Efficient intelligent detection method for delivery parcels based on transfer learning

Abstract

The application relates to the technical field of intelligent detection of delivery, and discloses an efficient intelligent detection method for delivery parcels based on transfer learning. The method acquires a parcel image dataset of a target delivery scene, screens relevant source domain parcel images from a cross-domain parcel image source based on a preset transfer learning framework, decouples features of the source domain parcel images, separates surface texture features, geometric structure features and motion blur features, forms a plurality of decoupled feature maps, constructs a transfer learning constraint condition suitable for a scene according to the target scene dataset, performs domain adaptation transformation on the decoupled feature maps by using the constraint condition, generates a target domain adaptation feature map, inputs the target domain adaptation feature map into a multi-branch interaction network for feature interaction learning, obtains a degeneration feature map with fused multi-dimensional characteristics, and finally performs state recognition on the degeneration feature map by using a preset parcel state classifier to output a detection result.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology for postal and parcel services, specifically to an efficient intelligent detection method for postal and parcel services based on transfer learning. Background Technology

[0002] With the rapid development of the modern logistics and delivery industry, the number and types of parcels are increasing daily, making the demand for automated parcel inspection more urgent. Parcel inspection involves judging the status of parcels and whether there are any anomalies, and is a crucial link in ensuring delivery security and improving logistics efficiency. However, existing parcel inspection methods face many challenges.

[0003] Different delivery scenarios vary significantly, such as warehousing and sorting, last-mile delivery, and transportation. These scenarios differ in lighting conditions, background environment, and package placement, leading to substantial domain-specific variations in package image data distribution. Traditional detection models are typically trained in a single, specific scenario. When transferred to other scenarios, the model's performance deteriorates significantly due to data distribution mismatch, limiting its generalization ability.

[0004] To achieve good detection results, traditional methods often rely on a large amount of labeled image data of packages in the target scene. However, in practical applications, large-scale labeling of package images in different scenarios requires significant manpower, resources, and time, especially for some emerging or special delivery scenarios, where it is difficult to quickly obtain sufficient labeled data, thus limiting model training. Regarding feature extraction, existing technologies mostly employ single feature extraction methods, failing to comprehensively capture the key characteristics of packages. Some methods only focus on the geometric structural features of the package, ignoring information such as material and markings contained in surface texture features; others do not consider the blurred features caused by motion during transportation or photography, which may reflect the package's transportation status or whether it has undergone special processing. This one-sidedness in feature extraction leads to insufficient model representation ability of packages, affecting detection accuracy.

[0005] The application of cross-domain transfer learning in package detection is still immature. Existing transfer learning methods lack adaptation mechanisms for delivery scenarios when performing cross-domain feature transfer, failing to effectively filter source domain data relevant to the target scenario, resulting in irrelevant source domain information interfering with model learning. Furthermore, during feature adaptation, the decoupled features are not specifically adjusted, making it difficult for source domain features to adapt well to the target scenario, further limiting the improvement of detection performance. These issues make it difficult for existing methods to meet the needs of efficient and accurate package detection in different delivery scenarios. Summary of the Invention

[0006] The purpose of this invention is to provide an efficient intelligent detection method for postal parcels based on transfer learning, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, this invention provides an efficient intelligent detection method for postal parcels based on transfer learning, the method comprising:

[0008] Obtain a dataset of package images for the target delivery scenario, and based on a preset transfer learning framework, filter source domain package images related to the target delivery scenario from cross-domain package image sources;

[0009] The source domain wrapping image is decoupled to separate the wrapping surface texture features, geometric structure features and motion blur features, forming multiple decoupled feature maps;

[0010] Based on the package image dataset of the target delivery scenario, scenario-adaptive transfer learning constraints are constructed, and the multiple decoupled feature maps are transformed by the transfer learning constraints to generate multiple target domain-adaptive feature maps.

[0011] The multiple target domain adaptation feature maps are input into a preset multi-branch interaction network for feature interaction learning to obtain a degenerate feature map that integrates the multi-dimensional characteristics of the package.

[0012] The degraded feature map is identified using a preset package status classifier, and the intelligent detection result of the delivered package is output.

[0013] Preferably, the source domain wrapping image is decoupled to separate the wrapping surface texture features, geometric structure features, and motion blur features, forming multiple decoupled feature maps, including:

[0014] Construct a feature decoupling network based on an adversarial generative mechanism, wherein the feature decoupling network includes a shared encoder and multiple independent decoders;

[0015] The shared encoder is used to perform deep feature extraction on the source domain wrapped image to generate a shared feature code;

[0016] The shared feature encoding is input into the first independent decoder, the second independent decoder, and the third independent decoder, respectively;

[0017] The surface texture features of the package are reconstructed using the first independent decoder, the geometric structure features are reconstructed using the second independent decoder, and the motion blur features are reconstructed using the third independent decoder.

[0018] A first decoupling feature map is generated based on the surface texture features of the package, a second decoupling feature map is generated based on the geometric structure features, and a third decoupling feature map is generated based on the motion blur features.

[0019] Preferably, based on the package image dataset of the target delivery scenario, scenario-adaptive transfer learning constraints are constructed, and the transfer learning constraints are used to perform domain adaptation transformation on the multiple decoupled feature maps to generate multiple target domain-adaptive feature maps, including:

[0020] Extract the distribution statistics of the package image dataset of the target delivery scenario, and calculate the domain difference measure between the distribution statistics and the distribution statistics of the source domain package images;

[0021] Based on the domain difference metric, a maximum mean difference loss function is constructed as the core transfer learning constraint.

[0022] The first decoupling feature map, the second decoupling feature map, and the third decoupling feature map are respectively input into the domain adaptation transformation module;

[0023] The parameters of the domain adaptation transformation module are optimized using the maximum mean difference loss function, so that the first target domain adaptation feature map, the second target domain adaptation feature map, and the third target domain adaptation feature map output by the domain adaptation transformation module meet the feature distribution requirements of the target delivery scenario.

[0024] Preferably, the multiple target domain adaptation feature maps are input into a preset multi-branch interaction network for feature interaction learning to obtain a degenerate feature map that fuses multi-dimensional characteristics of the package, including:

[0025] Configure a multi-branch interaction network with a spatial attention mechanism, wherein the multi-branch interaction network includes a first feature processing branch, a second feature processing branch and a third feature processing branch;

[0026] The first target domain adaptation feature map is input into the first feature processing branch for texture feature enhancement processing to generate a texture enhancement feature map.

[0027] The second target domain adaptation feature map is input into the second feature processing branch for geometric feature topology processing to generate a structural topology feature map.

[0028] The third target domain adaptation feature map is input into the third feature processing branch for motion feature compensation processing to generate a motion compensation feature map.

[0029] The degradation feature map is generated by fusing the texture enhancement feature map, structural topology feature map, and motion compensation feature map through a cross-branch feature gating mechanism.

