and microstructure designs to enhance the packaging's moisture resistance, aesthetics, and brand recognition. However, the diversity of inner lining paper types also leads to difficulties in identifying them during production. Currently, manual visual inspection or simple barcode / QR code identification is generally used, but these methods have significant errors and insufficient resolution. Furthermore, while supervised learning models are used for identification, these models heavily rely on large amounts of high-quality labeled data. In industrial practice, establishing a detailed labeled sample library for thousands of diverse inner lining paper types is extremely costly, and manual labeling is inefficient and inconsistent. Summary of the Invention

[0003] The embodiments of this disclosure provide a method and apparatus for recognizing inner lining paper images based on deep learning, which aims to solve one or more of the above-mentioned problems and other potential problems.

[0004] According to a first aspect of this disclosure, a deep learning-based image recognition method for inner lining paper is provided. The method includes acquiring a target image corresponding to the inner lining paper; outputting a recognition result based on the target image using a trained Lite-MTNet model, the recognition result including the type of inner lining paper and defective regions. The backbone network of the Lite-MTNet model includes an LA-Conv Block module for generating a spatial-channel joint attention map based on a local attention mechanism, an MGF module for multi-scale feature fusion, and a CTFB module for generating general features shared between segmentation and classification branches. The total loss function of the Lite-MTNet model is a weighted sum of the segmentation loss function and the classification loss function, with higher weights assigned to loss functions with larger losses; and adjusting the current transport route of the inner lining paper based on the type of inner lining paper, and generating warning information based on the defective regions.

[0005] According to a second aspect of this disclosure, a deep learning-based image recognition device for inner lining paper is provided. The device includes an image acquisition module configured to acquire a target image corresponding to the inner lining paper; a model processing module configured to output a recognition result based on the target image using a trained Lite-MTNet model, the recognition result including the type of inner lining paper and defective regions. The backbone network of the Lite-MTNet model includes an LA-Conv Block module for generating a spatial-channel joint attention map based on a local attention mechanism, an MGF module for multi-scale feature fusion, and a CTFB module for generating general features shared between segmentation and classification branches. The total loss function of the Lite-MTNet model is a weighted sum of the segmentation loss function and the classification loss function, with higher weights assigned to loss functions with larger losses; and a control module configured to adjust the current transport route of the inner lining paper based on the type of inner lining paper and generate warning information based on defective regions.

[0006] According to a third aspect of this disclosure, an electronic device is provided, including one or more processors and a memory associated with the one or more processors, the memory being used to store program instructions that, when read and executed by the one or more processors, perform a method provided according to a first scheme.

[0007] According to a fourth aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements the method provided according to the first aspect.

[0008] The embodiments provided in this specification allow for the processing of target images using a trained Lite-MTNet model. This model extracts low-level and high-level semantic features of paper images using a parameter-sharing approach. Combined with a dual-task joint loss function, it can jointly optimize classification and segmentation tasks. This enables rapid and continuous adaptation to new scenes and defect types after deployment through semi-supervised learning, without requiring a large amount of labeled data, thereby achieving high-performance recognition and detection of inner lining paper. Attached Figure Description

[0009] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0010] Figure 1 A flowchart illustrating a deep learning-based image recognition method for inner lining paper according to some embodiments of this disclosure is shown.

[0011] Figure 2 A schematic diagram illustrating the principle process of the Lite-MTNet model during the training phase of some embodiments of this disclosure is shown;

[0012] Figure 3 A schematic diagram of the structure of a deep learning-based image recognition device for inner lining paper according to some embodiments of the present disclosure is shown;

[0013] Figure 4 A schematic block diagram of an electronic device according to some embodiments of the present disclosure is shown. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0015] The terms “comprising” and “having”, and any variations thereof, in this specification, claims, and the foregoing drawings are intended to cover a non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the steps or units listed, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. Depending on the context, the word “if” as it applies herein may be interpreted as “when”, “in response to determination”, or “in response to detection”.

[0016] Figure 1 A flowchart illustrating a deep learning-based image recognition method 200 for inner lining paper, representing some embodiments of this disclosure, is shown. Method 100 can be executed, for example, by a terminal, which may include, but is not limited to, mobile phones, tablets, desktop computers, servers, etc. Figure 1 As shown in box 102, method 100 can obtain the target image corresponding to the inner lining paper.

