Logistics package state automatic detection method based on artificial intelligence

Abstract

The invention discloses a logistics parcel state automatic detection method based on artificial intelligence, and belongs to the field of intelligent logistics, and the method comprises the following steps: employing a multi-scale pyramid feature extraction algorithm to extract appearance features of logistics parcels, and obtaining the curved surface feature unified representation of the logistics parcels; eliminating the illumination change influence on the logistics parcels in transportation by adopting an illumination invariance feature transformation algorithm based on the uniform representation of the curved surface features to obtain illumination consistency curved surface features; on the basis of the illumination consistency curved surface features, a curved surface feature completion algorithm based on an attention mechanism is adopted to complete the surface features of the logistics parcels to obtain key curved surface features; fusing the key curved surface features and the illumination consistency curved surface features to obtain fused wrapped curved surface features; and inputting the fused parcel curved surface features into an anomaly detection model to determine the anomaly type and position of the parcel curved surface.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent logistics, and particularly relates to a method for automatically detecting the status of logistics parcels based on artificial intelligence. Background Art

[0002] During the automatic detection of the status of logistics parcels, the detection of parcel curved surfaces is a key but extremely challenging technical problem. Due to the diverse shapes and sizes of logistics parcels, significant differences in surface materials, colors, and textures, and the variable lighting conditions during transportation, it is difficult to extract and analyze the curved surface features of parcels. First, the diversity and complexity of parcel curved surfaces make it difficult to unify feature representation, balance the correlation and distinguishability between different curved surface features, and easily introduce redundancy and noise, affecting the detection accuracy. Second, the dynamic change of lighting conditions in the logistics environment directly leads to significant differences in the brightness distribution and contrast of the parcel surface images, posing higher requirements for the robustness of curved surface feature extraction. At the same time, the random stacking and occlusion of parcels during transportation further exacerbate the uncertainty of curved surface detection. Some key curved surface features may be covered or deformed, making it difficult to accurately capture and depict. In addition, the real-time detection of the status of a large number of logistics parcels also poses extremely high requirements for the efficiency and performance of the algorithm. However, the extraction and analysis of curved surface features are computationally intensive tasks, making it difficult to achieve real-time response in scenarios with a large amount of data. Therefore, how to improve the accuracy, robustness, and real-time performance of parcel curved surface detection in the logistics scenario is a technical problem that urgently needs to be solved and is the key to realizing the automatic detection of the status of logistics parcels based on artificial intelligence. Therefore, the present invention proposes a method for automatically detecting the status of logistics parcels based on artificial intelligence. Summary of the Invention

[0003] To solve the above technical problems, the present invention proposes a method for automatically detecting the status of logistics parcels based on artificial intelligence to solve the problems existing in the above prior art.

[0004] To achieve the above object, the present invention provides a method for automatically detecting the status of logistics parcels based on artificial intelligence, including the following steps:

[0005] Adopt a multi-scale pyramid feature extraction algorithm to extract the external features of the logistics parcel to obtain a unified representation of the curved surface features of the logistics parcel;

[0006] Based on the unified representation of the curved surface features, use an illumination-invariant feature transformation algorithm to eliminate the influence of illumination changes on the logistics parcel during transportation to obtain illumination-consistent curved surface features;

[0007] Based on the illumination-consistent curved surface features, use a curved surface feature completion algorithm based on the attention mechanism to complete the surface features of the logistics parcel to obtain key curved surface features;

[0008] Fuse the key surface features and the illumination consistency surface features to obtain the fused wrapped surface features;

[0009] Input the fused wrapped surface features into an anomaly detection model to determine the anomaly type and location of the wrapped surface.

[0010] Optionally, the process of obtaining the unified representation of the surface features of the logistics package includes:

[0011] Preprocess the logistics package image using a multi-scale pyramid feature extraction algorithm according to the shape, size, and surface material color of the logistics package to obtain surface features at different scales;

[0012] Obtain a pyramid feature map based on the surface features at different scales;

[0013] Calculate the weight coefficients of the features at each scale using an attention mechanism based on the pyramid feature map;

[0014] Adopt a weighted fusion strategy to perform weighted summation on the weight coefficients of the features at each scale to obtain the unified representation of the surface features of the logistics package.

[0015] Optionally, the process of obtaining the illumination consistency surface features includes:

[0016] Obtain the surface image of the logistics package, and use an illumination estimation algorithm based on spherical harmonics to model the brightness and shadow distribution in different regions of the surface image to obtain the illumination distribution map of the surface image;

[0017] Calculate the reflectivity of each pixel point in the illumination distribution map to obtain the illumination-normalized image;

[0018] Adopt a contrast enhancement algorithm based on histogram equalization to enhance the contrast of the illumination-normalized image to obtain the contrast-enhanced image;

[0019] Adopt an illumination-invariant feature transformation algorithm to extract features from the contrast-enhanced image to obtain the illumination consistency surface features.

