Special part identification abnormity identification method and device, equipment and storage medium

Abstract

The invention relates to the technical field of information processing, in particular to a special part identification abnormity identification method, device and equipment and a storage medium, by constructing a target detection model based on an RTMDet algorithm, the intelligent degree is improved, the detection speed and precision can be balanced, parameters of the target detection model are adjusted according to special part identification task requirements, and the identification accuracy of special parts is improved. Performing data enhancement processing on the labeled data set, training a target detection model by using the enhanced data set, dynamically adjusting sample weights in the training process of the target detection model by using a Focal Loss loss function, and optimizing the training process of the target detection model in combination with a cosine annealing learning rate attenuation strategy and a gradient cutting technology. And deploying the target detection model to a logistics sorting node to carry out real-time detection on a parcel image, and when the target detection model detects that the identifier of the special part is abnormal, recording an abnormal type and parcel position information, timely finding the parcel with the abnormal identifier of the special part and carrying out early warning, so that logistics management personnel can rapidly take corresponding measures.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information processing, and in particular to a special piece identification abnormality identification method, device, equipment and storage medium. BACKGROUND

[0002] In logistics transportation, accurately identifying the special piece identification on the package is crucial, and its absence, blur or non-standardization will make the transport personnel unable to correctly judge the nature of the package, and thus unable to take appropriate transportation, storage and emergency handling measures. At present, there are obvious deficiencies in the identification of special piece identification on the package, and the traditional manual detection method has been difficult to meet the efficient and safe transportation demand of modern logistics. The traditional method mainly relies on manual inspection, and it is difficult to efficiently and accurately identify the absence, blur or non-compliance of special piece identification, which makes the package with special piece identification abnormality easily flow into the transportation link, bringing great hidden dangers to transportation safety. SUMMARY

[0003] In order to overcome the deficiencies of the prior art, the purpose of the present application is to provide a special piece identification abnormality identification method, device, equipment and storage medium which uses a target detection model based on the RTMDet algorithm for abnormal identification, improves the intelligent degree, balances the detection speed and accuracy, discovers special piece identification abnormality package in time and gives an early warning, so that the logistics management personnel can take corresponding measures quickly.

[0004] The first aspect of the present application provides a special piece identification abnormality identification method, comprising: collecting historical special piece package image samples from the cameras of distribution centers and warehouse centers, using a labeling tool to label the special piece identification in the special piece package image samples to obtain a labeled data set; constructing a target detection model based on the RTMDet algorithm, adjusting the parameters of the target detection model according to the special piece identification identification task requirements; performing data enhancement processing on the labeled data set to obtain an enhanced data set, and training the target detection model using the enhanced data set; dynamically adjusting the sample weight in the training process of the target detection model using the Focal Loss loss function, and combining the cosine annealing learning rate decay strategy and the gradient clipping technology to optimize the training process of the target detection model; deploying the target detection model to a logistics sorting node for real-time detection of package images, and when the target detection model detects special piece identification abnormality, recording the abnormality type and package location information.

[0005] Optionally, in the first implementation manner of the first aspect, the collecting historical special piece package image samples from the cameras of the sorting center and the warehouse center, and labeling special piece identification in the special piece package image samples by using a labeling tool to obtain a labeled data set comprises: collecting historical package images photographed by the cameras of the sorting center and the warehouse center, and screening historical special piece package image samples containing special pieces from the historical package images; labeling a boundary box of a special piece identification area in the special piece package image samples by using a LabelImg or CVAT labeling tool to obtain labeled image samples; and dividing the labeled image samples into a training set, a verification set and a test set to form a labeled data set.

[0006] Optionally, in the second implementation manner of the first aspect, the constructing a target detection model based on an RTMDet algorithm, and adjusting parameters of the target detection model according to special piece identification recognition task requirements comprises: constructing a target detection model based on an RTMDet algorithm; obtaining a special piece identification recognition task requirement file, and parsing the special piece identification recognition task requirement file to obtain special piece identification sizes, a special piece identification aspect ratio and output categories; modifying a feature pyramid layer number of a neck part of the target detection model according to the special piece identification sizes; adjusting an anchor size ratio of the target detection model according to the special piece identification aspect ratio; and setting the output categories in a detection head of the target detection model, wherein the output categories include normal identification, missing identification, damaged identification and pasted identification errors.

[0007] Optionally, in the third implementation manner of the first aspect, the performing data enhancement processing on the labeled data set to obtain an enhanced data set, and training the target detection model by using the enhanced data set comprises: performing Gaussian noise injection and simulated occlusion processing on the labeled data set to obtain a first training set; performing mirror, rotation, scaling, translation and brightness adjustment transformation processing on the labeled data set by using AutoAugment to obtain a second training set; performing cutting on the labeled data set by using Cutout to obtain a third training set; mixing the labeled data set, the first training set, the second training set and the third training set by using mixup to obtain an enhanced data set; and training the target detection model by using the enhanced data set.

[0008] Optionally, in a fourth implementation form of the first aspect of the present application, the method further comprises: dynamically adjusting the sample weight in the training process of the target detection model by using the Focal Loss loss function; optimizing the training process of the target detection model by combining a cosine annealing learning rate decay strategy and a gradient clipping technique; and terminating the training of the target detection model by using an early stopping mechanism.

[0009] Optionally, in a fifth implementation form of the first aspect of the present application, the method further comprises: deploying the target detection model to a logistics sorting node for real-time detection of parcel images; when the target detection model detects a special piece identification abnormality, identifying an abnormal type of the special piece identification and determining a specific location of a current special piece parcel; and recording the abnormal type and the specific location to obtain abnormal type and parcel location information.

[0010] Optionally, in a sixth implementation form of the first aspect of the present application, the method further comprises: after deploying the target detection model to a logistics sorting node for real-time detection of parcel images and when the target detection model detects a special piece identification abnormality, recording abnormal type and parcel location information, the method further comprises: obtaining high-definition image information of an abnormal parcel; merging the high-definition image information and the abnormal type and parcel location information to generate early warning information; and pushing the early warning information to a sorting line control terminal, so that the sorting line control terminal generates and displays a visual early warning page based on the early warning information.