[0030] Preferably, the degradation feature map is generated by fusing the texture enhancement feature map, structural topology feature map, and motion compensation feature map through a cross-branch feature gating mechanism, including:

[0031] Calculate the first feature response correlation degree between the texture enhancement feature map and the structure topology feature map;

[0032] Calculate the second feature response correlation degree between the structural topology feature map and the motion compensation feature map;

[0033] A texture structure fusion weight matrix is ​​generated based on the first feature response correlation, and a structure motion fusion weight matrix is ​​generated based on the second feature response correlation.

[0034] The texture enhancement feature map and the structure topology feature map are weighted and fused using the texture structure fusion weight matrix to generate a primary fusion feature map.

[0035] The primary fused feature map and the motion compensation feature map are fused a second time using the structural motion fusion weight matrix to generate the degraded feature map.

[0036] Preferably, a preset package status classifier is used to perform status recognition on the degraded feature map, and intelligent detection results of the delivered package are output, including:

[0037] Construct a state classifier based on a dynamic prototype network, and initialize multiple packaged state prototype vectors in the state classifier;

[0038] The degraded feature map is mapped to the package state feature space to generate a package state feature vector to be classified.

[0039] Calculate the feature distance between the feature vector of the package state to be classified and the prototype vector of each package state;

[0040] The target package status category corresponding to the feature vector of the package to be classified is determined according to the principle of minimum feature distance.

[0041] The target package status category is output as the intelligent detection result of the package being delivered.

[0042] Preferably, after outputting the intelligent detection results of the delivered parcel, the method further includes:

[0043] Extract the incorrectly identified samples corresponding to the intelligent detection results, and calculate the feature offset between the incorrectly identified samples and the correct samples;

[0044] A prototype correction vector is constructed based on the feature offset, and the prototype correction vector is used to update the package state prototype vector in the state classifier.

[0045] The updated package state prototype vector is used to identify the state of the subsequent input degenerate feature map.

[0046] Preferably, the process involves extracting the incorrectly identified samples corresponding to the intelligent detection results and calculating the feature offset between the incorrectly identified samples and the correct samples, including:

[0047] The package image samples marked as misclassified in the intelligent detection results are obtained as misidentified samples;

[0048] Retrieve correct samples of the same type as the incorrectly identified samples from the package image dataset of the target delivery scenario;

[0049] Extract the first unclassified package state feature vector corresponding to the incorrectly identified sample and the second unclassified package state feature vector corresponding to the correct sample;

[0050] The vector offset between the first package state feature vector and the second package state feature vector in the feature space is calculated as the feature offset.

[0051] Preferably, constructing a prototype correction vector based on the feature offset, and updating the package state prototype vector in the state classifier using the prototype correction vector, includes:

[0052] The prototype movement direction is determined based on the direction of the feature offset, and the prototype movement step size is determined based on the magnitude of the feature offset.

[0053] Construct a prototype correction vector based on the prototype's movement direction and the prototype's movement step size;

[0054] The package state prototype vector of the corresponding category in the state classifier is superimposed with the prototype correction vector to generate the updated package state prototype vector.

[0055] Preferably, after acquiring the parcel image dataset of the target delivery scenario, the method further includes:

[0056] The target delivery scene's package image dataset is subjected to illumination consistency correction and perspective distortion correction to generate a standardized package image dataset;

[0057] The standardized package image dataset is used as input for the subsequent construction of transfer learning constraints.

[0058] Compared with the prior art, the beneficial effects of the present invention are:

[0059] This method, through a series of optimizations, brings multiple positive impacts to intelligent parcel detection. In the source domain image selection stage, based on a pre-defined transfer learning framework, source domain parcel images relevant to the target delivery scenario are selected from cross-domain parcel image sources. This reduces the interference of irrelevant source domain data on the model learning process. The removal of irrelevant data allows the model to focus its learning resources on information valuable to the target scenario, avoiding the problem of low learning efficiency caused by redundant data, and laying a more effective data foundation for subsequent feature processing and model training.

[0060] The feature decoupling step separates the source domain package image into surface texture features, geometric structure features, and motion blur features, forming multiple decoupled feature maps. This enables refined extraction of multi-dimensional features of the package. Texture features include detailed information such as the material of the package surface and printed markings; geometric structure features reflect the spatial attributes of the package, such as its shape, size, and outline; and motion blur features are related to the package's motion state and shooting conditions. This multi-feature decoupling approach breaks through the limitations of feature mixing in traditional feature extraction, allowing each dimension of features to be clearly expressed. This enables the model to more comprehensively perceive the essential characteristics of the package and avoids detection bias caused by the lack of key features.

[0061] The construction and application of scene-adaptive transfer learning constraints effectively reduces the impact of differences between the source and target domains. Image data from different delivery scenarios vary in terms of lighting, background, and noise. Directly using source domain features for transfer can lead to a mismatch between features and the target scene. However, by constructing constraints adapted to the target scene to perform domain adaptation transformation on the decoupled feature map, the source domain features can better fit the data distribution characteristics of the target scene while retaining key information. This makes the features more representative in the target scene and reduces the performance loss caused by domain differences.

[0062] The application of multi-branch interactive networks promotes information fusion among features of different dimensions. Decoupled features each carry different information, and the multi-branch structure enables targeted learning of each feature map, facilitating feature interaction during the learning process. This allows features to complement and reinforce each other. The interaction between texture features and geometric structure features helps the model more accurately determine whether a package is deformed, while the combination of geometric structure features and motion blur features aids in analyzing whether the package has undergone abnormal transportation. This feature interaction mechanism enhances the overall representational ability of the features, making the generated degenerate feature map more comprehensively reflect the multi-dimensional characteristics of the package. Attached Figure Description

[0063] Figure 1 This is a time sequence diagram of the efficient intelligent detection method for postal parcels based on transfer learning described in this invention.

[0064] Figure 2A flowchart for decoupling image features from the source domain;

[0065] Figure 3 A flowchart for feature learning in a multi-branch interactive network;

[0066] Figure 4 This is a flowchart for the state recognition of a package state classifier. Detailed Implementation

[0067] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0068] Please see Figure 1 This invention provides an efficient intelligent detection method for postal parcels based on transfer learning, the method comprising:

[0069] This process involves acquiring a dataset of package images from the target delivery scenario. Based on a pre-defined transfer learning framework, source domain package images relevant to the target delivery scenario are selected from cross-domain package image sources. Feature decoupling is performed on the source domain package images to separate surface texture features, geometric structure features, and motion blur features, forming multiple decoupled feature maps. Based on the package image dataset of the target delivery scenario, scene-adaptive transfer learning constraints are constructed. These constraints are then used to perform domain adaptation transformations on the multiple decoupled feature maps, generating multiple target domain-adapted feature maps. These target domain-adapted feature maps are input into a pre-defined multi-branch interactive network for feature interaction learning, obtaining a degraded feature map that integrates the multi-dimensional characteristics of the package. A pre-defined package state classifier is used to perform state recognition on the degraded feature map, outputting intelligent detection results for the delivered packages.