[0020] Optionally, after obtaining the illumination consistency surface features, it further includes identifying the position package based on the illumination consistency surface features; among them, the process included in the identification position includes:

[0021] Extract the surface features of the illumination consistency surface features based on the Gaussian pyramid and Laplacian pyramid to obtain the surface feature descriptor;

[0022] Adopt a feature matching algorithm based on Euclidean distance to calculate the Euclidean distance between the surface feature descriptor and each sample in the sample package feature library;

[0023] When the Euclidean distance is less than a preset threshold, it is determined that the current package is a known package, and historical transportation information of the known package is obtained according to the matching result;

[0024] When the Euclidean distance is greater than or equal to the preset threshold, it is determined that the current package is an unknown package, and the surface characteristics of the unknown package are added to the sample package feature library.

[0025] Optionally, the process of complementing the surface characteristics of the logistics package to obtain the key surface characteristics includes:

[0026] Obtain the three-dimensional point cloud data included in the logistics, and perform point cloud segmentation on the three-dimensional point cloud data to obtain a point cloud subset of a single package;

[0027] Calculate the normal vector of the single-package point cloud subset to obtain the normal vector information of each point;

[0028] Based on the normal vector information, divide the three-dimensional point cloud data into different surface regions;

[0029] Extract the feature descriptors of each surface region and construct a surface feature vector;

[0030] Perform normalization processing on the feature vector for illumination changes to obtain illumination-invariant surface characteristics;

[0031] Perform weighted aggregation on the illumination-invariant surface characteristics through an attention mechanism to obtain a key surface characteristic representation of the logistics package;

[0032] Judge feature missing based on the key surface characteristic representation, and complement the missing key surface characteristics through a convolutional neural network to obtain complete key surface characteristics.

[0033] Optionally, in the process of fusing the key surface characteristics and the illumination-consistent surface characteristics, multi-view geometric consistency constraints are used to optimize the surface characteristics to obtain the fused package surface characteristics.

[0034] Optionally, a reconstructed package surface is obtained based on the fused package surface characteristics, and the reconstructed package surface is optimized to obtain the overall structure and local details of the surface;

[0035] Among them, the process of obtaining the overall structure and local details of the reconstructed package surface includes:

[0036] Obtain the original three-dimensional data of the reconstructed package surface and perform preprocessing to obtain preprocessed surface data;

[0037] Extract multi-scale and multi-direction surface characteristics of the preprocessed surface data to obtain a surface characteristic representation;

[0038] Construct a topological graph for reconstructing the wrapped surface based on the surface feature representation;

[0039] Perform graph convolution operations on the topological graph using a surface feature fusion algorithm based on a graph convolutional neural network to obtain the initial overall structure and local detail information;

[0040] Through the stacking of multi-layer graph convolutional neural networks, layer by layer extract and fuse the high-level semantic features of the reconstructed wrapped surface to obtain the overall structure and local details of the reconstructed wrapped surface.

[0041] Compared with the prior art, the present invention has the following advantages and technical effects:

[0042] The present invention adopts a multi-scale pyramid feature extraction algorithm, combines illumination-invariant feature transformation and surface feature completion technology based on an attention mechanism. This technical solution realizes the efficient extraction and processing of the external features of logistics packages. By eliminating the influence of illumination changes, it ensures the illumination consistency of feature extraction, and at the same time accurately completes the surface features of the package, enhancing the integrity and robustness of the features. After fusing the key surface features and illumination consistency features, the obtained fused package surface features provide rich and accurate inputs for the anomaly detection model, effectively identifying and locating the types and positions of anomalies on the package surface, and greatly improving the accuracy and efficiency of logistics package detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The drawings forming a part of this application are used to provide a further understanding of this application. The schematic embodiments and descriptions thereof of this application are used to explain this application and do not constitute an improper limitation of this application. In the drawings:

[0044] Figure 1 It is a flowchart of an automatic logistics package status detection method based on artificial intelligence according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] It should be noted that, without conflict, the embodiments in this application and the features in the embodiments can be combined with each other. The following will refer to the drawings and combine the embodiments to detail this application.