[0011] The second aspect of the present application provides a special piece identification abnormality recognition device, comprising: a collection and labeling module, configured to collect historical special piece parcel image samples from cameras of distribution centers and warehouse centers, and label special piece identification in the special piece parcel image samples using a labeling tool to obtain a labeled data set; a construction and adjustment module, configured to construct a target detection model based on an RTMDet algorithm, and adjust parameters of the target detection model according to special piece identification recognition task requirements; an enhanced training module, configured to perform data enhancement processing on the labeled data set to obtain an enhanced data set, and train the target detection model using the enhanced data set; an adjustment and optimization module, configured to dynamically adjust sample weights in a training process of the target detection model using a Focal Loss loss function, and optimize the training process of the target detection model in combination with a cosine annealing learning rate decay strategy and a gradient clipping technique; and a deployment recording module, configured to deploy the target detection model to a logistics sorting node for real-time parcel image detection, and record abnormal type and parcel location information when the target detection model detects special piece identification abnormalities.

[0012] Optionally, in the first implementation manner of the second aspect of the present application, the collection and labeling module comprises: a collection and screening unit, configured to collect historical parcel images captured by cameras of distribution centers and warehouse centers, and screen historical special piece parcel image samples containing special pieces from the historical parcel images; a labeling unit, configured to perform bounding box labeling on special piece identification regions in the special piece parcel image samples using LabelImg or CVAT labeling tools to obtain labeled image samples; and a division unit, configured to divide the labeled image samples into a training set, a validation set and a test set to form a labeled data set.

[0013] Optionally, in the second implementation manner of the second aspect of the present application, the construction and adjustment module comprises: a construction unit, configured to construct a target detection model based on an RTMDet algorithm; an acquisition and analysis unit, configured to acquire a special piece identification recognition task requirement file, analyze the special piece identification recognition task requirement file to obtain special piece identification dimensions, special piece identification aspect ratios and output categories; a modification unit, configured to modify a feature pyramid layer number of a neck part of the target detection model according to the special piece identification dimensions; a first adjustment unit, configured to adjust anchor size ratios of the target detection model according to the special piece identification aspect ratios; and a setting unit, configured to set the output categories in detection heads of the target detection model, wherein the output categories include normal identification, missing identification, damaged identification and identification pasting errors.

[0014] Optionally, in a third implementation of the second aspect of the present invention, the augmentation training module includes: an occlusion injection unit, used to inject Gaussian noise and simulate occlusion processing into the labeled dataset to obtain a first training set; a transformation processing unit, used to perform mirroring, rotation, scaling, translation, and brightness adjustment transformation processing on the labeled dataset using AutoAugment to obtain a second training set; a cropping unit, used to crop the labeled dataset using Cutout to obtain a third training set; a mixing unit, used to mix the labeled dataset, the first training set, the second training set, and the third training set using mixup to obtain an augmented dataset; and a training unit, used to train the object detection model using the augmented dataset.

[0015] Optionally, in a fourth implementation of the second aspect of the present invention, the adjustment and optimization module includes: a second adjustment unit, used to dynamically adjust sample weights during the training process of the object detection model using the Focal Loss loss function; an optimization unit, used to optimize the training process of the object detection model by combining a cosine annealing learning rate decay strategy and gradient pruning technique; and a termination unit, used to terminate the training of the object detection model using an early stopping mechanism.

[0016] Optionally, in a fifth implementation of the second aspect of the present invention, the deployment recording module includes: a deployment unit, used to deploy the target detection model to a logistics sorting node for real-time detection of package images; an identification and determination unit, used to identify the anomaly type of the special item identifier and determine the specific location of the current special item package when the target detection model detects an anomaly in the special item identifier; and a recording unit, used to record the anomaly type and the specific location to obtain anomaly type and package location information.

[0017] Optionally, in a sixth implementation of the second aspect of the present invention, the method further includes: an acquisition module for acquiring high-definition image information of the abnormal package; a merging and generation module for merging the high-definition image information and the abnormality type and package location information to generate early warning information; and a push module for pushing the early warning information to the sorting line control terminal, so that the sorting line control terminal generates and displays a visual early warning page based on the early warning information.

[0018] A third aspect of the present invention provides a special part identification anomaly recognition device, the special part identification anomaly recognition device comprising: a memory and at least one processor, the memory storing instructions; at least one processor calling the instructions in the memory to cause the special part identification anomaly recognition device to perform each step of the special part identification anomaly recognition method described in any of the preceding claims.

[0019] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the special component identification anomaly identification method described in any of the preceding claims.

[0020] In the technical solution of this invention, a target detection model based on the RTMDet algorithm is constructed. The parameters of the target detection model are adjusted according to the requirements of the special item identification task. Data augmentation processing is performed on the labeled dataset. The target detection model is trained using the augmented dataset. The sample weights are dynamically adjusted during the training process of the target detection model using the Focal Loss loss function. The training process of the target detection model is optimized by combining the cosine annealing learning rate decay strategy and gradient clipping technology. The target detection model is deployed to the logistics sorting node for real-time detection of package images. The target detection model based on the RTMDet algorithm is used for anomaly identification, which improves the level of intelligence and can balance detection speed and accuracy. When the target detection model detects anomalies in special item identification, it records the anomaly type and package location information, promptly discovers packages with abnormal special item identification and issues warnings, enabling logistics managers to take corresponding measures quickly. Attached Figure Description

[0021] Figure 1 This is a first flowchart of a special component identification anomaly recognition method provided in an embodiment of the present invention;

[0022] Figure 2 This is a second flowchart of the special component identification anomaly recognition method provided in the embodiments of the present invention;

[0023] Figure 3 This is a third flowchart of the special component identification anomaly identification method provided in the embodiments of the present invention;

[0024] Figure 4 This is a fourth flowchart of the special component identification anomaly identification method provided in the embodiments of the present invention;

[0025] Figure 5 This is a schematic diagram of a special component identification anomaly recognition device provided in an embodiment of the present invention;

[0026] Figure 6 This is another structural schematic diagram of the special component identification anomaly recognition device provided in an embodiment of the present invention;

[0027] Figure 7 This is a schematic diagram of the structure of a special component identification anomaly recognition device provided in an embodiment of the present invention. Detailed Implementation

[0028] This invention provides a method, apparatus, device, and storage medium for identifying abnormal special item markings. It utilizes a target detection model based on the RTMDet algorithm to identify abnormal markings, thereby improving the level of intelligence, balancing detection speed and accuracy, and promptly detecting and issuing warnings for packages with abnormal special item markings, enabling logistics managers to take corresponding measures quickly.

