Train number, train container number and type identification device

Abstract

The invention discloses a train number, train container number and train type recognition device, which relates to the technical field of train recognition and comprises an edge disturbance sensing module, a fusion feature analysis module, a structure guide reconstruction module, a path evaluation recognition module and a dynamic path regulation and control module. And the structure guide reconstruction module is used for carrying out channel reordering and space adaptive amplification on the image based on the determined morphological fusion features, enhancing character structure coherence and boundary closure by utilizing a fusion feature guide reconstruction network, and extracting real container number character structure features from the image. According to the invention, by introducing multi-scale disturbance perception, morphological fusion feature analysis, structure guide reconstruction and dynamic path regulation and control mechanisms, accurate extraction and stable identification of a real character structure in a complex corrosion interference environment are realized, and the accuracy and robustness of container number identification are improved.

Description

Technical Field

[0001] This invention relates to the field of train identification technology, specifically to a train car number, train container number, and train model identification device. Background Technology

[0002] Train car number, container number, and train model recognition based on Deep Convolutional Neural Networks (DCNNs) is an intelligent visual recognition method that uses artificial intelligence image recognition technology to automatically extract and classify train-related identification information. This technology primarily involves deploying high-speed cameras to collect image data during train passage. Image preprocessing techniques are used to denoise, enhance, and correct the original images to ensure image quality suitable for subsequent recognition. Based on this, a deep convolutional neural network model is used to automatically detect and locate key information regions in the image, such as car number, container number regions, and car body structural outlines. Character recognition networks (such as CRNN and CTC) are used to extract and recognize characters in the number regions, while a classification network is combined to identify the train model. Existing recognition systems of this type typically include the following core components: image acquisition and preprocessing, target detection and localization, feature extraction and classification, result verification and output, etc. Some systems also incorporate modules such as temporal image fusion, blurred image compensation, and multi-angle image reconstruction to improve recognition robustness in complex scenarios. Throughout the process, deep convolutional neural networks not only have powerful expressive capabilities in image feature extraction, but can also be trained end-to-end, enabling the recognition of train car numbers, container numbers, and train models to still have good generalization ability and high accuracy under different lighting, speed, angle, and dirt conditions, thereby meeting the application needs of multiple scenarios such as railway transportation management, cargo tracking, and intelligent scheduling.

[0003] The existing technology has the following shortcomings: In the process of recognizing container numbers using deep convolutional neural networks, when there is extensive oxidation and corrosion around the container number characters on the container surface, irregular textured patterns are formed. These textures are extremely similar in shape to the character outlines and have almost no significant difference in grayscale levels from the character boundary areas. Some even overlap the actual edges of the characters, making the boundary between the characters and the background unclear. Since deep convolutional neural networks rely on edge information in the image to extract the structural features of the characters, when the edges of the corrosion textures highly overlap with the character outlines, the network has difficulty distinguishing between real characters and background interference in the feature space. Especially when such corrosion interference is lacking in the training samples, the network is more likely to classify the entire area as an invalid region, thus ignoring or misjudging the real characters. Therefore, existing technologies for recognizing train car numbers, train container numbers, and train models based on deep convolutional neural networks cannot accurately extract and identify the true container number characters based on the morphological fusion features when the edges of the container number characters are covered by rusted graphics. This results in missing or incorrect recognition results, which in turn prevents the container numbers from being correctly identified and recorded by the system. This affects subsequent automatic scheduling, cargo counting, and logistics tracking processes, and may ultimately lead to a series of serious logistics management problems such as misloading, missing loading, and interruption of cargo flow.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a train car number, train container number, and train model identification device to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a train car number, train container number and vehicle type recognition device, including an edge disturbance perception module, a fusion feature analysis module, a structure guidance reconstruction module, a path evaluation and recognition module and a dynamic path control module; The edge perturbation sensing module calculates the edge stability of each character edge region in the input image by constructing a local perturbation sensing network, and identifies whether the edge of the box number character is covered by the rust pattern based on the boundary response perturbation frequency and structural entropy change. The fusion feature analysis module, based on the identified area where the character edges are covered by rusted graphics, uses multi-scale contrast map convolution to extract edge extension ambiguity, structural cohesion splitting rate and texture contour heterogeneity, and determines the morphological fusion features when the box number character edges are covered by rusted graphics. The structure-guided reconstruction module, based on the determined morphological fusion features, performs channel reordering and spatial adaptive magnification on the image. It uses the fusion feature-guided reconstruction network to enhance the coherence of character structure and boundary closure, and extracts the true box number character structure features from the image. The path evaluation and recognition module takes the extracted real container number character structure features as input to evaluate and guide the recognition path selection structure. It completes the recognition of real container number characters based on character continuity score and stroke closure score, and completes the character sequence output through jump decoding. The dynamic path control module divides the recognition results into three types based on the fusion degree vector and the extracted confidence score, and executes dynamic control of the recognition path for direct output, structural enhancement recognition, or character relocation and re-recognition respectively to complete the generation of the final recognition result.