[0046] It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0047] Embodiment 1

[0048] As Figure 1 shown, in this embodiment, an automatic logistics package status detection method based on artificial intelligence is provided, including the following steps:

[0049] Step S101, the multi-scale pyramid feature extraction algorithm is used to extract the appearance features of the logistics package to obtain a unified representation of the surface features of the logistics package. Considering the diversity of the shape, size, surface material and color of the logistics package, the multi-scale pyramid feature extraction algorithm is adopted to adaptively extract the surface features at different scales, and the weights of the features at each scale are dynamically adjusted through a weighted fusion strategy to balance the correlation and discrimination between different surface features, and improve the unity and robustness of the feature representation to obtain a unified representation of the surface features.

[0050] Furthermore, considering the diversity of the shape, size, surface material and color of the logistics package, the multi-scale pyramid feature extraction algorithm is used to preprocess the logistics package image. For the surface features at different scales, multi-scale features are extracted through a convolutional neural network to obtain a pyramid feature map. According to the correlation and discrimination of each scale feature map, an attention mechanism is used to calculate the weight coefficients of each scale feature. Through the weighted fusion strategy, the features at each scale are weighted and summed according to the weight coefficients to obtain a unified feature representation.

[0051] Even further, considering the diversity of the shape, size, surface material and color of the logistics package, first, the multi-scale pyramid feature extraction algorithm is used to preprocess the logistics package image. This algorithm generates multiple copies of the original image with different resolutions by constructing an image pyramid. For example, if the size of the original image is 1024×1024, images of different scales such as 512×512, 256×256, and 128×128 are generated. Then, for the surface features at different scales, multi-scale features are extracted through a convolutional neural network to obtain a pyramid feature map. The convolutional neural network uses the VGG-16 model, which contains 13 convolutional layers and 3 fully connected layers. Through convolutional operations with a convolutional kernel size of 3×3 and a stride of 1, local features at different scales are extracted, and then the features are aggregated through max pooling operations to obtain pyramid feature maps with scales of 1 / 2, 1 / 4, and 1 / 8. According to the correlation and discrimination of each scale feature map, an attention mechanism is used to calculate the weight coefficients of each scale feature. The attention mechanism generates a query vector, performs a dot product operation with each scale feature map to obtain attention weights, and then normalizes them through the Softmax function to obtain weight coefficients. The larger the weight coefficient, the greater the contribution of the feature at this scale to classification. Through the weighted fusion strategy, the features at each scale are weighted and summed according to the weight coefficients to obtain a unified feature representation with a dimension of 512.

[0052] Step S102: Based on the unified representation of surface features, use the illumination-invariant feature transformation algorithm to eliminate the influence of illumination changes on the logistics packages during transportation and obtain illumination-consistent surface features. After obtaining the unified representation of surface features, in response to the dynamic changes in illumination conditions during the logistics transportation process, introduce the illumination-invariant feature transformation algorithm. By performing illumination normalization and contrast enhancement on the surface images of the packages, eliminate the influence of illumination changes on the extraction of surface features, ensure consistent surface feature representations under different illumination conditions, and obtain illumination-consistent surface features.

[0053] Furthermore, the process of obtaining the illumination-consistent surface features includes: acquiring the surface image of the logistics package, using the illumination estimation algorithm based on spherical harmonics to model the brightness and shadow distribution in different regions of the surface image to obtain the illumination distribution map of the surface image; calculating the reflectivity of each pixel point in the illumination distribution map to obtain the illumination-normalized image; using the contrast enhancement algorithm based on histogram equalization to enhance the contrast of the illumination-normalized image to obtain the contrast-enhanced image; using the illumination-invariant feature transformation algorithm to extract features from the contrast-enhanced image to obtain the illumination-consistent surface features.

[0054] Even further, after acquiring the surface image of the package, first use the illumination estimation algorithm based on spherical harmonics. By modeling the brightness and shadow distribution in different regions of the image, estimate the illumination distribution in the image and generate an illumination distribution map with a resolution of 1024×1024. Then, use the illumination normalization algorithm to process the image. By calculating the reflectivity of each pixel point in the image and normalizing it to the range of [0,1], eliminate the influence of uneven illumination on the image quality. Next, use the contrast enhancement algorithm based on histogram equalization. By stretching the grayscale histogram of the image, increase the contrast of the image to 5 times the original, making the surface features in the image more prominent to obtain the illumination-consistent surface features. On this basis, use the multi-scale pyramid feature extraction algorithm. By constructing the Gaussian pyramid and Laplacian pyramid of the image, extract the surface features of the illumination-consistent surface features at different scales and generate a 128-dimensional surface feature descriptor. Finally, use the feature matching algorithm based on Euclidean distance to match the surface features of the current package with the pre-established package surface feature library containing 1000 samples. Calculate the Euclidean distance between the current package and each sample in the feature library. If the minimum distance is less than the threshold 8, then determine that the current package is a known package, and obtain the historical transportation information of the package according to the matching result for optimizing the subsequent logistics transportation path and timeliness; otherwise, identify the package as an unknown package, add its surface features to the feature library, and cluster and update the feature library according to the feature similarity to improve the feature library and enhance the accuracy of subsequent package identification, providing data support for realizing the intelligent sorting and scheduling of packages.