[0029] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0030] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the special component identification anomaly recognition method in this invention includes:

[0031] 101. Collect historical special package image samples from cameras in distribution centers and warehousing centers, and use annotation tools to annotate the special item identifiers in the special package image samples to obtain an annotated dataset;

[0032] In this embodiment, the system connects to the video surveillance system databases of the distribution center and the warehousing center, retrieves historical video data, and uses image recognition algorithms to initially filter out package images containing special items (such as fragile items, dangerous goods, valuables, cold chain goods, etc. with specific markings). Then, using professional annotation tools such as LabelImg or CVAT, the special item marking areas are accurately located on these images, marked with rectangular bounding boxes, and their corresponding categories (normal markings or abnormal types to be identified later) are specified. The annotation process must ensure that the bounding boxes completely cover the markings and that the categories are accurate. Finally, all the annotated images are divided into training sets, validation sets, and test sets according to a preset ratio to form a structured annotated dataset.

[0033] 102. Construct a target detection model based on the RTMDet algorithm, and adjust the parameters of the target detection model according to the requirements of the special part identification task;

[0034] In this embodiment, the RTMDet algorithm is selected as the basic architecture. The special part identification task requirement file provided by the logistics management party is obtained, and after parsing, the key parameters are identified, including the physical size range of typical special part identifications, the common aspect ratios of the identifications, and the output categories to be detected. Based on the parsed identification size, the number of layers in the feature pyramid (FPN) of the model's neck part is adjusted to ensure that the model can effectively capture the identification features of different scales (especially smaller sizes). Based on the aspect ratio of the identification, the size ratio of the anchor box preset by the model is optimized to make it more in line with the actual identification shape and improve the quality of candidate box generation. The number of output categories (e.g., 4 categories: normal, missing, damaged, and pasted incorrectly) is set in the detection head of the model.

[0035] 103. Perform data augmentation on the labeled dataset to obtain the augmented dataset, and use the augmented dataset to train the target detection model;

[0036] In this embodiment, firstly, Gaussian noise is injected into the original dataset and simulated occlusion blocks (such as simulated labels being partially covered or stains) are randomly added to obtain the first training set. Secondly, the AutoAugment strategy is applied to automatically search for the optimal combination of enhancement strategies (including mirror flipping, random rotation, scaling at different scales, translation, and brightness / contrast adjustment) to generate the transformed second training set. Thirdly, the Cutout technique is used to randomly crop rectangular regions in the image to generate the third training set. Finally, the mixup technique is used to mix the original dataset, the first training set, the second training set, and the third training set in a random ratio to generate a diverse and rich augmented dataset that covers various potential interference scenarios. This augmented dataset is then used to train the RTMDet model with adjusted parameters.

[0037] 104. Utilize the Focal Loss function to dynamically adjust sample weights during the training process of the object detection model, and combine cosine annealing learning rate decay strategy and gradient clipping technique to optimize the training process of the object detection model.

[0038] In this embodiment, Focal Loss is used as the loss function to dynamically adjust sample weights, automatically reducing the loss contribution of a large number of easily classified negative samples (background or normal labels). This allows the model training to focus more on difficult-to-classify samples (such as small objects, blurry, partially occluded abnormal labels) and positive samples (the abnormal labels themselves), thereby effectively alleviating class imbalance and improving the detection accuracy of abnormal labels. At the same time, a cosine annealing learning rate decay strategy is adopted to keep the learning rate high in the early stage of training so that it decreases rapidly, and then smoothly decreases it to near zero according to the cosine function in the later stage. This helps the model escape local optima and converge to a better solution. In addition, gradient pruning technology is applied to limit the maximum value of the gradient during backpropagation, prevent gradient explosion, and ensure the stability of the training process. An early stopping mechanism is also implemented during training to continuously monitor the performance indicators on the validation set. When the indicators no longer improve within several consecutive epochs, training is automatically terminated to avoid overfitting and to save the model weights with the best performance on the validation set.

[0039] 105. Deploy the target detection model to the logistics sorting node for real-time package image detection. When the target detection model detects an abnormal special item identifier, record the abnormality type and package location information.

[0040] In this embodiment, the trained and optimized RTMDet model is converted into a deployment-friendly format and integrated into the image processing system of the logistics sorting node. This system captures images of packages on the conveyor belt in real time using industrial cameras installed above or to the side of the sorting line. The deployed model performs inference on each frame of the image, detects special item identification areas on the packages, and determines their status category (normal or specific abnormal type). When the model detects abnormalities such as missing, damaged, or incorrectly pasted identifications, the system accurately identifies the specific type of the abnormality (e.g., "missing identification"). At the same time, combined with the package tracking system (e.g., barcode scanning, RFID, or image-based location association), the system determines the specific location of the abnormal special item package on the sorting line (e.g., which sorting port it is located at, and the coordinates of the conveyor belt). The system records the detected abnormality type (e.g., "damaged identification") and package location information (e.g., "sorting port 3, coordinates X:1200, Y:300") in real time to a log database or message queue.