[0007] Preferably, the edge disturbance sensing module is implemented as follows: A local perturbation sensing network is constructed by establishing a hierarchical sensing structure containing multi-scale convolution kernels and multi-directional perturbation kernels for each character edge region in the input image. During the convolution process, perturbation parameters with variable amplitudes are introduced to perform perturbation response mapping on the pixels in the character edge region, generating an edge response matrix containing multi-directional perturbation results for edge stability calculation. Statistical calculations are performed on the perturbation results of each character edge region based on the edge response matrix. The variance and mean difference of pixel response intensity under different perturbation conditions are calculated, and the edge stability value is defined by the inverse function of response volatility. An edge stability distribution map is constructed on the image plane to represent the structural stability of character edges under local perturbation. The boundary response perturbation frequency is calculated based on the edge stability distribution map, and a joint feature vector is generated by combining the structural entropy change. Through cluster analysis and threshold discrimination of the joint feature vector, a highly correlated distribution pattern between the boundary response perturbation frequency and the structural entropy change is identified. When the boundary response perturbation frequency increases and the structural entropy change increases synchronously, it is determined that the edge of the box number character is covered by the rust pattern in the input image.

[0008] Preferably, the implementation of the fusion feature parsing module is as follows: Based on the region where the identified character edges are covered by the rusted graphics, a multi-scale contrast map convolutional structure is constructed. By simultaneously applying directional variable contrast map convolutional kernels at multiple scales, edge response intensity maps and local background contrast maps are generated, and the response results at each scale are uniformly encoded into an edge structure response sequence. Convolution response characteristic analysis is performed on the edge structure response sequence. Edge extension ambiguity is calculated by local gradient change trend, structural cohesion splitting rate is calculated by statistical analysis of character structure continuity breakpoints, and texture contour heterogeneity is calculated by texture subspace separation degree between contour region and adjacent background region. These characteristics represent character edge ambiguity, structural integrity and texture difference, respectively. A three-dimensional fusion feature space is formed by combining edge extension ambiguity, structural cohesion splitting rate and texture contour heterogeneity. The dense area of ​​fusion features is determined by density clustering and distribution hierarchical comparison. The fusion parameter at the center of the dense area of ​​fusion features is used as the morphological fusion feature, thereby determining the morphological fusion feature when the edge of the box number character is covered by the rusted graphic.