[0017] In this embodiment, an industrial camera with a resolution of at least 12 megapixels can be used to acquire a target image of the location of the inner lining paper. During image acquisition, a multi-angle LED combination light source (white light + blue light) can be used to create a stable lighting environment to suppress reflections and shadow interference, ensuring that the paper image and defects are clearly visible. The target image can be represented as:

[0018]

[0019] Where H and W represent the height and width of the image, respectively, and c=3 represents the three RGB channels.

[0020] In box 104, method 100 can output recognition results based on the target image using a trained Lite-MTNet model. The recognition results include the type of inner lining paper and the defect area. The backbone network of the Lite-MTNet model includes the LA-Conv Block module for generating a spatial-channel joint attention map based on the local attention mechanism, the MGF module for multi-scale feature fusion, and the CTFB module for generating general features shared between the segmentation branch and the classification branch. The total loss function of the Lite-MTNet model is a weighted sum of the segmentation loss function and the classification loss function, with the larger the loss function, the higher the weight assigned.

[0021] In this embodiment, the target image can be processed according to the Lite-MTNet model to obtain the recognition result output by the model. The recognition result includes the type of inner lining paper. and defect areas The prediction results are as follows. The Lite-MTNet model's backbone network (New Backbone) consists of three key modules: Local Attention Guided Lightweight Convolutional Block (LA-Conv Block), Multi-Scale Guided Fusion Module (MGF), and Cross-Task Feature Bridging Mechanism (CTFB). These three modules work together to improve the model's ability to represent complex patterns and micro-defects, making it particularly suitable for scenarios where large backgrounds and micro-defects coexist in industrial paper images.

[0022] Assuming the preprocessed target image is After the target image is processed by a shallow convolutional encoder, an initial feature map is obtained:

[0023]

[0024] in, This represents the initially extracted low-level features, which serve as input for subsequent modules.

[0025] In the LA-Conv Block module, the input features are denoted as This represents an intermediate feature map output from the previous layer or encoder. A local attention mechanism will be used to generate a spatial-channel joint attention map to adjust the convolutional response.

[0026]

[0027] in, It is a 1 x 1 convolution, used to compress or transform the channel dimensions; It is a Sigmoid activation function that generates normalized attention weights; The generated attention map is multiplied element-wise with the input features to guide the local response. The output attention map will be used to adjust the response intensity of subsequent depthwise separable convolutions, thereby improving texture and edge modeling capabilities.

[0028] In the MGF module, convolution extraction is performed on the input features at multiple scales to obtain branch features at multiple scales:

[0029]

[0030] in, This represents the semantic features under different receptive fields.

[0031] Next, the gating weights are calculated for each dimensional branch:

[0032]

[0033] in, It is a learnable convolutional kernel.

[0034] Finally, the multi-scale features are weighted and fused into the output:

[0035]

[0036] in, This represents the fusion gating weight, used to measure the importance of the i-th scale; The resulting multi-scale feature map contains rich contextual information.

[0037] The CTFB (Cross-Task Feature Bridge) module facilitates feature sharing between classification and segmentation tasks, and its input consists of features from two branches: This represents semantic features derived from the classification branch. This represents the spatial features derived from the segmented branches. After concatenation, a joint representation is formed:

[0038]

[0039] Next, the channel selector uses a 1x1 convolution and the Gumbel-Softmax method to select the channel key:

[0040]

[0041] in, This is a general feature representation used for sharing between two tasks.

[0042] Furthermore, to ensure the model's continued adaptability to new scenarios and defect types after deployment, a dual-task joint loss function can be used to jointly optimize the classification and segmentation tasks. The weighted sum of the segmentation and classification loss functions is used as the total loss function, with functions exhibiting larger losses receiving higher weights. This automatically allocates optimization focus based on the magnitude of the segmentation and classification losses in each round. The specific segmentation and classification loss functions can be set according to actual needs; for example, they can be set to the cross-entropy loss function.

[0043] By leveraging the lightweight attention network structure and loss function of the Lite-MTNet model, a joint optimization strategy can be employed to fully utilize limited labeled data and combine it with unsupervised image data to achieve an image recognition process that balances accuracy and efficiency. This significantly improves the automation and intelligence level of the packaging process and demonstrates high accuracy in recognizing different types of inner lining paper.