[0070] Example 1: See Figure 2A feature decoupling network based on an adversarial generative mechanism is constructed, comprising a shared encoder and multiple independent decoders. The shared encoder employs a deep convolutional neural network structure to extract deep features from the source domain wrapping image. The shared encoder consists of multiple convolutional and pooling layers. The convolutional layers use 3x3 or 5x5 convolutional kernels to extract low-level features such as edges and corners through convolution operations. The pooling layers then reduce feature dimensionality through max pooling or average pooling, reducing computational complexity. As the network depth increases, the shared encoder extracts high-level abstract features of the wrapping, generating shared feature codes. These shared feature codes contain a general representation of the wrapping, covering material, shape, and dynamic information, providing a unified input basis for reconstructing different features. During the parameter training of the shared encoder, the adversarial generative principle is used to optimize feature extraction accuracy through the interaction between the generator and the discriminator. The discriminator distinguishes between real and fake feature samples, prompting the generator (encoder) to generate more realistic feature codes.

[0071] The system uses a shared feature encoding input from multiple independent decoders, including a first, second, and third independent decoder. These decoders have different structures to accommodate the reconstruction of different feature types. The first independent decoder focuses on reconstructing the surface texture features of the package, employing upsampling and deconvolution operations. Upsampling expands the feature map size through bilinear interpolation or transposed convolution to recover high-resolution information; deconvolution uses stride and padding operations to progressively reconstruct texture details. The first independent decoder includes a dedicated texture feature reconstruction layer that combines Gabor filters or LBP operators to analyze the material, pattern, and color variations of the package surface, outputting the package surface texture features. The second independent decoder handles geometric features, employing different upsampling strategies, focusing on maintaining shape accuracy. The geometric feature extraction layer of the second independent decoder integrates boundary detection algorithms such as the Canny edge detector, describing the package's 3D structure and contour information through feature point matching and curvature analysis. The third independent decoder targets motion blur features; the upsampling process incorporates deblurring techniques, and the motion feature analysis layer combines optical flow computation methods to estimate the package's velocity direction in the logistics environment, generating motion blur features.

[0072] The surface texture features, geometric structure features, and motion blur features are fed into the feature map generation module, where they are converted into decoupled feature maps with a unified format. The feature map generation module first performs normalization processing, using min-max normalization or standardization methods to scale each feature value to the same range, eliminating amplitude differences. Subsequently, the module adjusts the feature dimensions, unifying the map size to a predetermined size, such as 224x224 pixels, through fully connected layers or matrix transformations. Surface texture features generate the first decoupled feature map, geometric structure features generate the second decoupled feature map, and motion blur features generate the third decoupled feature map. The entire process ensures compatibility between the maps and avoids feature conflicts.

[0073] The training of the shared encoder relies on a source-domain wrapped image dataset. The input images undergo preprocessing, such as grayscale conversion or noise removal, to enhance data quality. The training process employs iterative optimization. In each iteration, the shared encoder processes a batch of image samples, calculates feature loss, and jointly updates weights with the decoder. Decoder parameters are trained independently. The first independent decoder minimizes reconstruction differences using texture reconstruction loss functions such as perceptual loss or structural similarity indices. The second independent decoder relies on geometric constraint losses, such as Chamfer distance, to ensure shape consistency. The third independent decoder uses motion consistency losses, such as optical flow error terms. An adversarial mechanism is introduced into the discriminator network. The discriminator evaluates the authenticity of the decoder's output features, driving the encoder and decoder to improve together and enhancing the robustness of feature decoupling. The output of the feature map generation module is stored in a specific format for easy access and processing later.

[0074] The feature decoupling network architecture prioritizes computational efficiency, employing a parallel computing mechanism where the feature codes output by the shared encoder can be simultaneously distributed to each independent decoder. The first independent decoder is configured with single or multiple output channels to accommodate fine-grained texture representation; the second independent decoder applies graph neural network principles to map geometric features to a low-dimensional embedding space; and the third independent decoder integrates time-series analysis to predict motion trajectories. The feature map generation module supports batch processing, with batch normalization accelerating convergence. The entire network is deployed on dedicated hardware such as GPUs, leveraging hardware acceleration to shorten processing time.

[0075] The generation of shared feature encoding relies on the size adaptation of the input image. The source domain wrapping image is uniformly scaled to a standard resolution before input to avoid feature loss. Depth control of the shared encoder avoids overfitting by incorporating Dropout layers or regularization terms. The activation function of the first independent decoder is ReLU or sigmoid, highlighting texture details in nonlinear transformations. The architecture of the second independent decoder includes skip connections to preserve low-level geometric information. The deblurring layer of the third independent decoder uses an iterative optimization method with a fixed number of iterations. The normalization standard of the feature map generation module is based on the input feature distribution, dynamically adjusting the normalization range. In the data flow, the source domain wrapping image is passed through the shared encoder only once, and the feature encoding is distributed to multiple decoders for processing, saving computational resources. The error handling mechanism includes a feature verification step, checking for map dimension errors and restarting the generation process. During the debugging phase, visualization tools can be used to verify the output feature maps, ensuring that the feature decoupling quality meets expected standards. System logs record the processing time and resource consumption of each step, facilitating maintenance and upgrades.