[0009] Preferably, convolutional response characteristic analysis is performed on the edge structure response sequence. Edge extension ambiguity is calculated through local gradient change trends, structural cohesion splitting rate is calculated statistically through character structure continuity breakpoints, and texture contour heterogeneity is calculated through the texture subspace separation degree between the contour region and the adjacent background region. Specifically: The local gradient change trend of each pixel in the edge structure response sequence in the multi-scale convolutional layer is extracted, the gradient direction switching frequency and intensity fluctuation range between adjacent positions are calculated, and the gradient attenuation magnitude is obtained by comparative analysis between the continuous edge region and the blurred region, which is used to calculate the edge extension blur and represent the degree of transition of the character edge from clear to blurred. Based on the structural connectivity graph in the edge structure response sequence, the distribution density of breakpoints in the same structural path of the character graphic is statistically analyzed. Combining the projection position of the breakpoint on the structural principal axis and the distribution of continuous segment length, the continuity information of the character structure is extracted, the structural cohesion splitting rate is calculated, and the degree of fragmentation of the character outline in the spatial structure is represented. For the character contour region and the adjacent background region that have been divided in the edge structure response sequence, the texture feature vectors of the character contour region and the texture feature vectors of the adjacent background region are extracted within a fixed receptive field, respectively, and a high-dimensional texture feature distribution map is constructed. By calculating the difference between the mean vector Euclidean distance and the covariance matrix between the set of texture feature vectors of the character contour region and the set of texture feature vectors of the adjacent background region, the texture subspace separation degree is obtained, and the texture contour heterogeneity is quantified accordingly to characterize the difference in texture structure between the character boundary and the background.

[0010] Preferably, the implementation of the structure-guided reconstruction module is as follows: Based on the determined morphological fusion features, channel reordering is performed on the image input feature map. The importance score of each dimension feature in the morphological fusion features to the image channel is used as the reordering criterion to construct a feature priority sequence and rearrange the order of each channel of the feature map. At the same time, a spatial adaptive amplification operator based on local response density is applied to the rearranged feature map. The image feature map that has undergone channel reordering and spatial adaptive magnification is input into the fusion feature-guided reconstruction network. The morphological fusion features are encoded as structural enhancement factors through the fusion feature-guided pathway to participate in the feature convolution process, guiding the distribution of convolution weights to focus on regions with boundary continuity and structural consistency. Extract the backbone structure contour of the connected regions from the image features output by the fusion feature-guided reconstruction network. Combine structural connectivity analysis and boundary closure discrimination algorithms to filter out interference regions that do not have structural closure properties, and retain the region features with strong connectivity and closed boundaries in the character contour region to complete the extraction of the true box number character structure features.

[0011] Preferably, the path evaluation and identification module is implemented as follows: The extracted structural features of the real box number characters are input into the evaluation and guidance recognition path selection structure. By performing feature map channel compression and temporal feature expansion operations on the real box number character structural features, a spatiotemporal joint feature map containing stroke arrangement order, boundary closure attributes and local connection weights is formed. A path scoring unit is constructed in the evaluation and guidance recognition path selection structure to score the structural expressive ability of different path candidate sequences, which is used to initialize the recognition path. Based on the path candidate sequence output by the evaluation-guided identification path selection structure, a character continuity score and a stroke closure score model are constructed. The character continuity score is calculated by analyzing the stroke sequence connection stability and the cosine similarity of adjacent stroke directions in the structural features of the real box number characters. The stroke closure score is measured by the energy gradient intensity distribution and the rate of change of closure curvature in the identified contour closure path. The score results are fed back to the path selection structure as a weighting factor to complete the identification of real box number characters. The identified character probability distribution sequence is input into the jump decoding unit. Based on the probability of whitespace characters and the marking of repeated characters in the temporal features, the jump decoding unit performs a forward merging search by constructing a non-overlapping jump candidate point sequence, dynamically prunes the character sequence, removes redundant paths, and outputs the optimal character sequence to achieve the character sequence output.