[0044] In box 106, method 100 can adjust the current transport route of the inner liner paper based on the type of inner liner paper and generate warning information based on defective areas.

[0045] In this embodiment, based on the identification results, the production line corresponding to the type of inner liner paper can be determined, and the current conveyor route can be adjusted so that the inner liner paper can be sent to the corresponding production line according to its type. Simultaneously, if a defective area exists, a warning message can be generated based on the defective area and sent to the relevant production line staff. This allows staff to know which part of the paper has a defect, enabling precise removal of the contaminated area or removal of the entire paper as a defective product, thus ensuring production quality.

[0046] In one possible implementation, acquiring the target image corresponding to the inner lining paper includes:

[0047] The original images of the inner lining paper from different perspectives are obtained, and the original images are registered based on the homography matrix to obtain the target image.

[0048] In this embodiment, to achieve complete image coverage and preservation of edge details, original images of the inner lining paper from different perspectives can be acquired, and then processed using a homography matrix. The images are registered and assembled to map the points in each original image to corresponding points on the same image plane, thus achieving seamless stitching of the original images to generate a complete paper image, i.e., the target image.

[0049] In one possible implementation, after acquiring the target image corresponding to the inner lining paper, the method further includes:

[0050] The target image is preprocessed, including normalization, local contrast enhancement, edge response extraction, and resolution scaling.

[0051] In this embodiment, in order to enhance the target image The availability and feature expressiveness of the target image provide a unified and enhanced input for the Lite-MTNet model, and it can also preprocess the target image. First, the target image can be normalized and standardized, mapping pixel values ​​to [0, 1] or a standard normal distribution, improving training stability. The normalization process is as follows:

[0052]

[0053] in , Let represent the mean and standard deviation of the c-th channel, respectively.

[0054] Subsequently, CLAHE (Contrast Limited Adaptive Histogram Equalization) can be used to enhance local contrast, highlighting subtle defects and texture details on the paper. At the same time, bilateral filtering or nonlocal mean filtering (NLM) can be used to reduce noise interference and preserve structural edges.

[0055] Furthermore, edge responses E(x,y) can be extracted using Sobel or Laplacian operators to enhance the clarity of segmented target contours. To adapt to the model input size, all target images can be uniformly scaled to a fixed resolution. Padding is used to ensure that the aspect ratio is not distorted.

[0056] In one possible implementation, in the Lite-MTNet model, the input data is processed sequentially through the LA-Conv Block module, the MGF module, the segmentation branch / classification branch, and the CTFB module. The resulting feature maps are then fed into the classification head and the segmentation head, respectively, to generate recognition results.

[0057] In this embodiment, after three stages of feature enhancement from LA-Conv to MGF to CTFB, the output unified semantic feature map is as follows:

[0058]

[0059] The feature map is fed into a classification head (Multi-layer Perceptron / GAP + FC), which obtains image-level labels through global average pooling and fully connected layers, and a segmentation head (depth-separable convolution + upsampling module), which obtains a pixel-by-pixel prediction mask through a lightweight decoder, and finally obtains the recognition result.

[0060] In one possible implementation, the segmentation loss function of the segmentation branch is a cross-entropy loss based on the semantic category of the pixel location, and the classification loss function of the classification branch is a cross-entropy loss based on the image category.

[0061] In this embodiment, the segmentation branch uses pixel-wise multi-class cross-entropy loss as a supervision signal to measure the difference between the model's prediction and the true segmentation label. The specific loss function is:

[0062]

[0063] Where H and W represent the height and width of the input image, used to control the total number of pixels; K represents the number of semantic categories, which can be set to K=2, representing "normal region" and "defect region" respectively; This represents the true label at pixel position (i, j). If the pixel belongs to class k, it is 1; otherwise, it is 0. This represents the softmax probability that the model predicts at pixel position (i, j) that it belongs to the kth class. The natural logarithm function is used to amplify prediction bias. This loss term evaluates the class prediction for each pixel location in the image, encouraging the model to accurately locate small defect regions, thereby improving semantic segmentation accuracy.