[0076] After feature maps are generated, each decoupled feature map is stored in memory or an external database, with indexes indicating the corresponding feature type. Network parameters are initialized using the Xavier or He method to balance weight distribution and accelerate training convergence. Training samples are selected to balance different wrapping types and avoid data skew. The first decoupled feature map primarily stores surface texture information, the second decoupled feature map stores structural data in matrix form, and the third decoupled feature map records motion-related parameters. After decoupling, the maps can be used for subsequent processing stages, and the network output interface is defined in a standard API format. During testing, the overall model consisting of a shared encoder and decoder is encapsulated as an independent module, directly invoked via the input image path. Performance evaluation uses internal monitoring metrics but avoids reliance on external experimental reports. The integration of the feature decoupled network supports modular updates; individual decoders can be replaced independently without affecting overall functionality. The network runtime environment is compatible with mainstream frameworks such as TensorFlow or PyTorch, and deployment scripts configure memory optimization parameters. In the processing pipeline, the input rate of the source domain wrapping image matches the computational capacity to avoid queue backlog. Input data preprocessing steps include size alignment and format conversion to reduce compatibility errors. The shared feature encoding maintains a unified dimension and data type, and the decoder input format adaptation incurs no conversion overhead. The output of the feature map generation module is directly passed to subsequent modules, ensuring consistent data format and supporting chained processing. The overall design meets real-time requirements, with network inference time controlled at the millisecond level. The network architecture's scalability supports the addition of decoders to address the need for different feature types. Data security measures, such as encrypted storage, protect the feature map content. The user interface provides configuration options to modify the number of decoder layers or the range of feature dimensions. Model persistence ensures that model files are periodically saved to prevent data loss due to interruption. The feature decoupling network is integrated into the overall system as a core component, and its processing flow is documented for easy reference by implementation personnel. Network debugging tools assist in parameter tuning, but the results are only used for internal verification and do not generate external reports. System resource monitoring ensures that the network operates within safe thresholds, and a hardware load balancing mechanism automatically allocates computing resources. Finally, the feature decoupling network outputs three decoupling feature maps, each representing a high-level representation of specific package characteristics, for access and use in subsequent stages. The feature map storage management mechanism supports batch reading and writing, and the standardized data format facilitates integration. The reliability of the processing link is improved through redundancy design and error checking, and the network output quality monitoring system logs abnormal events. During implementation, key modules of the feature decoupling network are tested in isolation, and the outputs of each decoder are visualized for qualitative verification, but no quantitative experimental conclusions are generated. The network upgrade path supports incremental training, with the model adaptively adjusting to new data inputs. The input and output protocols of the entire feature decoupling process are clearly defined, making it easy to integrate into transfer learning frameworks. Network operating parameters are configurable, optimizing computational resource allocation. A data flow control mechanism coordinates the synchronous execution of the encoder and decoder, avoiding deadlock or latency issues.The efficiency of feature map generation relies on hardware acceleration, with GPU parallel computing optimizing processing speed. Version control of the feature decoupling network records change history for easy backtracking and troubleshooting. Implementation details comprehensively cover network construction, operation processes, and maintenance mechanisms. The output interface design supports third-party calls, exporting decoupling feature maps to a universal file format. Feature map content verification is performed through internal consistency checks, such as feature value range verification. Processing latency optimization uses asynchronous I / O and caching techniques to improve throughput. The network's fault tolerance is achieved through timeout mechanisms and retry logic, ensuring processing continuity. The shared encoder's feature extraction algorithm is based on a standard convolutional model, with parameter sharing reducing redundant computation. The decoder design emphasizes domain adaptation, with the specialization of texture, geometric, and motion features optimizing output accuracy. The feature map generation module's dimensionality adjustment algorithm dynamically matches input features and automatically scales to a preset size. Resource consumption monitoring tools for the decoupling process report memory and CPU usage, guiding performance optimization. Network training data employs augmentation strategies such as random rotation or cropping to improve feature generalization ability without relying on external data augmentation libraries. The adversarial mechanism is implemented based on classic generative adversarial networks (GANs) principles, but without involving mathematical derivation. The normalization algorithm for the feature maps depends on the input data type, and parameters are adjustable to adapt to the characteristics of different wrapped images. Periodic summaries of the network's operational status are generated after deployment, used for system health checks rather than performance verification reports. All code in the implementation process is modularly written, and unit tests cover core functional logic. Network parameter initialization rules are standardized, automatically adapting to different deployment environments. A concurrency control mechanism for the data processing pipeline manages image batches, ensuring fair scheduling. Access permissions for the output feature maps are set with a permission management policy, authorizing specific modules to use them. The complete lifecycle of the feature decoupling network, from construction to obsolescence, is clearly managed, and implementers follow documented steps. Finally, the decoupled feature maps are passed as intermediate products to the next processing stage, with seamless data flow.

[0077] Example 2: See Figure 3 This paper constructs a computational process for domain difference measurement, extracting distribution statistics parameters from the target delivery scenario package image dataset. These distribution statistics include pixel-level mean, channel variance, and higher-order moments of the spatial distribution. Simultaneously, corresponding statistical parameters from the source domain package images are collected, and both sets of statistics are input into the difference analysis module. This module employs the maximum mean difference algorithm, using a preset kernel function to map the feature space and comparing distribution differences through nonlinear transformation. During algorithm execution, the distance between the source and target domain samples is calculated after kernel function processing, yielding a numerical domain difference measurement result. Based on this measurement, a maximum mean difference loss function is constructed. This function serves as a core part of the transfer learning constraints, guiding the subsequent domain adaptation process.

[0078] The topology of the domain adaptation transformation module is designed, consisting of three independent multilayer perceptrons, corresponding to the inputs of the first, second, and third decoupled feature maps, respectively. Each multilayer perceptron contains five to seven fully connected layers, using a combination of ReLU and Tanh activation functions. The feature map data is flattened into a one-dimensional vector and input into the perceptron, where a nonlinear transformation is performed in the hidden space. The maximum mean difference loss function is connected to the perceptron output, and the perceptron parameters are optimized through backpropagation. The optimization process is iterative; in each iteration, the loss function calculates the difference between the current output features and the target domain distribution, generating a gradient signal to adjust the network weights. After parameter optimization, the perceptron outputs the first, second, and third target domain adaptation feature maps, whose feature distributions conform to the data patterns of the target delivery scenario.

[0079] The basic architecture of a multi-branch interactive network is configured, which integrates a spatial attention mechanism and is divided into three functional branches. The first feature processing branch contains a texture feature enhancement unit, configured with a four-layer convolutional structure, using a combination of 3×3 and 1×1 kernel sizes. The branch input receives the first target domain adaptation feature map. After initial convolutional layers extract local features, the spatial attention submodule generates a feature weight map, enhancing important texture regions through pixel-level multiplication. Subsequent convolutional layers stack the processed signals, ultimately outputting a texture-enhanced feature map. The second feature processing branch configures a geometric feature topology processing unit, whose core component is a graph neural network topology. The input second target domain adaptation feature map is converted into a node graph after region segmentation. Node features include geometric parameters, and edge features describe spatial relationships. The graph convolutional layers perform three message passing operations, aggregating neighborhood information and updating node states, outputting a structural topology feature map. The third feature processing branch deploys a motion feature compensation processing unit, which integrates a time series analysis module and an optical flow estimation algorithm. After the third target domain adaptation feature map is input, the optical flow field estimation module calculates the pixel displacement vector, the displacement compensation matrix corrects the position of the feature value, the motion artifact suppression algorithm eliminates residual noise, and finally outputs the motion compensation feature map.

[0080] A cross-branch feature gating fusion mechanism is established, which processes texture enhancement feature maps, structural topology feature maps, and motion compensation feature maps. In the initial stage, the spatial correlation between the texture enhancement feature map and the structural topology feature map is calculated, and a first correlation matrix is ​​generated using a pixel-wise cosine similarity algorithm. The similarity values ​​are normalized to a texture-structure fusion weight matrix in the 0-1 interval. Simultaneously, the mutual information between the structural topology feature map and the motion compensation feature map is calculated, and the information entropy of local regions is statistically analyzed using a sliding window to generate a structural-motion fusion weight matrix. The first-stage fusion operation is performed: corresponding pixels of the texture enhancement feature map and the structural topology feature map are weighted and summed according to the texture-structure fusion weight matrix to form a primary fused feature map. The second-stage fusion operation combines the primary fused feature map and the motion compensation feature map with the structural-motion fusion weight matrix, performing channel-level weighted superposition to generate a degenerate feature map. This feature map fuses texture, geometry, and motion three-dimensional characteristics while maintaining the spatial resolution of the original feature map. During the fusion mechanism processing, the three feature maps need to undergo data format alignment, including size scaling, channel number unification, and numerical normalization preprocessing.