[0055] Step S103, based on the surface features of the illumination consistency, use the surface feature completion algorithm based on the attention mechanism to complete the surface features of the logistics package to obtain the key surface features. On the basis of obtaining the unified representation of the surface features, aiming at the problems of random stacking of logistics packages and occlusion of key surface features, a surface feature completion algorithm based on the attention mechanism is adopted. By learning the context information and structural priors of the package surface, the occluded or missing key surface features are automatically inferred and completed, improving the integrity and accuracy of surface feature extraction.

[0056] Furthermore, obtain the three-dimensional point cloud data of the logistics package, segment the point cloud data of a single package through a point cloud segmentation algorithm to obtain the point cloud subset of a single package; calculate the normal vector of the single package point cloud subset to obtain the normal vector information of each point, and divide the point cloud according to the normal vector information to obtain different surface regions; extract the feature descriptors of each surface region, construct a surface feature vector, normalize the feature vector for the illumination change to obtain the illumination-invariant surface feature; perform weighted aggregation on the surface feature through the attention mechanism to highlight the key surface features and suppress the non-key surface features to obtain the key surface feature representation of the package; determine whether there is occlusion or missing of the key surface features. If so, according to the context information and structural priors of the package surface, complete the missing key surface features through a convolutional neural network to obtain the complete key surface features; fuse the completed key surface features with the illumination-invariant surface features, and optimize the surface features through multi-view geometric consistency constraints to improve the accuracy and robustness of the surface features; based on the optimized surface features, identify and classify the logistics package through classification algorithms such as support vector machines to realize the intelligent processing and management of the logistics package.

[0057] Furthermore, a 3D laser scanner is deployed in the logistics warehouse to scan the packages entering the warehouse and obtain the 3D point cloud data of the packages. The region-growing based point cloud segmentation algorithm is adopted, and the normal vector angle threshold of the seed point is set to 15° and the curvature threshold is set to 1 to segment the point cloud data and obtain the point cloud subset of a single package. The PCA algorithm is used to calculate the normal vector of the point cloud subset, and the point cloud is divided into different surface regions according to the normal vector angle, and the angle threshold is set to 30°. A series of features, including curvature, normal vector, texture, etc., are extracted from each surface region to construct a 128-dimensional surface feature vector. In response to the light changes in the warehouse, the maximum and minimum normalization processing is performed on the feature vector, and the feature values are mapped to the interval [0,1] to obtain the light-invariant surface feature. By introducing the attention mechanism, the different surface features are weighted and aggregated, and the weights are automatically learned through the convolutional neural network to highlight the feature representation of the key surfaces. In the case of package occlusion or surface missing, the convolutional neural network is used to extract the context information and structural prior of the package surface to complete the missing key surface features, and the completion accuracy reaches more than 95%. The completed key surface features are cascaded with the light-invariant features, and the surface features are optimized through multi-view Figure 1 consistency constraints. The RANSAC algorithm is used to estimate the transformation relationship between different viewpoints, and the outliers are removed to improve the accuracy of the features. Finally, the optimized surface features are input into the support vector machine for training and classification, and the 5-fold cross-validation is used to evaluate the model performance. On the test set of 5000 packages, the recognition accuracy of 95% is achieved, meeting the requirements of intelligent processing of logistics packages.

[0058] The attention mechanism is used to weight and aggregate the surface features to highlight the key surface features and suppress the non-key surface features, obtaining the key surface feature representation of the package.

[0059] Obtain surface feature data, calculate the attention weight value of each surface feature according to a preset attention mechanism model to obtain an attention weight vector; perform a weighted aggregation operation on the surface features according to the attention weight vector to obtain an aggregated surface feature vector; determine whether each eigenvalue in the aggregated surface feature vector is greater than a preset key feature threshold. If it is greater, determine the feature as a key surface feature; if it is less, determine the feature as a non-key surface feature; construct a feature masking vector according to the determined key surface features and non-key surface features, where the vector value corresponding to the key surface feature is 1 and the vector value corresponding to the non-key surface feature is 0; perform a dot product operation on the aggregated surface feature vector and the feature masking vector to obtain a masked surface feature vector, achieving the effect of highlighting key surface features and suppressing non-key surface features; input the masked surface feature vector into a preset feature encoding model, and encode the masked surface feature vector through the feature encoding model to obtain a wrapped key surface feature representation; use the obtained wrapped key surface feature representation as the input for subsequent tasks to guide the execution of relevant tasks and improve task performance.