[0012] Preferably, based on the path candidate sequence output by the evaluation-guided path selection structure, a character continuity score and a stroke closure score model are constructed. The character continuity score is calculated by analyzing the stroke sequence connection stability and the cosine similarity of adjacent stroke directions in the structural features of the actual box number characters. The stroke closure score is measured by identifying the energy gradient intensity distribution and the rate of change of closure curvature in the closed contour path. Specifically: Based on the path candidate sequence output by the evaluation-guided path selection structure, the stroke connection trajectory in each candidate path is extracted, the direction vector of each stroke connection trajectory is modeled, and the direction cosine similarity and connection point response stability between adjacent strokes are calculated. By statistically analyzing the connection continuity coefficient and structural jump probability between strokes, a character continuity scoring model is constructed to measure the overall arrangement coherence of character strokes. Based on the character continuity scoring model, the contour closure region in the path candidate sequence is identified, the pixel gradient response map and local curvature distribution map of the closed path are extracted, the average value of the energy gradient intensity and the standard deviation of the curvature change rate on the closed path are calculated, and the stroke closure scoring model is constructed to evaluate the degree of contour closure and edge closure consistency in each character candidate. The character continuity score and stroke closure score results are normalized and fused to generate a unified path score vector, which is fed back as a weighting factor to the evaluation guide recognition path selection structure. The recognition priority of the path candidate sequence is reordered, and the path with the most complete structure and the highest stroke coherence is selected based on the comprehensive score to complete the recognition output of the real box number character.

[0013] Preferably, the implementation of the dynamic path control module is as follows: Based on the fusion degree vector and the extracted confidence score, a fusion confidence mapping model is constructed to generate a fusion score value for each recognition result. The fusion score value is obtained by weighted summation of the features of each dimension of the fusion degree vector and the fusion of the extracted confidence score. Based on the preset fusion score threshold range, the recognition results are divided into three recognition path categories: direct output path, structure enhancement recognition path, and character relocation and re-recognition path. Recognition results with fusion score values ​​higher than the first threshold in the fusion score threshold range are included in the direct output path. Confidence values ​​are confirmed for the character recognition results in the direct output path and they are directly output as the final recognition results, ending the subsequent processing flow and ensuring the rapid completion and stable generation of high-confidence paths. Recognition results with fusion scores between the first and second thresholds are assigned to the structure enhancement recognition path. The image feature map is subjected to structure enhancement processing to improve the coherence of character structure and boundary closure before the recognition result is regenerated. Recognition results with fusion scores below the second threshold are assigned to the character relocation and re-recognition path. The character image region is relocated and image features are extracted. The structure reconstruction and confidence enhancement recognition processing are performed to generate new recognition results. The fusion score corresponding to the new recognition result is input into the fusion confidence mapping model to update the path classification criteria. The path division and recognition are performed iteratively until the final recognition result that meets the output conditions is generated.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention constructs a staged deep convolutional neural network recognition structure, including an edge perturbation perception module, a fusion feature parsing module, a structure-guided reconstruction module, a path evaluation and recognition module, and a dynamic path control module. This structure enables stable extraction and accurate recognition of character structural features in complex image scenarios where the edges of box number characters are extensively covered by oxidation and corrosion. Compared to traditional recognition methods that rely on edge sharpness and grayscale contrast, this technical solution introduces multi-dimensional analysis methods such as edge stability analysis, structural fuzzy modeling, and fusion feature reconstruction. This allows for accurate identification of the degree of structural interference in the corroded area and reconstruction of the character's main structure at the feature level. This ensures that the network's ability to perceive the true character boundaries during extraction is not affected by occlusion interference, improving the stability and robustness of character recognition in complex backgrounds.

[0015] 2. This invention introduces an adaptive recognition path strategy by fusing a confidence scoring mechanism with a dynamic path control structure. During the recognition process, different recognition paths are dynamically divided based on the degree of morphological fusion and the confidence score, achieving differentiated and accurate processing of the recognition output. This mechanism can flexibly perform structural enhancement or character relocation operations based on the scoring results when facing different interference intensities, incomplete structures, or feature occlusion. It also optimizes the recognition path selection by iteratively updating the fused score, thereby avoiding misjudgments and information omissions in high-interference scenarios. This recognition strategy improves the processing efficiency and adaptability of deep networks in unstructured scenarios, and enhances the recognition accuracy and practical deployment robustness of box number characters in complex rust and interference environments. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a schematic diagram of the modules for identifying train car numbers, train container numbers, and train type of vehicles according to the present invention. Detailed Implementation