[0081] A hardware acceleration scheme for a multi-branch interactive network is deployed, with the three feature processing branches employing a parallel computing architecture. Convolutional operations in the texture feature enhancement branch are accelerated using GPUs, with the Winograd algorithm optimizing convolution efficiency. The graph neural network operations in the geometric feature topology processing branch utilize a dedicated graph computing accelerator, employing a sparse matrix storage format to save memory resources. Optical flow calculations in the motion feature compensation branch utilize a hardware-level optical flow engine, with fixed functional units handling motion vector estimation. Matrix operations for the cross-branch feature gating mechanism are performed by tensor processing units, with weight calculation and feature fusion handled in separate pipeline stages. A memory management module dynamically allocates computing resources for each branch and monitors GPU memory usage in real time. Input feature maps employ a block loading mechanism, dividing large-scale feature maps into sub-block sequences for processing. The network output degenerate feature maps are stored in a GPU memory buffer, awaiting subsequent calls from the state classification module.

[0082] Multiple quality control nodes were set up during implementation. During the domain adaptation transformation phase, a feature distribution monitor continuously measured the statistical differences between the output features and the target domain. When the difference exceeded a set threshold, a recalculation command was triggered. Each processing unit of the multi-branch interactive network has a built-in feature quality evaluator: the texture enhancement processing branch detects texture sharpness indicators, the geometric topology processing branch measures structural integrity parameters, and the motion compensation processing branch calculates motion artifact residue values. A feature consistency check mechanism was deployed during the cross-branch fusion phase to compare the dimensionality matching of the input and output feature maps. The network is configured with a dynamic tuning interface, allowing online adjustment of the sensitivity coefficient of the spatial attention submodule, the number of iterations of the graph convolutional layer, and the optical flow estimation accuracy parameters. Debugging data, including the mean, variance, and distribution histogram of the feature maps at each stage, was recorded through a log system as internal status monitoring data but not used in performance verification reports. The operating status of each processing unit was fed back to the central scheduler in real time, and abnormal events triggered adaptive processing strategy switching. The degraded feature maps output by the multi-branch interactive network were standardized before being transmitted to the state classifier, with strict format validation rules defined for the data interface.

[0083] The entire processing flow is modularly encapsulated, with the domain difference measurement module, domain adaptation transformation module, and multi-branch interaction network forming independent computing units. Data exchange between units adopts a standardized data bus protocol, and feature maps transmit additional metadata description information. A parameter storage unit stores the final weight parameters of the domain adaptation transformation module and the configuration parameters of the multi-branch network, automatically loaded during system initialization. A computing resource scheduler coordinates the execution order of each module and establishes a dependency tree to manage the processing flow. Error handling mechanisms cover abnormal scenarios such as feature map size mismatch, data overflow, and computation timeouts; error recovery strategies include data retransmission and parameter reset. The execution efficiency of the processing pipeline is optimized through a time-slice round-robin mechanism, with priority scheduling allocated to critical path operations. System resource monitoring tools track CPU, GPU, and memory usage, maintaining stable resource utilization. The implementation environment is compatible with mainstream operating systems and computing frameworks, and the deployment scheme supports containerized encapsulation and cloud platform expansion. The processing flow documentation details the input / output specifications, parameter configuration ranges, and exception code definitions for each step. A version control system manages algorithm iteration updates, and each module has independent upgrade capabilities without affecting overall functionality.

[0084] Example 3: See Figure 4 A feature response correlation calculation module is constructed, which handles the matching relationship between texture enhancement feature maps and structural topology feature maps. The texture enhancement feature map is represented as... ,in Indicates the feature map height. Indicates the width of the feature map. This represents the number of channels. The structural topology feature map is represented as follows: To facilitate computation, the same dimensional specifications are maintained. The first feature response correlation is calculated using an improved cosine similarity method, comparing feature vectors position-by-position in the spatial dimension. For a feature map with coordinates... For a local region, the correlation degree is defined as follows:

[0085]

[0086] in: Indicates position The correlation value at the location, A small constant is used to prevent division by zero errors. The calculation results constitute the first characteristic response correlation matrix. The matrix elements take values ​​in the range [-1, 1]. Similarly, the structural topological feature map... With motion compensation feature map Calculate the correlation matrix of the second characteristic response. The same calculation logic is used, but the channel weight parameters are adjusted.

[0087] The design incorporates a feature fusion weight generation mechanism to transform the correlation matrix into an operable weight distribution. The first feature response correlation matrix... After being processed by a normalization function, it is converted into a texture structure fusion weight matrix. The normalization process employs a variant of the hyperbolic tangent function, preserving the relative magnitudes of the weights while limiting the output range. The second feature response correlation matrix... The structural motion fusion weight matrix is ​​transformed by a sigmoid function. The design emphasizes the complementary relationship between motion and geometric features. Spatial smoothing constraints are introduced during the weight matrix generation process, and Gaussian filtering is used to eliminate local abrupt changes, ensuring the continuity of the weight distribution.

[0088] A two-stage feature fusion operation is implemented. The first stage processes the texture enhancement feature map. With structural topology feature map Fusion. At each spatial location of the feature map. At each location, according to the weight value of the corresponding position. Perform a weighted combination. Specifically, this operation is defined as channel-wise linear interpolation to generate a primary fused feature map. The second stage of fusion processing involves the initial fusion feature map. With motion compensation feature map The combination of structural motion fusion weight matrix is ​​used. The fusion process is guided. This stage introduces a channel attention mechanism to dynamically adjust the fusion ratio of different feature channels, ultimately outputting a degraded feature map. The fusion operation preserves the spatial structure information of the original feature map while achieving the organic integration of multi-dimensional features.

[0089] Construct a state classifier for a dynamic prototype network and initialize it. A prototype vector of package status Each prototype vector This represents a typical feature pattern of a packaged state. Degenerate feature map. The spatial dimension is compressed by a global average pooling layer, generating a feature vector of the package state to be classified. The feature distance calculation module compares the input feature vectors. With each prototype vector The similarity is measured using Mahalanobis distance:

[0090]

[0091] in: Indicates up to the The distance of the prototype This is the covariance matrix for the corresponding category. The distance calculation result is input into the softmax function to generate the category probability distribution, and the category with the highest probability is selected as the target package state category for output. The classifier has a built-in prototype vector update mechanism that dynamically adjusts the spatial distribution of the prototype vectors based on the classification results.

[0092] A hardware acceleration scheme for feature fusion and classification is implemented. Feature response correlation calculation is deployed on a tensor processing unit, and a matrix multiplication accelerator is used to optimize computational efficiency. The weight matrix generation stage calls a dedicated function unit to perform nonlinear transformations and filtering operations. The two-stage feature fusion operation is implemented through pixel-level parallel processing, and the GPU thread block allocation strategy matches the spatial dimension of the feature maps. Vector distance calculation for the dynamic prototype network uses batch processing mode, simultaneously calculating the distances between multiple samples and all prototypes. The covariance matrix inversion operation is accelerated using Cholesky decomposition, and the decomposition results are pre-calculated and cached to reduce redundant computation overhead.

[0093] The system includes quality control and debugging mechanisms. During the feature fusion phase, feature value range checks are implemented to prevent numerical overflow or underflow. The correlation matrix generation process records statistical distribution parameters and monitors the dispersion of matrix elements. A weight matrix visualization tool assists in observing spatial weight distribution patterns and verifies the rationality of weight generation. During prototype network runtime, the system monitors changes in feature distance distribution to identify anomalous deviations. The classification decision process records the top-K candidate classes and their probability values ​​to support subsequent error analysis. System logs detail the computation time and resource usage of each step for performance optimization reference.