[0060] Based on the optimized surface features, identify and classify the logistics parcels through classification algorithms such as support vector machines.

[0061] Obtain the three-dimensional model data of the logistics parcel and extract the surface features of the three-dimensional model; perform optimization processing on the extracted surface features to remove redundant features and highlight key features; construct a support vector machine classification model according to the optimized surface features and use training data to train the model; obtain the three-dimensional model data of the logistics parcel to be identified and classified, extract its surface features and perform optimization processing; input the optimized surface features into the trained support vector machine classification model to obtain the identification and classification results of the logistics parcel; if the confidence level of the classification result is lower than the preset threshold, obtain auxiliary features through correlation analysis and re-perform classification judgment in combination with the surface features; perform automatic sorting, storage, and transportation on the logistics parcel according to the identification and classification results of the logistics parcel to improve logistics efficiency.

[0062] In step S104, to further enhance the robustness of surface feature extraction, introduce a feature enhancement algorithm based on adversarial learning. Automatically generate parcel surface samples with different materials, textures, and lighting conditions through a generative adversarial network and use them as training data to enhance the original feature extraction model, improving the model's adaptability and generalization ability to complex surface features.

[0063] According to the preset surface feature type, a multi-view scanning method is used to obtain the three-dimensional point cloud data of the target object, and a three-dimensional model of the wrapped surface is obtained through point cloud filtering and surface reconstruction algorithms. For the three-dimensional wrapped surface model, a deep learning-based feature extraction algorithm is used to extract multi-scale and multi-directional surface features. According to the extracted feature vectors, a support vector machine classifier is trained to obtain an initial recognition model of the surface features. A feature enhancement network based on adversarial learning is constructed, which includes two parts: a generator and a discriminator. The generator receives random noise and labels as inputs and generates three-dimensional wrapped surfaces similar to real samples. The discriminator receives real samples and generated samples and judges their authenticity. Influence factors such as material, texture, and lighting conditions are introduced into the generator, and the diversity of the generated samples is controlled by adjusting the network parameters. According to the preset material type, texture complexity, and lighting change range, a large number of three-dimensional wrapped surface samples with different attribute combinations are randomly generated. The generated three-dimensional wrapped surface samples are added to the original training data, and the same feature extraction and classification algorithms as the initial model are used to retrain the surface feature recognition model. By increasing the diversity of the data, the generalization ability and robustness of the model are improved. For the new surface to be recognized, its three-dimensional point cloud is extracted and the wrapped surface model is reconstructed, and the trained surface feature recognition model is used for prediction to obtain the feature category to which the surface belongs. The recognition result is compared with that of the initial model to evaluate the effect of the feature enhancement algorithm. If the recognition accuracy after feature enhancement does not meet the expectation, the generator is returned to adjust the material, texture, and lighting parameters, regenerate the training samples for reinforcement learning, and continuously iterate and optimize until the recognition accuracy meets the application requirements.

[0064] Specifically, according to the preset surface feature types, such as plane, cylindrical surface, spherical surface, etc., a multi-view scanning method is used to obtain the three-dimensional point cloud data of the target object. The Gaussian filtering algorithm is used to remove the noise and outliers in the point cloud data, and then the Poisson surface reconstruction algorithm is used to obtain the three-dimensional model of the wrapped surface, with the reconstruction accuracy controlled within 1 mm. For the three-dimensional wrapped surface model, a feature extraction algorithm based on PointNet++ is used to extract 512-dimensional local features and 1024-dimensional global features at different scales, and aggregate them in 6 directions to obtain a 6144-dimensional surface feature vector. According to the extracted feature vector, a support vector machine classifier with a radial basis kernel function is used for training, and the accuracy of 5-fold cross-validation reaches 93%, obtaining the initial recognition model of the surface features. A feature enhancement network based on WGAN-GP is constructed. The generator receives 100-dimensional Gaussian noise and 10-dimensional one-hot labels as inputs, and through 5 fully connected layers and 3 3D transposed convolution layers, generates a 128×128×128 three-dimensional wrapped surface similar to the real samples. The discriminator uses 5 3D convolutional layers and 3 fully connected layers to perform binary classification on the real samples and the generated samples, and adds a gradient penalty term to improve the training stability. Material parameters, normal maps, and HDR environment light maps are introduced into the generator, and by randomly sampling in the parameter space, the diversity of the generated samples is controlled. According to the preset 5 material types, 3 texture complexities, and 10 lighting change ranges, 10,000 three-dimensional wrapped surface samples with different attribute combinations are randomly generated. The generated three-dimensional wrapped surface samples are added to the original training data, and PointNet++ is used to extract features and SVM is used for classification to retrain the surface feature recognition model. Through data augmentation, the accuracy of the model on the test set is increased to 91%, and the generalization ability and robustness are improved. For the new surface to be recognized, its three-dimensional point cloud is extracted and the wrapped surface model is obtained through Poisson reconstruction, and the trained surface feature recognition model is used for prediction, with the accuracy reaching 97%, which is 4% higher than the initial model. To further improve the recognition accuracy, the generator is returned to adjust the material, texture, and lighting parameters, and 5000 training samples are regenerated for reinforcement learning. After 3 iterations, the recognition accuracy is increased to 92%, meeting the application requirements of automatic sorting of logistics parcels.