[0064] In the classification branch, to achieve accurate identification of different types of inner lining paper (such as kraft paper, whiteboard paper, composite paper, etc.), a multi-class cross-entropy loss function is used as a supervision signal to optimize the model's classification performance in the paper type identification task. The specific loss function is as follows:

[0065]

[0066] Where K represents the total number of lining paper categories involved in the classification task (3 or more categories). The one-hot encoding of the image's true label is 1 if the image belongs to class k, and 0 otherwise. The probability that the model predicts an image belongs to the k-th class is obtained by normalization using the softmax function; The natural logarithm function is used to quantify the difference between predicted values ​​and true labels. This loss function can effectively guide the model to learn the macroscopic differences in texture structure, fiber arrangement, and printing patterns of different inner lining papers, thereby achieving fast and accurate paper type classification.

[0067] In one possible implementation, in the total loss function, the weight of the segmentation loss function is the ratio between the segmentation loss function and the sum of loss functions, the weight of the classification loss function is the ratio between the classification loss function and the sum of loss functions, and the sum of loss functions is the sum of the segmentation loss function and the classification loss function.

[0068] In this embodiment, the semantic segmentation task of the segmentation branch can be taken as the dominant task to guide the overall optimization process, and the total loss function can be:

[0069]

[0070] in, This is the total loss function, used to guide online fine-tuning; For semantic segmentation loss; For classification loss.

[0071] To achieve dynamic adaptation during task optimization, this application designs a dynamic weighting strategy based on loss normalization, enabling the optimization focus of the two tasks to automatically adjust according to the actual training situation. Specifically, the loss weights for classification and segmentation... and It will be dynamically updated with each training round t, always satisfying The calculation method is as follows:

[0072]

[0073] This mechanism can automatically allocate optimization focus based on the magnitude of the loss from segmentation and classification in each round. If a task currently has a high loss, it will be automatically given greater weight for correction.

[0074] In one possible implementation, the method further includes:

[0075] Based on historical data, a training set is determined, which includes target image samples and recognition result samples corresponding to the target image samples.

[0076] Based on target image samples, the initial model generates predicted recognition results; and

[0077] Using the identification result samples as supervision signals, the initial model is trained for at least one round to obtain the Lite-MTNet model.

[0078] In this embodiment, as Figure 2 As shown, Figure 2A schematic diagram 200 illustrating the principle process 200 of the Lite-MTNet model during the training phase of some embodiments of this disclosure is shown. Using historical acquisition data, target image samples 210-1 of the inner lining paper captured by the camera under different backgrounds during historical image acquisition can be obtained, along with recognition result samples 210-2 obtained after labeling the target image samples 210-1. The labels include segmentation labels and classification labels. The segmentation labels are in the form of binary mask images. , This indicates that the pixel belongs to a defective area, and vice versa for a normal paper area. The classification label corresponds to a one-hot encoded category vector. Where K is the total number of paper types. A training set can be constructed based on the target image sample 210-1 and the recognition result sample 210-2. The backbone network of the initial model used for training can be a lightweight CNN network such as MobileNetV3. During the training of the Lite-MTNet model 230 using the training set, the generator 220 of the model can generate a predicted recognition result 221 based on the target image sample 210-1. Based on the comparison between the target image sample 210-1 and the predicted recognition result 221, a generator loss 222 can be obtained. The generator loss 222 is used to determine the loss of the predicted recognition result 221, and this loss can be determined using the aforementioned loss function. The generator loss 222 can be a large value, and then it can be backpropagated to the generator 220 to guide the optimization of the generator 220's parameters, achieving one round of supervised training for the generator 220. This training process can be iterated round by round until the generator 220 can generate a more accurate predicted recognition result 221, that is, until the loss value calculated by the loss function is smaller. After training, the Lite-MTNet model 230 can output recognition results 240.

[0079] The supervised training process can employ a semi-supervised strategy. This involves first using a small number of labeled samples for supervised learning, and then using unlabeled samples combined with consistency regularization and pseudo-labels for auxiliary training. This allows for high-performance recognition and detection under limited labeling conditions. Furthermore, users can upload new manually labeled samples or automatically collected high-confidence samples through the interface to build incremental datasets and periodically perform fine-tuning training. This enables the model to continuously adapt to new paper types, printing styles, or defect patterns, possessing a long-lifecycle intelligent evolution capability.