[0094] The processing workflow achieves end-to-end modular integration, forming a coherent processing chain from feature response correlation calculation, weight matrix generation, feature fusion, and state classification. Data interfaces define strict format specifications, and feature map transmission includes metadata describing its dimensions and numerical range. An error handling mechanism covers scenarios such as numerical anomalies, dimension mismatches, and computation timeouts, automatically triggering recovery processes. The system resource manager dynamically allocates computing resources, prioritizing the execution efficiency of critical path operations. A version control system manages algorithm parameter iterations, supporting online updates of prototype vectors and fusion weights. The deployment scheme is compatible with multiple hardware platforms, and the core computing module provides multi-precision computation options to adapt to different scenario requirements. Processing results are output in a standardized format as a detection report, including classification results and interpretable analysis data of related features.

[0095] Example 4: The system first filters out erroneous samples from the intelligent detection results. After the package status classifier outputs the detection results, erroneously identified samples are extracted through a preset labeling mechanism. These samples typically exhibit classification confidence levels below a threshold or inconsistencies with manual verification results. For example, a package that is actually classified as "fragile" is misclassified as "ordinary item," and the system records the package image and its corresponding unclassified package status feature vector as an erroneously identified sample. Simultaneously, the system retrieves samples from the package image dataset of the target delivery scenario that belong to the same category as the erroneously identified samples and are correctly classified. For example, all package images correctly classified as "fragile" and their feature vectors are retrieved as correct samples. The feature offset calculation process between erroneously identified samples and correct samples is shown in the table below. Table 1 shows the analysis data of three typical erroneous samples.

[0096] Table 1: Analysis data of three typical error samples.

[0097] Sample number Error categories Correct category Feature Dimension Offset direction Offset modulus Key Difference Dimensions E-0452 ordinary items fragile items 256 [0.12,-0.08,...,0.05] 1.75 12,45,189 E-1078 document Electronic products 256 [-0.21,0.15,...,-0.03] 2.31 7,63,201 E-2093 Large cargo cold chain items 256 [0.08,0.12,...,-0.11] 1.98 34,78,215

[0098] The table data shows that each erroneous sample exhibits significant differences across specific feature dimensions. Sample E-0452 shows marked deviations from correct samples in dimensions 12, 45, and 189; these dimensions may correspond to the characteristics of the package's cushioning material. The system generates a feature offset by calculating the difference between the first unclassified package state feature vector of the erroneous sample and the second unclassified package state feature vector of the correct sample. This offset includes orientation and magnitude information; the orientation information represents the numerical trend of each dimension in the 256-dimensional vector space, while the magnitude reflects the overall degree of deviation.

[0099] The construction of the prototype correction vector is based on the analysis results of the feature offsets. For sample E-0452, the system first analyzes its offset direction, determining that the prototype vector needs to be moved towards positive dimensions 12 and 189, and negative dimension 45. The movement step size is scaled according to the offset magnitude of 1.75, and a logarithmic compression function is used to map the magnitude to an appropriate adjustment range. The final generated prototype correction vector has a correction value of +0.15 in the key dimension 12, -0.10 in dimension 45, and +0.08 in dimension 189. The adjustment range for the other dimensions is gradually reduced according to the decay coefficient. This selective adjustment method can avoid the distortion of the classification boundary caused by over-correction.

[0100] The prototype vector update for the state classifier employs an incremental adjustment strategy. Taking the prototype vector for the "fragile" category as an example, the original vector has a value of 0.76 in dimension 12, which is updated to 0.91 after correction vector adjustment. The system sets a maximum adjustment limit to prevent a single sample from excessively influencing the prototype vector. After each update, the new prototype vector is immediately applied to subsequent state recognition tasks. Simultaneously, the system retains historical versions of the prototype vector; when multiple erroneous samples of the same type appear consecutively, the adjustment trend can be analyzed and the correction strategy optimized.

[0101] The error sample analysis module's workflow includes multiple quality control steps. During the feature offset calculation phase, the system checks the dimensionality consistency of two feature vectors and prunes outliers. In the generation of the prototype correction vector, a direction verification step is included to ensure the offset direction aligns with the direction of classification performance improvement. The updated prototype vector needs to be tested on a validation set to confirm that it can correct target error samples while maintaining its original correct classification ability. The system records detailed parameters for each prototype adjustment, including adjustment time, operator, and number of affected samples, forming a complete adjustment log.

[0102] Data management during implementation employs a hierarchical storage strategy. Frequently accessed current prototype vectors are stored in memory, historical prototype vectors and error sample analysis data are stored in a cache, and complete operation logs are written to a persistent database. The system maintains an independent prototype vector update channel for each package status category, supporting parallel adjustments to prototypes of different categories. Error sample feature vectors are stored in a compressed format to reduce memory usage. All data access operations are performed through a unified interface, ensuring data consistency and security.

[0103] Hardware resource configuration has been optimized for feature vector operations. Prototype vectors are stored in GPU memory, accelerating distance calculations. Feature offset calculation tasks are distributed across multiple computing cores for parallel processing. The system monitors resource usage and automatically initiates distributed computing mode when processing large-scale error sample sets. The memory management module dynamically allocates computing buffers and adjusts resource quotas for each processing stage based on task requirements. Network communication optimizations reduce data transmission latency between nodes, maintaining efficient collaborative processing capabilities in a cluster deployment environment.

[0104] The user interface of the error sample analysis system provides visualization tools. Operators can view the comparison curves of error samples and correct samples across key feature dimensions to intuitively understand the classification bias. The adjustment process of the prototype vector is displayed in animation, showing the vector's movement trajectory in the feature space. The system generates a schematic diagram of the changes in the classification boundary to help understand the impact of adjustments on overall classification performance. All visualization results support interactive exploration, allowing users to focus on specific dimensions or sample sets for in-depth analysis.

[0105] In terms of system integration, the error sample analysis module and the state classifier form a closed-loop control. Error samples output by the classifier automatically trigger the analysis process, and the generated prototype correction vectors are directly fed back to the classifier. This real-time adjustment mechanism enables the classifier to adapt to changes in data distribution. The module provides an application programming interface (API) that allows external systems to query the current prototype vector status or submit batch adjustment requests. Integration with the data acquisition system ensures that newly collected package images can be quickly used to verify the prototype adjustment effect.

[0106] The version control mechanism manages the iteration process of prototype vectors. Each major adjustment creates a new version number, recording complete adjustment parameters and impact assessments. The system supports rollback to any historical version to address performance degradation after adjustments. Version difference analysis tools compare the distribution changes of prototype vectors across different versions in the feature space, aiding in optimal version selection. All version data is stored encrypted to ensure intellectual property security.