[0065] Step S105, after obtaining robust and complete surface features, introduce a surface feature fusion algorithm based on graph convolutional neural network. By constructing a topological graph of the wrapped surface, use graph convolutional operations to adaptively aggregate and fuse surface features in different regions, capture the overall structure and local details of the surface, and further improve the accuracy and robustness of surface detection.

[0066] Furthermore, a reconstructed wrapped surface is obtained based on the fused wrapped surface features, and the overall structure and local details of the surface are optimized; wherein, the process of obtaining the overall structure and local details of the reconstructed wrapped surface includes: obtaining the original three-dimensional data of the reconstructed wrapped surface and performing preprocessing to obtain preprocessed surface data; extracting multi-scale and multi-directional surface features of the preprocessed surface data to obtain a surface feature representation; constructing a topological graph of the reconstructed wrapped surface based on the surface feature representation; performing graph convolution operations on the topological graph using a surface feature fusion algorithm based on a graph convolutional neural network to obtain initial overall structure and local detail information; and through the stacking of multi-layer graph convolutional neural networks, layer by layer extracting and fusing the high-level semantic features of the reconstructed wrapped surface to obtain the overall structure and local details of the reconstructed wrapped surface.

[0067] Further, Step 1: Obtain the original three-dimensional data of the reconstructed wrapped surface and perform preprocessing on it, including operations such as denoising and smoothing, to obtain preprocessed surface data. Step 2: According to the preprocessed surface data, extract multi-scale and multi-directional surface features, including geometric features such as curvature and normal vector, as well as appearance features such as texture and color, to construct a robust and complete surface feature representation. Step 3: For the extracted surface feature representation, construct a topological graph model of the reconstructed wrapped surface, where the nodes of the graph represent the sampling points on the surface, the edges represent the connection relationships between the nodes, and the feature of the node is the surface feature vector of that point. Step 4: Design a surface feature fusion algorithm based on a graph convolutional neural network, and by applying graph convolution operations on the topological graph, adaptively aggregate and fuse the surface features of the nodes to capture the overall structure and local detail information of the surface. Step 5: During the graph convolution process, use an attention mechanism to dynamically adjust the weights of different node features, highlight the features of key regions, suppress the influence of noise and redundant information, and improve the accuracy and robustness of feature fusion. Step 6: Through the stacking of multi-layer graph convolutional neural networks, layer by layer extract and fuse the high-level semantic features of the surface to establish a hierarchical feature representation that characterizes the overall structure and local details of the surface. Step 7: Input the fused surface features into a classifier or a regressor to determine the surface category to which each sampling point belongs or predict its three-dimensional coordinates, thereby realizing the precise detection and reconstruction of the surface and obtaining the final surface detection result.

[0068] First, preprocess the surface to be detected. Use the bilateral filtering algorithm to remove high-frequency noise, and set the filtering window size to 5×5. Then, use the least squares method to fit the surface to obtain the smoothed three-dimensional model data. On this basis, extract the multi-scale geometric features of the surface, including the mean curvature, Gaussian curvature, and shape index at different scales, and the scale range is from 1 to 10 times the average side length of the surface. At the same time, extract the 72-dimensional color histogram feature and GLCM texture feature in the HSV color space to form a 128-dimensional appearance feature vector. Combine the geometric features and appearance features to obtain a 256-dimensional surface feature representation. Construct a topological graph model based on the feature vector, with the surface sampling points as nodes and the geodesic distance between nodes less than the threshold of 0.5 as edges. Apply spectral graph convolution on the topological graph, which is realized by the Chebyshev polynomial expansion, with the convolution kernel size of 3×3 and the feature dimension of 128. An attention mechanism is introduced during the convolution process, and the weights of different node features are controlled by a gating unit. The weight coefficients are generated by the attention matrix, where the rows of the matrix represent the central nodes and the columns represent the neighboring nodes. Stack 8 graph convolution layers and attention layers, and finally map to the target category or three-dimensional coordinates through a fully connected layer. Compare the prediction results with the ground truth manually annotated, and calculate the average IoU and chamfer distance. Test this method on the ModelNet40 dataset, and the surface detection accuracy of common objects such as airplanes, cars, and chairs reaches 93%, which is 5 percentage points better than the traditional method, proving the effectiveness and robustness of this method.