[0041] In this embodiment of the invention, a customized target detection model based on the RTMDet algorithm is constructed, and the model parameters are precisely adjusted according to the requirements of the special part identification task. A robust augmented dataset is constructed by combining multi-strategy data augmentation for model training. Focal Loss is used to dynamically focus on difficult samples, and cosine annealing learning rate and gradient clipping are combined to optimize the training process. This effectively solves the problems of low efficiency, easy omissions, and difficulty in dealing with complex environments in traditional manual detection. Finally, the optimized model is deployed to the sorting node for real-time detection and recording of anomaly types and location information. This significantly improves the intelligent identification capability and real-time early warning efficiency of abnormal states such as missing, damaged, and mislabeled special part identification, providing strong technical support for safe logistics transportation.

[0042] Please see Figure 2 The second embodiment of the special component identification anomaly recognition method in this invention includes:

[0043] 201. Collect historical package images captured by cameras in distribution centers and warehousing centers, and filter historical special package images containing special items from the historical package images;

[0044] In this embodiment, the system batch exports historical package image data from the video surveillance storage servers of the distribution center and the warehousing center. Using a pre-trained package classification model or a rule-based filter (such as images containing specific color areas or shape templates), the system automatically identifies and extracts package images with special item labels (such as fragile, upside down, dangerous goods, biological products, cold chain, etc.) from the massive historical images, forming a preliminary sample set of historical special item package images.

[0045] 202. Use LabelImg or CVAT annotation tools to annotate the bounding box of the special part identification area in the special part package image sample to obtain the annotated image sample;

[0046] In this embodiment, graphical annotation tools such as LabelImg or CVAT are used to carefully draw rectangular boxes on the selected special item package image samples to select the complete special item label (such as the "glass" icon + "fragile" text). At the same time, a corresponding category label is selected for each bounding box, which includes not only the "special item label" itself, but more importantly, its status: normal label (clear and complete), label missing (the label is missing where it should be), label damaged (the label is torn, soiled, or partially missing), label incorrectly pasted (such as the "upward" arrow pasted upside down, or the dangerous goods label being pasted in the wrong category). After completing the annotation of all samples, an annotated image sample with precise location and status category information is obtained.

[0047] 203. Divide the labeled image samples into training set, validation set and test set to form a labeled dataset;

[0048] In this embodiment, all labeled image samples are randomly shuffled and then divided according to a preset ratio (e.g., 70% for training, 15% for validation, and 15% for testing). The training set is used to directly train the model parameters, the validation set is used to monitor model performance, adjust hyperparameters (such as learning rate), and determine early stopping time during training, and the test set is used to finally evaluate the model's generalization performance on completely unknown data. These three parts together constitute the labeled dataset used for model training and evaluation. The division process ensures that the distribution ratio of each category (normal, missing, broken, and incorrect) in the three subsets is basically consistent.

[0049] 204. Construct a target detection model based on the RTMDet algorithm;

[0050] In this embodiment, RTMDet is selected as the basic object detection framework. RTMDet includes an efficient backbone network, a feature pyramid network (Neck, such as PAFPN), and a detection head. Based on the hardware deployment environment (such as edge computing devices, GPU servers) and real-time requirements (such as sorting line speed), an RTMDet pre-trained model of appropriate size is selected as the starting point for construction.

[0051] 205. Obtain the special parts identification task requirement file, parse the special parts identification task requirement file, and obtain the special parts identification size, the aspect ratio of the special parts identification, and the output category;

[0052] In this embodiment, a detailed special parts identification task requirement document is obtained from the logistics business department or system requirement document. This document clearly defines the targets that the system needs to identify, including common special parts identification types (fragile, upside down, dangerous goods classification labels, cold chain labels, etc.), the typical physical size range of these labels (e.g., small labels 5cm x 5cm, large labels 10cm x 15cm), common aspect ratios of the labels (e.g., 1:1 square, 2:1 or 3:1 rectangle), and the specific categories that the system needs to output (normal labels, missing labels, damaged labels, incorrectly pasted labels). The parsing program reads this document and extracts three key parameters: special parts label size, special parts label aspect ratio, and output category list.

[0053] 206. Modify the number of feature pyramid layers in the neck part of the target detection model according to the special part identification size;

[0054] In this embodiment, the neck part of the model (usually a Feature Pyramid Network (FPN) or a PAFPN) is responsible for fusing feature maps from different layers of the backbone network to detect targets of different scales. The parsed label size information (especially the minimum size) determines the minimum target scale that the model needs to detect. If the labels are generally small, the number of layers in the feature pyramid needs to be increased or the feature fusion strategy needs to be adjusted to ensure that the high-resolution feature maps in the shallow layers (containing details of small targets) can be effectively utilized. If the labels are generally large, the pyramid structure can be simplified appropriately to improve speed. According to the actual label size distribution, the configuration of the neck part (such as the number of layers in the PAFPN and the level of the input feature map) can be adjusted to optimize the model's ability to detect small-sized labels.

[0055] 207. Adjust the anchor size ratio of the target detection model according to the aspect ratio of the special part markings;

[0056] In this embodiment, the object detection model presets a set of anchor boxes of different sizes and aspect ratios as references for candidate regions. The typical aspect ratios of the labels obtained by parsing (such as 1:1 for squares, 2:1 and 3:1 for rectangles) are the key basis for adjusting the anchors. The generation parameters of the anchors in the model configuration file are adjusted so that the aspect ratio distribution generated is closer to the shape of the actual special label (for example, increasing the proportion of 1:1 and 2:1 anchors and reducing uncommon proportions such as 4:1). This makes the initial candidate boxes (Proposals / Priors) generated by the model fit the shape of the real label more closely, reduces invalid candidates, and improves the accuracy and efficiency of subsequent classification and regression.

[0057] 208. Set the output categories in the detection head of the target detection model. The output categories include normal labels, missing labels, damaged labels, and incorrectly pasted labels.

[0058] In this embodiment, the model's head is responsible for classifying candidate regions and performing bounding box regression. Based on the output category list determined by the requirements analysis (normal label, missing label, damaged label, incorrect label pasting), the number of output channels in the classification branch of the head is modified. For example, the original model may output 80 categories (COCO dataset), but now it needs to be modified to output 4 categories. At the same time, it is ensured that the bounding box regression branch remains unchanged or is finely adjusted according to requirements.