[0107] Standardizing operational procedures is a crucial aspect of implementation. The criteria for labeling erroneous samples clearly define what constitutes a classification error. The operational procedures for calculating feature offsets detail data preprocessing steps and anomaly handling methods. The prototype calibration approval process specifies the authorization levels required for adjustments of varying magnitudes. The system maintenance manual contains comprehensive troubleshooting guidelines, covering various scenarios ranging from data anomalies to hardware failures. Training materials demonstrate standard operational procedures and solutions to common problems through real-world case studies.

[0108] The implementation of the error sample analysis system significantly improved the adaptive capability of the state classifier. By continuously analyzing classification errors and adjusting prototype vectors accordingly, the system progressively optimizes the decision boundaries for each category. This dynamic adjustment mechanism enables the classifier to adapt to changes in package characteristics, such as the introduction of new packaging materials or changes in transportation conditions. The system design focuses on balancing immediate adjustments with global stability, ensuring that each correction improves classification performance without introducing new biases. The entire implementation process established a complete workflow for error analysis, prototype adjustment, and performance verification, providing technical support for the continuous optimization of the intelligent detection system.

[0109] Example 5: Standardized preprocessing of a package image dataset starting with the target delivery scene. This preprocessing includes two key operations: illumination consistency correction and perspective distortion correction. In the illumination consistency correction stage, the image processing system first analyzes the brightness distribution histogram of all images, calculating the average brightness level and contrast range. Based on this analysis, the system establishes an illumination compensation model, using adaptive histogram equalization to enhance details in underexposed areas and applying gamma transform to compress highlights in overexposed areas. Intermediate results are stored after this process for subsequent steps. The illumination compensation model parameters are dynamically adjusted during batch processing to adapt to package images acquired at different times. In the perspective distortion correction stage, a calibration algorithm is deployed to identify planar reference objects in the package images. The system calculates perspective distortion parameters based on the known dimensions of the reference objects, generating the corresponding inverse transformation matrix. Image pixels are resampled to the correction coordinates through bilinear interpolation to generate a standardized package image dataset. The standardized dataset is stored in a dedicated buffer area, and metadata records the correction parameters and historical operation logs.

[0110] The construction of transfer learning constraints is based on a standardized dataset as input. The feature distribution analysis module calculates the multidimensional statistical features of the dataset, including channel mean, cross-channel covariance matrix, and higher-order moment distribution parameters. The system encodes the statistical results into transfer descriptor vectors, which serve as the benchmark for subsequent domain adaptation transformations. The descriptor vectors are automatically updated every 24 hours to dynamically track changes in scene features. This vector is then transmitted to the constraint generation engine, which compares it with the source domain feature statistics to generate a domain difference report.

[0111] In the prototype vector adjustment process, feature offsets are handled using a geometric analysis approach. The system uses the feature vector of the incorrectly identified sample as the starting point and the feature vector of the correct sample as the ending point; the direction of the line connecting the two points is defined as the prototype movement direction. The direction vector is normalized to eliminate the influence of length. The magnitude of the feature offset serves as the basic measure of the movement distance, but it needs to be converted into the actual movement step size through a non-linear mapping. The system calls a logarithmic compression function to process the original magnitude value, preventing over-adjustment caused by abnormally large offsets. The compression parameters are adaptively configured according to the current classifier state: the compression intensity is increased when the overall classification error rate is high, and the compression intensity is decreased when the error rate is low. The learning rate parameter is stored in the system configuration file, and operators can adjust it within the range of 0.01 to 0.1 through the management interface.

[0112] The prototype correction vector is synthesized based on the movement direction and adjustment step size. The system multiplies the unit direction vector by the adjustment step size to generate the basic correction amount. Under specific conditions, a random perturbation component is added: if three consecutive adjustments are made in the same direction, a small-amplitude random change in the orthogonal direction is introduced to prevent the prototype vector from getting trapped in a local optimum. After the correction vector is generated, a boundary check procedure is executed to confirm whether it exceeds the preset feature space range. For out-of-bounds vectors, the system automatically reduces the magnitude proportionally to keep it within the effective space. All correction operations generate detailed transaction logs, recording the timestamp of vector changes, operator identification, and impact analysis results.

[0113] The prototype vector update for the packaged state in the state classifier employs an incremental write mechanism. The system locks the target prototype vector and performs atomic operations: read the current value, add a correction vector, and write back the new value. In high-concurrency environments, a version number control mechanism prevents conflicting writes. After each update, a vector backup program is triggered, saving the state before adjustment in the historical version repository. The historical version retention strategy is set as follows: the most recent 100 modifications are retained in their entirety; for any subsequent modifications, only the differential data is retained. The prototype vector storage area uses a double-buffered structure to achieve seamless switching, ensuring that the update process does not affect the real-time classification task.

[0114] The input process for the standardized dataset in the transfer learning framework is implemented through a standardized interface. The dataset is read in segments according to a predetermined batch size and transmitted to the distributed statistics module via a high-speed data channel. The statistics module deploys a streaming processing algorithm to calculate cumulative statistics and update constraint parameters in real time. To maintain responsiveness, the system allocates dedicated computing resources to handle constraint update tasks, including an independent thread pool and memory cache. When new constraints are generated, a system event is published to notify the relevant modules, triggering parameter synchronization operations in each processing unit.

[0115] The preprocessing and vector adjustment system employs a multi-level monitoring mechanism for operation and management. In the image preprocessing stage, the system detects input image quality indicators (such as sharpness and noise level), issuing a re-acquisition request when they fall below a threshold. In the perspective correction stage, verification points are deployed, and the geometric distortion of the corrected image is sampled and measured. The prototype vector adjustment process monitors changes in feature space distribution in real time, activating an alarm when the distance between different prototype categories falls below a safety threshold. Resource consumption monitoring covers memory usage, CPU load, and disk I / O, triggering resource reclamation procedures in case of anomalies. Operation logs are stored in a structured manner, and key performance indicators (such as average preprocessing latency and vector update frequency) are included in the system health report.

[0116] A complete anomaly handling framework was built during implementation. Overprocessing protection rules were defined during the illumination correction phase: processing was paused and manually reviewed when the pixel value variation coefficient exceeded a set upper limit. The retry logic for distortion correction failures included three automatic attempts and a rollback strategy: rolling back from the perspective model to affine transformation. Collision detection was implemented during prototype vector adjustment: if two categories of correction paths intersected, the system paused adjustment and executed a conflict resolution procedure. All hardware failure events triggered an emergency protocol: the preprocessing task switched to degraded mode, and the prototype vector was rolled back to the latest stable version.

[0117] The standardized deployment architecture supports multi-node scaling. The image preprocessing unit can be horizontally scaled to multiple compute nodes, and the task scheduler automatically allocates resources based on the load. The migration constraint generator is designed in a master-slave mode, with the standby node taking over the computing tasks when the master node fails. The prototype vector storage uses a distributed key-value database, supporting cross-datacenter replication to ensure disaster recovery capabilities. Each functional module is connected via a standardized API, and interface version management is compatible with forward scaling. The deployment toolchain includes automated configuration scripts, a container image generator, and resource template files.