[0080] Figure 3Schematic diagrams of a deep learning-based image recognition device 300 for inner lining paper, representing some embodiments of this disclosure, are shown. The various embodiments in this specification are described in a progressive manner; similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the device embodiments, since they are basically similar to the method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments. Figure 3 As shown, the device 300 includes an image acquisition module 301, configured to acquire a target image corresponding to the inner lining paper; a model processing module 302, configured to output a recognition result based on the target image using a trained Lite-MTNet model, the recognition result including the type of inner lining paper and defective regions, the backbone network of the Lite-MTNet model including an LA-Conv Block module for generating a spatial-channel joint attention map based on a local attention mechanism, an MGF module for multi-scale feature fusion, and a CTFB module for generating common features shared between the segmentation branch and the classification branch, the total loss function of the Lite-MTNet model is a weighted sum of the segmentation loss function and the classification loss function, with the larger the loss function, the higher the weight assigned; and a control module 303, configured to adjust the current conveying route of the inner lining paper based on the type of inner lining paper and generate warning information based on the defective regions.

[0081] In one possible implementation, the image acquisition module 301 is further configured to acquire original images of the inner lining paper from different perspectives, and perform image registration on each original image based on the homography matrix to obtain the target image.

[0082] In one possible implementation, the image acquisition module 301 is further configured to preprocess the target image, the preprocessing including normalization, local contrast enhancement, edge response extraction, and resolution scaling.

[0083] In one possible implementation, in the Lite-MTNet model, the input data is processed sequentially through the LA-Conv Block module, the MGF module, the segmentation branch / classification branch, and the CTFB module. The resulting feature maps are then fed into the classification head and the segmentation head, respectively, to generate recognition results.

[0084] In one possible implementation, the segmentation loss function of the segmentation branch is a cross-entropy loss based on the semantic category of the pixel location, and the classification loss function of the classification branch is a cross-entropy loss based on the image category.

[0085] In one possible implementation, in the total loss function, the weight of the segmentation loss function is the ratio between the segmentation loss function and the sum of loss functions, the weight of the classification loss function is the ratio between the classification loss function and the sum of loss functions, and the sum of loss functions is the sum of the segmentation loss function and the classification loss function.

[0086] In one possible implementation, the device further includes a training module configured to determine a training set based on historically acquired data, the training set including target image samples and recognition result samples corresponding to the target image samples; generate predicted recognition results from an initial model based on the target image samples; and perform at least one round of model training on the initial model using the recognition result samples as supervision signals to obtain a Lite-MTNet model.

[0087] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this specification is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., Digital Versatile Discs (DVDs)), or semiconductor media (e.g., Solid State Disks (SSDs)).

[0088] Figure 4 A block diagram of an electronic device 400 that can implement various embodiments of the present disclosure is shown. For example... Figure 4 As shown, the electronic device 400 includes a processor 410, a disk drive 420, an input / output interface 430, a network interface 440, and a memory 450. The processor 410, disk drive 420, input / output interface 430, network interface 440, and memory 450 can communicate with each other via a communication bus 460.

[0089] The processor 410 can be implemented using a general-purpose CPU, microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits to execute relevant programs and implement the technical solution provided in this application.

[0090] The memory 450 can be implemented in the form of ROM (Read Only Memory), RAM (Read Access Memory), static memory, dynamic storage devices, etc. The memory 450 can store the operating system 451 used to control the operation of the electronic device 400, and the basic input / output system (BIOS) 452 used to control the low-level operations of the electronic device 400. Additionally, it can store a web browser 453, a data storage management system 454, etc. In summary, when the technical solution provided in this application is implemented through software or firmware, the relevant program code is stored in the memory 450 and is called and executed by the processor 410.

[0091] Input / output interface 430 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices may include displays, speakers, vibrators, indicator lights, etc.

[0092] Network interface 440 is used to connect a communication module (not shown in the figure) to enable communication and interaction between the device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0093] Bus 460 includes a pathway for transmitting information between various components of the device, such as processor 410, disk drive 420, input / input interface 430, network interface 440, and memory 450.