[0059] In this embodiment of the invention, historical images of special items are systematically collected and filtered from cameras at logistics nodes. Professional tools are used for precise bounding box annotation and state classification to construct a structured, high-quality labeled dataset. Based on this, an RTMDet model is built and the business requirements document is deeply analyzed. The core structural parameters of the model are optimized according to the actual physical characteristics of the identifiers, and specific anomaly categories are accurately set in the detection head. This process ensures that the target detection model is closely aligned with the specific business scenarios and physical constraints of special item identification from the underlying design, laying a solid foundation for high-precision, high-efficiency training and application of the model. It effectively solves the problems of poor model generalization and difficulty in adapting to specific identifier characteristics.

[0060] Please see Figure 3 The third embodiment of the special component identification anomaly recognition method in this invention includes:

[0061] 301. Gaussian noise injection and simulated occlusion processing are applied to the labeled dataset to obtain the first training set;

[0062] In this embodiment, to address the noise (camera sensor noise, uneven lighting) and occlusion (packet stacking, conveyor belt mechanical parts obstructing, labels partially obscured by other items) problems that may exist in images in the actual sorting environment, the original labeled dataset (mainly the training set) is processed. Gaussian noise injection is applied, superimposing random noise conforming to a Gaussian distribution onto the image pixel values. Simulated occlusion processing is applied, placing gray or mosaic blocks at random positions on the image, or using a black mask of random shape to cover part of the image area (which may cover the label area). The processed image and its corresponding annotation information (boundary boxes and category labels need to be adjusted according to the occlusion situation; if the label is completely occluded, it may be removed; if it is partially occluded, it is retained) constitute the first training set.

[0063] 302. Use AutoAugment to perform mirroring, rotation, scaling, translation, and brightness adjustment transformations on the labeled dataset to obtain the second training set;

[0064] In this embodiment, the AutoAugment strategy is used, which is a method that automatically searches for the best combination of data augmentation strategies suitable for a specific dataset. It automatically applies a series of random geometric transformations (horizontal / vertical mirroring, rotation within a certain angle range, scaling at different ratios, translation in the horizontal and vertical directions) and photometric transformations (brightness enhancement / decrease, contrast adjustment) to the input labeled dataset (training set). The original image is processed by applying these automatically generated transformation strategies to obtain a second training set with significantly increased geometric and photometric diversity.

[0065] 303. Use Cutout to crop the labeled dataset to obtain the third training set;

[0066] In this embodiment, the Cutout enhancement technique is applied. This technique randomly selects one or more square regions on the images of the original labeled dataset (training set) and sets the pixel values ​​in these regions to zero (or fills them with mean / random noise). Cutout simulates the scenario where some information is lost or severely occluded during image acquisition. By forcing the model to make judgments without relying on local patches of the image (which may just contain the key parts of the label), it encourages the model to learn to use more global and robust features, thereby improving the ability to recognize cases where the label is missing or severely occluded. The processed images form the third training set.

[0067] 304. Use mixup to mix the labeled dataset, the first training set, the second training set, and the third training set to obtain the augmented dataset;

[0068] In this embodiment, a mixup augmentation strategy is adopted. Instead of using a single image sample directly, mixup randomly selects two images (which may come from the original labeled dataset, the first training set, the second training set, or the third training set) and performs linear weighted mixing at a random ratio. Correspondingly, the labels (boundary box coordinates and class) of the two images are also mixed at the same ratio. This creates a large number of "virtual" samples between the original samples, especially samples on the class boundaries. Finally, the set of all samples processed by mixup constitutes the final augmented dataset.

[0069] 305. Train the object detection model using augmented datasets;

[0070] In this embodiment, the comprehensive augmented dataset obtained in step 304 is used to train the RTMDet object detection model whose parameters have been constructed and adjusted in step 102.

[0071] 306. Dynamically adjust sample weights during the training process of the object detection model using the Focal Loss function;

[0072] In this embodiment, Focal Loss is used as the loss function for the classification task during the backpropagation process of model training. The core idea of ​​Focal Loss is to dynamically adjust the loss weight of samples according to the difficulty of classification. For a large number of easily classified background samples or clear and normal label samples, Focal Loss will significantly reduce their loss contribution. However, for samples that are difficult to classify (such as blurry, small targets, partially occluded abnormal labels, or foreground targets), even if their prediction probability is low, their loss contribution will be relatively high. By adjusting the parameters to reduce the weight of easily classified samples, the focus of model training is concentrated on those difficult and important samples, which significantly improves the detection ability of rare abnormal samples.

[0073] 307. Optimize the training process of the object detection model by combining the cosine annealing learning rate decay strategy and gradient clipping technique;

[0074] In this embodiment, a cosine annealing learning rate decay strategy is adopted, setting an initial learning rate and a minimum learning rate. The learning rate decays according to the shape of a cosine function during training. This strategy maintains a high learning rate and decreases rapidly in the early stage of training, and then gradually decreases to the minimum value in the later stage. This helps the model escape local minima, find better solutions, and converge stably. At the same time, gradient clipping technology is applied. After each backpropagation calculation of the gradient, the norm of the gradient is checked. If the gradient norm exceeds a preset threshold, the entire gradient vector is scaled proportionally to make its norm equal to the threshold. This effectively prevents training instability or divergence caused by gradient explosion during training.

[0075] 308. Employ an early stopping mechanism to terminate the training of the target detection model;

[0076] In this embodiment, an early stopping mechanism is implemented to prevent overfitting. During training, after each epoch, the model's performance metrics are evaluated using an independent validation set, and changes in these metrics are monitored. A patience value is defined. If the validation set performance metrics do not improve (or even decrease) within consecutive patience epochs, it is determined that the model performance has stopped improving and may have begun to overfit. At this point, training is automatically terminated. The model that is finally saved is the model weights of the epoch with the best performance on the validation set, rather than the weights of the last epoch. This ensures that the deployed model has the best generalization ability.