[0118] The user interface integrates key control points. The preprocessing module offers a manual correction mode, allowing operators to fine-tune the automated results. The prototype vector management view displays spatial relationships between categories and an adjustment timeline. The system setup wizard guides operators through configuring sensitive parameters, including the shape of the learning rate curve and the range of boundary constraints. The diagnostic panel displays intermediate images and feature space visualizations in real time during the preprocessing workflow, aiding in problem diagnosis.

[0119] Strict operational procedures are established for the execution process. A timed calibration process for image acquisition equipment ensures input data quality, and camera parameter verification is performed daily. The prototype adjustment approval workflow has tiered access permissions: low-risk modifications are executed automatically, while high-risk operations require superior approval. Periodic maintenance tasks include monthly cleanup of temporary files and quarterly recalibration of feature space reference points. Operator training courses cover preprocessing principles and vector adjustment mechanisms, and analyze handling processes using historical fault case studies. All these implementation details constitute a complete technical solution, supporting the stable operation and continuous optimization of the intelligent parcel detection system.

[0120] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0121] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A highly efficient intelligent detection method for postal parcels based on transfer learning, characterized in that, The method includes: Obtain a dataset of package images for the target delivery scenario, and based on a preset transfer learning framework, filter source domain package images related to the target delivery scenario from cross-domain package image sources; Construct a feature decoupling network based on an adversarial generative mechanism, wherein the feature decoupling network includes a shared encoder and multiple independent decoders; The shared encoder is used to perform deep feature extraction on the source domain wrapped image to generate a shared feature code; The shared feature encoding is input into the first independent decoder, the second independent decoder, and the third independent decoder, respectively. The surface texture features of the package are reconstructed using the first independent decoder, the geometric structure features are reconstructed using the second independent decoder, and the motion blur features are reconstructed using the third independent decoder. A first decoupling feature map is generated based on the surface texture features of the package; a second decoupling feature map is generated based on the geometric structure features; and a third decoupling feature map is generated based on the motion blur features. Based on the package image dataset of the target delivery scenario, scenario-adaptive transfer learning constraints are constructed, and the multiple decoupled feature maps are transformed by the transfer learning constraints to generate multiple target domain-adaptive feature maps. Configure a multi-branch interaction network with a spatial attention mechanism, wherein the multi-branch interaction network includes a first feature processing branch, a second feature processing branch and a third feature processing branch; The first target domain adaptation feature map is input into the first feature processing branch for texture feature enhancement processing to generate a texture enhancement feature map. The second target domain adaptation feature map is input into the second feature processing branch for geometric feature topology processing to generate a structural topology feature map. The third target domain adaptation feature map is input into the third feature processing branch for motion feature compensation processing to generate a motion compensation feature map. Calculate the first feature response correlation degree between the texture enhancement feature map and the structure topology feature map; Calculate the second feature response correlation degree between the structural topology feature map and the motion compensation feature map; A texture structure fusion weight matrix is ​​generated based on the first feature response correlation, and a structure motion fusion weight matrix is ​​generated based on the second feature response correlation. The texture enhancement feature map and the structure topology feature map are weighted and fused using the texture structure fusion weight matrix to generate a primary fusion feature map. The structural motion fusion weight matrix is ​​used to perform a secondary fusion of the primary fusion feature map and the motion compensation feature map to generate a degenerate feature map; The degraded feature map is identified using a preset package status classifier, and the intelligent detection result of the delivered package is output.

2. The efficient intelligent detection method for postal parcels based on transfer learning as described in claim 1, characterized in that, Based on the package image dataset of the target delivery scenario, scenario-adaptive transfer learning constraints are constructed. These constraints are then used to perform domain adaptation transformation on the multiple decoupled feature maps, generating multiple target domain-adaptive feature maps, including: Extract the distribution statistics of the package image dataset of the target delivery scenario, and calculate the domain difference measure between the distribution statistics and the distribution statistics of the source domain package images; Based on the domain difference metric, a maximum mean difference loss function is constructed as the core transfer learning constraint. The first decoupling feature map, the second decoupling feature map, and the third decoupling feature map are respectively input into the domain adaptation transformation module; The parameters of the domain adaptation transformation module are optimized using the maximum mean difference loss function, so that the first target domain adaptation feature map, the second target domain adaptation feature map, and the third target domain adaptation feature map output by the domain adaptation transformation module meet the feature distribution requirements of the target delivery scenario.

3. The efficient intelligent detection method for postal parcels based on transfer learning as described in claim 1, characterized in that, The degraded feature map is analyzed using a pre-defined package status classifier to identify its status and output intelligent detection results for the delivered packages, including: Construct a state classifier based on a dynamic prototype network, and initialize multiple packaged state prototype vectors in the state classifier; The degraded feature map is mapped to the package state feature space to generate a package state feature vector to be classified. Calculate the feature distance between the feature vector of the package state to be classified and the prototype vector of each package state; The target package status category corresponding to the feature vector of the package to be classified is determined according to the principle of minimum feature distance. The target package status category is output as the intelligent detection result of the package being delivered.

4. The efficient intelligent detection method for postal parcels based on transfer learning as described in claim 3, characterized in that, After outputting the intelligent detection results of the delivered parcels, the method further includes: Extract the incorrectly identified samples corresponding to the intelligent detection results, and calculate the feature offset between the incorrectly identified samples and the correct samples; A prototype correction vector is constructed based on the feature offset, and the prototype correction vector is used to update the package state prototype vector in the state classifier. The updated package state prototype vector is used to perform state recognition on the subsequent input degenerate feature map.

5. The efficient intelligent detection method for postal parcels based on transfer learning as described in claim 4, characterized in that, Extracting the incorrectly identified samples corresponding to the intelligent detection results, and calculating the feature offset between the incorrectly identified samples and the correct samples, including: The package image samples marked as misclassified in the intelligent detection results are obtained as misidentified samples; Retrieve correct samples of the same type as the incorrectly identified samples from the package image dataset of the target delivery scenario; Extract the first unclassified package state feature vector corresponding to the incorrectly identified sample and the second unclassified package state feature vector corresponding to the correct sample; The vector offset between the first package state feature vector and the second package state feature vector in the feature space is calculated as the feature offset.

6. The efficient intelligent detection method for postal parcels based on transfer learning as described in claim 5, characterized in that, Constructing a prototype correction vector based on the feature offset, and updating the package state prototype vector in the state classifier using the prototype correction vector, includes: The prototype movement direction is determined based on the direction of the feature offset, and the prototype movement step size is determined based on the magnitude of the feature offset. Construct a prototype correction vector based on the prototype's movement direction and the prototype's movement step size; The package state prototype vector of the corresponding category in the state classifier is superimposed with the prototype correction vector to generate the updated package state prototype vector.

7. The efficient intelligent detection method for postal parcels based on transfer learning as described in claim 1, characterized in that, After acquiring the parcel image dataset of the target delivery scenario, the method further includes: The target delivery scene's package image dataset is subjected to illumination consistency correction and perspective distortion correction to generate a standardized package image dataset; The standardized package image dataset is used as input for the subsequent construction of transfer learning constraints.

Images