[0094] It should be noted that although the above-described device only shows the processor 410, disk drive 420, input / output interface 430, network interface 440, memory 450, bus 460, etc., in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the method of this application, and does not necessarily include all the components shown in the figures.

[0095] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0096] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing. Furthermore, although operations are depicted in a specific order, this should be understood as requiring that such operations be performed in the specific order shown or in sequential order, or requiring that all illustrated operations be performed to achieve the desired result. In certain environments, multitasking and parallel processing may be advantageous. Similarly, while several specific implementation details are included in the foregoing discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations.

[0097] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A deep learning-based image recognition method for inner lining paper, characterized in that, The method includes: Obtain the target image corresponding to the inner lining paper; Based on the target image, the trained Lite-MTNet model outputs recognition results, including the type of inner lining paper and defective regions. The backbone network of the Lite-MTNet model includes an LA-Conv Block module for generating a spatial-channel joint attention map based on a local attention mechanism, an MGF module for multi-scale feature fusion, and a CTFB module for generating common features shared between the segmentation and classification branches. The total loss function of the Lite-MTNet model is a weighted sum of the segmentation loss function and the classification loss function; the larger the loss function, the higher its weight. The current conveying route of the inner liner paper is adjusted based on the type of inner liner paper, and warning information is generated based on the defective area.

2. The deep learning-based image recognition method for inner lining paper according to claim 1, characterized in that, The step of obtaining the target image corresponding to the inner lining paper includes: The original images of the inner lining paper from different perspectives are obtained, and the original images are registered based on the homography matrix to obtain the target image.

3. The deep learning-based image recognition method for inner lining paper according to claim 1, characterized in that, After obtaining the target image corresponding to the inner lining paper, the process further includes: The target image is preprocessed, including normalization, local contrast enhancement, edge response extraction, and resolution scaling.

4. The deep learning-based image recognition method for inner lining paper according to claim 1, characterized in that, In the Lite-MTNet model, the input data is processed sequentially through the LA-Conv Block module, MGF module, segmentation branch / classification branch and CTFB module. The resulting feature maps are then fed into the classification head and segmentation head to generate recognition results.

5. The deep learning-based image recognition method for inner lining paper according to claim 1, characterized in that, The segmentation loss function of the segmentation branch is a cross-entropy loss based on the semantic category of the pixel location, and the classification loss function of the classification branch is a cross-entropy loss based on the image category.

6. The deep learning-based image recognition method for inner lining paper according to claim 1, characterized in that, In the total loss function, the weight of the segmentation loss function is the ratio between the segmentation loss function and the sum of the loss functions, the weight of the classification loss function is the ratio between the classification loss function and the sum of the loss functions, and the sum of the loss functions is the sum of the segmentation loss function and the classification loss function.

7. The deep learning-based image recognition method for inner lining paper according to claim 1, characterized in that, The method further includes: Based on historical data, a training set is determined, which includes target image samples and recognition result samples corresponding to the target image samples; Based on the target image samples, the initial model generates predicted recognition results; and Using the identified sample as a supervision signal, the initial model is trained for at least one round to obtain the Lite-MTNet model.

8. A deep learning-based image recognition device for inner lining paper, characterized in that, The device includes: The image acquisition module is configured to acquire the target image corresponding to the inner lining paper. The model processing module is configured to output recognition results from a trained Lite-MTNet model based on the target image. The recognition results include the type of inner lining paper and defect areas. The backbone network of the Lite-MTNet model includes an LA-Conv Block module for generating a spatial-channel joint attention map based on a local attention mechanism, an MGF module for multi-scale feature fusion, and a CTFB module for generating common features shared between the segmentation and classification branches. The total loss function of the Lite-MTNet model is a weighted sum of the segmentation loss function and the classification loss function; the larger the loss function, the higher its weight. The control module is configured to adjust the current conveying route of the inner liner paper based on the type of inner liner paper, and generate warning information based on the defective area.

9. An electronic device, comprising: One or more processors, and A memory associated with the one or more processors, the memory being used to store program instructions that, when read and executed by the one or more processors, perform the steps of the deep learning-based image recognition method for inner lining paper according to any one of claims 1-7.

10. A computer program product, comprising a computer program that, when executed by a processor, implements a deep learning-based image recognition method for inner lining paper according to any one of claims 1-7.

Images