[0069] Step S106: Input the fused package surface features into the anomaly detection model based on multi-task learning. By jointly optimizing the surface classification and anomaly region localization tasks, simultaneously judge the anomaly type and location of the package surface, and realize the comprehensive and automatic detection of the logistics package status, providing reliable decision support for the intelligent logistics system.

[0070] Furthermore, based on the fusion features of the wrapped surface, an anomaly detection model based on a convolutional neural network is constructed. The loss functions of the surface classification and anomaly region localization tasks are simultaneously optimized through a multi-task learning framework to achieve end-to-end anomaly detection. The attention mechanism is used to weight the local regions of the wrapped surface to highlight the feature representation of the anomaly region and improve the accuracy of anomaly localization. Prior knowledge, such as the shape and texture features of normal wrapped surfaces, is introduced into the anomaly detection model, and the risk of overfitting is reduced through regularization constraints to improve the generalization ability of the model. For different types of anomalies, such as damage, stains, deformation, etc., targeted feature extraction methods, such as multi-scale texture descriptors, shape context descriptors, etc., are designed to improve the accuracy of anomaly type judgment. During the model training process, data augmentation methods such as rotation, scaling, noise addition, etc., are used to expand the diversity of anomaly samples and improve the robustness of the model. The anomaly detection results are combined with the attribute information of the logistics package, such as weight, volume, transportation mode, etc., to construct a decision support system to provide a comprehensive assessment and disposal suggestions for the package status in intelligent logistics. Based on the detected anomaly type and location information, the subsequent processing flow of the logistics package is optimized, such as automatic sorting, targeted inspection, etc., to improve the logistics efficiency and accuracy.

[0071] Furthermore, during the construction of the anomaly detection model, a convolutional neural network is used as the basic architecture, and a multi-task learning framework is designed to simultaneously optimize the two tasks of surface classification and anomaly region localization. Among them, the cross-entropy loss function is used for the surface classification task, and the focal loss function is used for the anomaly region localization task. The weight ratio of the loss functions of the two tasks is set to 1:2, and the model parameters are jointly optimized through the backpropagation algorithm. In the feature extraction part of the model, targeted feature descriptors are designed for different types of anomalies. For example, Gabor filters are used to extract multi-scale texture features, and shape context descriptors are used to describe local shape features. The extracted feature vectors are concatenated and then input into the fully connected layer for fusion. During the model training process, data augmentation techniques such as random rotation of ±10° and Gaussian noise (mean of 0, standard deviation of 0.1) are used to expand the diversity of anomaly samples and improve the robustness of the model. At the same time, regularization terms such as L1 norm and L2 norm are introduced to constrain the sparsity and smoothness of the model parameters, reduce the risk of overfitting, and improve the generalization ability. In the inference stage, the attention mechanism is used to weight the local regions of the wrapped surface to generate a heat map to highlight the anomaly region and improve the accuracy of anomaly localization. Finally, the anomaly detection results are fused with the attribute information of the logistics package to construct a decision support system for comprehensive assessment and disposal suggestions of the logistics package. For example, for a package detected with a damage anomaly, the system automatically sorts it to the manual inspection channel for targeted inspection and repair of the damaged part to improve the logistics efficiency and accuracy.

[0072] In view of the problems of the diversity of the shapes and materials of logistics parcels, as well as the light changes, random stacking, and occlusion during the transportation process, the present invention adaptively extracts surface features by using a multi-scale pyramid feature extraction algorithm, and introduces a light-invariant transformation and a feature completion algorithm based on an attention mechanism to eliminate the influence of light and infer occlusion features. The generalization ability of the feature extraction model is enhanced through adversarial learning, and then the graph convolutional neural network is used to fuse the surface features to capture the overall structure and local details. Finally, anomaly detection is realized based on multi-task learning, and at the same time, the type and location of the anomaly are judged. The present invention improves the robustness, integrity, and accuracy of the extraction of the surface features of logistics parcels, realizes the comprehensive automatic detection of the parcel status, and provides reliable decision-making support for the intelligent logistics system.