[0077] In this embodiment of the invention, a comprehensive augmented dataset covering various actual sorting interference scenarios is constructed by implementing multi-level and multi-type data augmentation strategies. In the model training stage, the Focal Loss loss function is innovatively applied to dynamically adjust the sample weights, effectively alleviating the class imbalance problem and focusing on difficult-to-identify samples. Combined with the cosine annealing learning rate strategy, better convergence is achieved. Gradient pruning technology is used to ensure training stability, and an early stopping mechanism is introduced to prevent overfitting. These optimization methods work together to significantly improve the accuracy and stability of the trained RTMDet model in the complex, variable, and sample-imbalanced actual logistics environment for identifying special item identifiers and their abnormal states (especially rare abnormal samples).

[0078] Please see Figure 4 The fourth embodiment of the special component identification anomaly recognition method in this invention includes:

[0079] 401. Deploy the target detection model to the logistics sorting node for real-time package image detection;

[0080] In this embodiment, the target detection model is deployed to the image acquisition and processing system of the logistics sorting node. By deploying industrial cameras at key locations on the sorting line (such as above the entrance and diversion port), images of moving packages on the conveyor belt are captured in real time (or near real time). The converted model continuously infers the incoming image stream, detects whether there are special item identifiers in the images, and identifies their status.

[0081] 402. When the target detection model detects an anomaly in the special part identification, it identifies the anomaly type of the special part identification and determines the specific location of the current special part package.

[0082] In this embodiment, after the target detection model infers for each frame of image, it outputs the detection results, including the coordinates of the detected bounding box (marker position), category labels (normal marker, missing marker, damaged marker, incorrect marker pasting), and confidence score. The system sets a confidence threshold (e.g., 0.7). When the model outputs a detection result for a certain type of anomaly (missing, damaged, incorrect pasting) and its confidence score exceeds the threshold, it is determined that a special item mark anomaly has been detected. At this time, the system records the specific type of the anomaly (e.g., "damaged marker"). At the same time, in order to accurately locate the problem package, the system needs to combine the package tracking information. The system uses the package tracking information to determine the specific location of the current special item package with the anomaly (e.g., "Package ID: SN123456, currently located 2 meters in front of sorting port 5").

[0083] 403. Record the anomaly type and its specific location to obtain anomaly type and package location information;

[0084] In this embodiment, the system combines the anomaly type (such as "damaged marker") identified in step 402 and the specific location (such as "package ID: SN123456, location: conveyor belt C line, coordinates Zone5, Marker120") into a structured record. This anomaly type and package location information record will be written to the system's event log database or message queue in real time, providing a data foundation for subsequent early warning, interception and manual review.

[0085] 404. Obtain high-resolution image information of the abnormal package;

[0086] In this embodiment, after an abnormal identifier is detected and a record is generated, the system triggers high-definition image acquisition. The system acquires and stores high-definition image information containing clearly visible abnormal identifiers (or missing identifier areas).

[0087] 405. Combine high-definition image information with anomaly type and package location information to generate early warning information;

[0088] In this embodiment, the system associates and merges the high-definition image information (as visual evidence) obtained in step 404 with the anomaly type and package location information (including anomaly type, package ID, and current location) generated in step 403, and packages them to generate a rich warning message.

[0089] 406. Push the warning information to the sorting line control terminal so that the sorting line control terminal can generate and display a visual warning page based on the warning information;

[0090] In this embodiment, the system pushes the generated early warning information to the sorting line control terminal (such as the sorting supervisor's monitoring computer, the on-site operator's tablet computer, or a large screen display) in real time via an internal network (such as MQTT protocol, WebSocket, or REST API). After receiving the early warning information, the application on the control terminal parses it and generates a visual early warning page based on the received information (abnormality type, package ID, location, and image). This page may mark the package location on the map in a conspicuous manner (such as a red border or flashing icon), clearly display the abnormality type (text + icon), and show a high-definition image of the abnormal package for manual confirmation. This page is displayed to logistics managers or on-site operators, enabling them to intuitively and quickly understand the abnormal situation and its details, so as to take corresponding measures quickly (such as manually sorting and removing the package from the line, suspending the sorting line, and notifying the quality inspector for review).

[0091] In this embodiment of the invention, the optimized RTMDet model is efficiently deployed to the logistics sorting node to achieve real-time detection of package images on the conveyor belt. When the model identifies an anomaly, the system accurately determines the anomaly type and, in conjunction with package tracking technology, precisely locates the problematic package. Subsequently, the system automatically acquires a high-definition image of the abnormal package as evidence, merges it with the anomaly type and package location information to generate structured early warning information, and pushes it to the sorting line control terminal in real time. The terminal generates and displays a visual early warning page that integrates anomaly details, location map, and high-definition images. This closed-loop process achieves fully automated processing from "intelligent anomaly identification" to "precise location positioning" and then to "intuitive information early warning," enabling logistics managers to obtain key information and take intervention measures immediately. This effectively solves the problems of delayed response and poor information transmission in traditional methods, and greatly improves the efficiency and safety of handling special item anomalies in the logistics sorting process.

[0092] The above describes the method for identifying abnormal special part markings in the embodiments of the present invention. The following describes the device for identifying abnormal special part markings in the embodiments of the present invention. Please refer to [link / reference]. Figure 5 One embodiment of the special component identification anomaly recognition device of the present invention includes:

[0093] The collection and annotation module 501 is used to collect historical special item package image samples from cameras in the distribution center and the warehousing center, and use annotation tools to annotate the special item identifiers in the special item package image samples to obtain an annotated dataset;

[0094] The adjustment module 502 is used to build a target detection model based on the RTMDet algorithm and adjust the parameters of the target detection model according to the requirements of the special part identification task.

[0095] The augmentation training module 503 is used to perform data augmentation on the labeled dataset to obtain an augmented dataset, and to train the target detection model using the augmented dataset.