[0073] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed in the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An automatic detection method for logistics package status based on artificial intelligence, characterized in that: The following steps are involved: The multi-scale pyramid feature extraction algorithm is used to extract the surface features of the logistics package to obtain a unified representation of the surface features of the logistics package; Based on the unified representation of the surface features, an illumination invariant feature transformation algorithm is used to eliminate the influence of illumination changes on the logistics packages in transportation to obtain illumination consistent surface features; Based on the illumination consistency surface features, a surface feature completion algorithm based on an attention mechanism is used to complete the surface features of the logistics package to obtain key surface features; Fusing the key surface feature and the illumination consistency surface feature to obtain a fused wrapping surface feature; The fused wrapped surface features are input into an anomaly detection model to determine the anomaly type and location of the wrapped surface.

2. The method for automatic detection of logistics package status based on artificial intelligence according to claim 1 is characterized in that: The process of obtaining a unified representation of the surface features of a logistics package includes: According to the shape, size and surface material color of the logistics package, a multi-scale pyramid feature extraction algorithm is used to pre-process the logistics package image to obtain surface features at different scales; Obtaining a pyramid feature map based on the surface features at different scales; Based on the pyramid feature map, an attention mechanism is used to calculate the weight coefficient of each scale feature; A weighted fusion strategy is adopted to perform weighted summation on the weight coefficients of each scale feature to obtain a unified representation of the surface features of the logistics package.

3. The method for automatic detection of logistics package status based on artificial intelligence according to claim 1 is characterized in that: The process of obtaining lighting consistent surface features includes: Acquire a surface image of the logistics package, and use a spherical harmonic function-based illumination estimation algorithm to model the brightness and shadow distribution of different areas in the surface image to obtain an illumination distribution map of the surface image; Calculating the reflectivity of each pixel in the illumination distribution map to obtain an illumination normalized image; Using a contrast enhancement algorithm based on histogram equalization to perform contrast enhancement on the illumination normalized image to obtain a contrast enhanced image; An illumination invariant feature transformation algorithm is used to extract features from the contrast enhanced image to obtain illumination consistent surface features.

4. The method for automatic detection of logistics package status based on artificial intelligence according to claim 3 is characterized in that: After obtaining the illumination consistency surface feature, the method further includes identifying a position package based on the illumination consistency surface feature; wherein the process of identifying the position includes: Extracting the surface features of the illumination consistent surface features based on Gaussian pyramid and Laplacian pyramid to obtain a surface feature descriptor; A feature matching algorithm based on Euclidean distance is used to calculate the Euclidean distance between the surface feature descriptor and each sample in the sample wrapping feature library; When the Euclidean distance is less than a preset threshold, the current package is determined to be a known package, and the historical transportation information of the known package is obtained according to the matching result; When the Euclidean distance is greater than or equal to a preset threshold, the current package is determined to be an unknown package, and the surface features of the unknown package are added to the sample package feature library.

5. The method for automatic detection of logistics package status based on artificial intelligence according to claim 1 is characterized in that: The process of completing the surface features of the logistics package to obtain key surface features includes: Acquire three-dimensional point cloud data included in the logistics, and perform point cloud segmentation on the three-dimensional point cloud data to obtain a point cloud subset of a single package; Calculate the normal vector of a single parcel point cloud subset to obtain the normal vector information of each point; Based on the normal vector information, the three-dimensional point cloud data is divided into regions to obtain different surface regions; Extract feature descriptors of each surface area and construct surface feature vectors; Normalize the feature vector according to the illumination change to obtain the illumination-invariant surface features; The illumination-invariant surface features are weighted and aggregated through the attention mechanism to obtain the key surface feature representation of the logistics package. Based on the key surface feature representation, feature missing is judged, and the missing key surface features are supplemented through a convolutional neural network to obtain complete key surface features.

6. The method for automatic detection of logistics package status based on artificial intelligence according to claim 1 is characterized in that: In the process of fusing the key surface features and the illumination consistency surface features, multi-view geometric consistency constraints are used to optimize the surface features to obtain the fused wrapping surface features.

7. The method for automatic detection of logistics package status based on artificial intelligence according to claim 6 is characterized in that: A reconstructed wrapping surface is obtained based on the fused wrapping surface features, and the reconstructed wrapping surface is optimized to obtain the overall structure and local details of the surface; The process of obtaining the overall structure and local details of the reconstructed wrapped surface includes: Acquire the original three-dimensional data of the reconstructed wrapped surface and perform preprocessing to obtain preprocessed surface data; Extracting multi-scale and multi-directional surface features of the pre-processed surface data to obtain a surface feature representation; Constructing a topological map of the reconstructed wrapped surface based on the surface feature representation; A surface feature fusion algorithm based on a graph convolutional neural network is used to perform a graph convolution operation on the topological graph to obtain an initial overall structure and local detail information; By stacking multiple layers of graph convolutional neural networks, the high-level semantic features of the reconstructed wrapped surface are extracted and fused layer by layer to obtain the overall structure and local details of the reconstructed wrapped surface.

Images