[0096] The adjustment and optimization module 504 is used to dynamically adjust the sample weights during the training process of the object detection model using the Focal Loss loss function, and to optimize the training process of the object detection model by combining the cosine annealing learning rate decay strategy and gradient clipping technique.

[0097] The deployment recording module 505 is used to deploy the target detection model to the logistics sorting node for real-time detection of package images. When the target detection model detects an abnormality in the special item identification, it records the abnormality type and package location information.

[0098] In this embodiment, a target detection model based on the RTMDet algorithm is constructed. The parameters of the target detection model are adjusted according to the requirements of the special item identification task. Data augmentation processing is performed on the labeled dataset. The target detection model is trained using the augmented dataset. The Focal Loss loss function is used to dynamically adjust the sample weights during the training process of the target detection model. The training process of the target detection model is optimized by combining the cosine annealing learning rate decay strategy and gradient clipping technology. The target detection model is deployed to the logistics sorting node for real-time detection of package images. The target detection model based on the RTMDet algorithm is used for anomaly identification to improve the level of intelligence and balance detection speed and accuracy. When the target detection model detects anomalies in special item identification, it records the anomaly type and package location information, promptly discovers packages with abnormal special item identification and issues warnings, enabling logistics managers to take corresponding measures quickly.

[0099] Please see Figure 6 Another embodiment of the special component identification anomaly recognition device in this invention includes:

[0100] The collection and annotation module 501 is used to collect historical special item package image samples from cameras in the distribution center and the warehousing center, and use annotation tools to annotate the special item identifiers in the special item package image samples to obtain an annotated dataset;

[0101] The adjustment module 502 is used to build a target detection model based on the RTMDet algorithm and adjust the parameters of the target detection model according to the requirements of the special part identification task.

[0102] The augmentation training module 503 is used to perform data augmentation on the labeled dataset to obtain an augmented dataset, and to train the target detection model using the augmented dataset.

[0103] The adjustment and optimization module 504 is used to dynamically adjust the sample weights during the training process of the object detection model using the Focal Loss loss function, and to optimize the training process of the object detection model by combining the cosine annealing learning rate decay strategy and gradient clipping technique.

[0104] The deployment recording module 505 is used to deploy the target detection model to the logistics sorting node for real-time detection of package images. When the target detection model detects an abnormal special item identifier, it records the abnormality type and package location information.

[0105] In this embodiment, the collection and annotation module 501 includes: a collection and filtering unit 5011, used to collect historical package images captured by cameras in the distribution center and the warehousing center, and to filter historical special package image samples containing special items from the historical package images; an annotation unit 5012, used to annotate the special item identification area in the special item package image samples with a LabelImg or CVAT annotation tool to obtain annotated image samples; and a partitioning unit 5013, used to partition the annotated image samples into a training set, a validation set, and a test set to form an annotated dataset.

[0106] In this embodiment, the construction and adjustment module 502 includes: a construction unit 5021, used to construct a target detection model based on the RTMDet algorithm; an acquisition and parsing unit 5022, used to acquire a special part identification task requirement file, parse the special part identification task requirement file, and obtain the special part identification size, the aspect ratio of the special part identification, and the output category; a modification unit 5023, used to modify the number of feature pyramid layers in the neck part of the target detection model according to the special part identification size; a first adjustment unit 5024, used to adjust the anchor size ratio of the target detection model according to the aspect ratio of the special part identification; and a setting unit 5025, used to set the output category in the detection head of the target detection model, wherein the output category includes normal identification, missing identification, damaged identification, and incorrectly pasted identification.

[0107] In this embodiment, the augmentation training module 503 includes: an injection occlusion unit 5031, used to inject Gaussian noise and simulate occlusion processing into the labeled dataset to obtain a first training set; a transformation processing unit 5032, used to perform mirroring, rotation, scaling, translation, and brightness adjustment transformation processing on the labeled dataset using AutoAugment to obtain a second training set; a cropping unit 5033, used to crop the labeled dataset using Cutout to obtain a third training set; a mixing unit 5034, used to mix the labeled dataset, the first training set, the second training set, and the third training set using mixup to obtain an augmented dataset; and a training unit 5035, used to train an object detection model using the augmented dataset.

[0108] In this embodiment, the adjustment and optimization module 504 includes: a second adjustment unit 5041, used to dynamically adjust sample weights during the training process of the object detection model using the Focal Loss loss function; an optimization unit 5042, used to optimize the training process of the object detection model by combining a cosine annealing learning rate decay strategy and gradient pruning technique; and a termination unit 5043, used to terminate the training of the object detection model using an early stopping mechanism.

[0109] In this embodiment, the deployment recording module 505 includes: a deployment unit 5051, used to deploy the target detection model to the logistics sorting node for real-time detection of package images; an identification and determination unit 5052, used to identify the abnormality type of the special item identifier and determine the specific location of the current special item package when the target detection model detects an abnormality of the special item identifier; and a recording unit 5053, used to record the abnormality type and specific location to obtain abnormality type and package location information.

[0110] In this embodiment, the system further includes: an acquisition module 506, used to acquire high-definition image information of abnormal packages; a merging and generation module 507, used to merge the high-definition image information and the abnormality type and package location information to generate early warning information; and a push module 508, used to push the early warning information to the sorting line control terminal, so that the sorting line control terminal can generate and display a visual early warning page based on the early warning information.

[0111] above Figure 5 and Figure 6 The special part identification anomaly recognition device in the embodiments of the present invention will be described in detail from the perspective of modular functional entities. The special part identification anomaly recognition device in the embodiments of the present invention will be described in detail from the perspective of hardware processing.

[0112] Figure 7 This is a schematic diagram of the structure of a special part identification anomaly recognition device 600 provided in an embodiment of the present invention. The special part identification anomaly recognition device 600 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 610 (e.g., one or more processors) and a memory 620, and one or more storage media 630 (e.g., one or more mass storage devices) storing application programs 633 or data 632. The memory 620 and storage media 630 can be temporary or persistent storage. The program stored in the storage media 630 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the special part identification anomaly recognition device 600. Furthermore, the processor 610 may be configured to communicate with the storage media 630 and execute a series of instruction operations in the storage media 630 on the special part identification anomaly recognition device 600 to implement the steps of the special part identification anomaly recognition method provided in the above-described method embodiments.

[0113] The special component identification and anomaly detection device 600 may also include one or more power supplies 640, one or more wired or wireless network interfaces 650, one or more input / output interfaces 660, and / or one or more operating systems 631, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 7 The structure of the special part identification anomaly recognition device shown does not constitute a limitation on the special part identification anomaly recognition device. It may include more or fewer parts than shown, or combine certain parts, or have different part arrangements.

[0114] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of a special component identification anomaly recognition method.

[0115] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0116] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0117] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method of identifying abnormality of a special part, characterized by, The application relates to a special piece identification method and device based on a target detection model. Collecting historical special piece parcel image samples from cameras of distribution centers and warehouse centers, labeling special piece identification in the special piece parcel image samples by using a labeling tool to obtain a labeled data set; Constructing a target detection model based on an RTMDet algorithm, and adjusting parameters of the target detection model according to special piece identification task requirements; Performing data enhancement processing on the labeled data set to obtain an enhanced data set, and training the target detection model by using the enhanced data set; In the training process of the target detection model, the sample weight is dynamically adjusted by using a Focal Loss loss function, and the training process of the target detection model is optimized by combining a cosine annealing learning rate decay strategy and a gradient clipping technology; Deploying the target detection model to a logistics sorting node for real-time parcel image detection, and recording abnormal type and parcel position information when the target detection model detects special piece identification abnormalities.

2. The method of claim 1, wherein The method comprises the following steps: Collecting historical parcel images captured by cameras of distribution centers and warehouse centers, and screening historical special piece parcel image samples containing special pieces from the historical parcel images; Labeling a boundary box of a special piece identification region in the special piece parcel image samples by using a LabelImg or CVAT labeling tool to obtain labeled image samples; Dividing the labeled image samples into a training set, a verification set and a test set to form a labeled data set.

3. The method of claim 1, wherein The method comprises the following steps: Constructing a target detection model based on an RTMDet algorithm; Obtaining a special piece identification task requirement file, analyzing the special piece identification task requirement file to obtain special piece identification size, special piece identification aspect ratio and output category; Modifying the feature pyramid layer number of the neck part of the target detection model according to the special piece identification size; Adjusting the anchor size ratio of the target detection model according to the special piece identification aspect ratio; Setting the output category in the detection head of the target detection model, and the output category comprises a normal identification, an identification loss, an identification damage and an identification pasting error.

4. The method of claim 1, wherein The method comprises the following steps: Performing Gaussian noise injection and simulated occlusion processing on the labeled data set to obtain a first training set; Performing mirror, rotation, scaling, translation and brightness adjustment transformation processing on the labeled data set by using AutoAugment to obtain a second training set; Performing cutting on the labeled data set by using Cutout to obtain a third training set; Mixing the labeled data set, the first training set, the second training set and the third training set by using mixup to obtain an enhanced data set; training the target detection model by using the enhanced dataset.

5. The method of claim 1, wherein The Focal Loss loss function is used to dynamically adjust the sample weight in the training process of the target detection model, and the cosine annealing learning rate decay strategy and gradient clipping technology are combined to optimize the training process of the target detection model, including: The Focal Loss loss function is used to dynamically adjust the sample weight in the training process of the target detection model; The cosine annealing learning rate decay strategy and gradient clipping technology are combined to optimize the training process of the target detection model; An early stopping mechanism is adopted to terminate the training of the target detection model.

6. The method of claim 1, wherein The target detection model is deployed to the logistics sorting node for real-time detection of parcel images, and when the target detection model detects special piece identification abnormalities, the abnormal type and parcel location information are recorded, including: The target detection model is deployed to the logistics sorting node for real-time detection of parcel images; When the target detection model detects special piece identification abnormalities, the abnormal type of the special piece identification is identified, and the specific location of the current special piece parcel is determined; The abnormal type and the specific location are recorded to obtain abnormal type and parcel location information.

7. The method of claim 1, wherein After the target detection model is deployed to the logistics sorting node for real-time detection of parcel images, and when the target detection model detects special piece identification abnormalities, the abnormal type and parcel location information are recorded, the method further includes: Obtain high-definition image information of the abnormal parcel; Merge the high-definition image information and the abnormal type and parcel location information to generate early warning information; The early warning information is pushed to the sorting line control terminal, so that the sorting line control terminal generates and displays a visual early warning page based on the early warning information.

8. A special component identification anomaly recognition device, characterized in that, It includes: A collection module is used to collect historical special piece parcel image samples from the cameras of the sorting center and the warehouse center, and to label the special piece identification in the special piece parcel image samples using a labeling tool to obtain a labeled dataset; An adjustment module is used to construct a target detection model based on the RTMDet algorithm and adjust the parameters of the target detection model according to the special piece identification recognition task requirements; An enhanced training module is used to perform data enhancement processing on the labeled dataset to obtain an enhanced dataset, and to train the target detection model by using the enhanced dataset; An adjustment and optimization module is used to dynamically adjust the sample weight in the training process of the target detection model by using the Focal Loss loss function, and to optimize the training process of the target detection model by combining the cosine annealing learning rate decay strategy and the gradient clipping technology; A deployment and recording module is used to deploy the target detection model to the logistics sorting node for real-time detection of parcel images, and when the target detection model detects special piece identification abnormalities, the abnormal type and parcel location information are recorded.

9. A special part identification anomaly recognition apparatus characterized by comprising: The special piece identification abnormality identification device includes a memory and at least one processor, and the memory stores instructions. The at least one processor invokes the instructions in the memory to cause the special part identification anomaly recognition device to perform the steps of the special part identification anomaly recognition method of any one of claims 1-7.

10. A computer-readable storage medium having stored thereon instructions, the instructions comprising, The instructions, when executed by a processor, implement the steps of the special part identification anomaly recognition method of any one of claims 1-7.

Images