[0018] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0019] This invention provides, for example Figure 1The train car number, train container number, and vehicle type recognition device shown include an edge disturbance perception module, a fusion feature analysis module, a structure guidance reconstruction module, a path evaluation and recognition module, and a dynamic path control module; The edge perturbation sensing module calculates the edge stability of each character edge region in the input image by constructing a local perturbation sensing network, and identifies whether the edge of the box number character is covered by the rust pattern based on the boundary response perturbation frequency and structural entropy change. In this embodiment, the edge disturbance sensing module is implemented as follows: A local perturbation sensing network is constructed by establishing a hierarchical sensing structure containing multi-scale convolution kernels and multi-directional perturbation kernels for each character edge region in the input image. During the convolution process, perturbation parameters with variable amplitudes are introduced to perform perturbation response mapping on the pixels in the character edge region, generating an edge response matrix containing multi-directional perturbation results for edge stability calculation. Constructing a local perturbation-aware network can be achieved by designing multi-layer convolutional structures in the character edge regions of the input image, with each layer embedding convolutional kernels of different scales and perturbation kernels of different directions. In practical implementation, three receptive field scales—small, medium, and large—can be used, each corresponding to a different convolutional kernel size. For example, a 3×3 convolutional kernel for detail detection, a 5×5 convolutional kernel for medium-scale feature extraction, and a 7×7 convolutional kernel for structure awareness can be used in parallel on the character edge regions. Simultaneously, perturbation kernels of specific directions, such as horizontal, vertical, and diagonal perturbations, are introduced into each convolutional kernel group. By introducing small offsets and intensity perturbations within the neighborhood of each pixel, response maps under different perturbation directions are generated. The responses of each type of perturbation are collected and integrated into a multi-channel edge response matrix, which reflects the trend of response intensity changes of character edges under different perturbation conditions. Subsequent analysis of this edge response matrix can effectively observe the stability of character edges under local perturbations, thus providing high-dimensional, multi-view data support for edge stability calculation. For example, if the edge region of a character exhibits similar response fluctuations under all perturbation directions, it indicates that the structure of the region is stable. Conversely, if the response differences in each direction are significant, there may be edge structure interference.

[0020] Local perturbation-aware networks (LPNs) are specific convolutional structures with high-resolution spatial perception capabilities. Their core function is to capture the response patterns of local image regions under various perturbation conditions. Multi-scale convolutional kernels are used to extract structural features within different receptive fields. Small-scale kernels focus on preserving edge details, while large-scale kernels capture the overall character outline and contrast relationships of large regions, enabling the network to comprehensively express both local details and global structure when processing character edges. Multi-directional perturbation kernels simulate image perturbations in different directions, giving the network directional sensitivity. This sensitivity is a crucial indicator of edge stability. For example, if a character edge shows significant response changes under horizontal perturbation but remains stable in the vertical direction, it indicates directional inconsistency in the structure of that region, potentially indicating rust texture interference. Hierarchical perceptual structures involve fusing and stacking convolutional responses of different scales and directions to form a hierarchical response map, allowing the network to capture response differences in character edge regions across multiple dimensions. This multi-dimensional response stacking structure not only enhances the network's structural perception capabilities but also lays an accurate response foundation for subsequent edge stability calculations and rust interference recognition.

[0021] Statistical calculations are performed on the perturbation results of each character edge region based on the edge response matrix. The variance and mean difference of pixel response intensity under different perturbation conditions are calculated, and the edge stability value is defined by the inverse function of response volatility. An edge stability distribution map is constructed on the image plane to represent the structural stability of character edges under local perturbation. When performing statistical calculations on character edge regions based on the edge response matrix, the response intensity of each pixel under all perturbation directions can be considered as a response sequence. Variance analysis and mean difference calculations are then performed on this sequence. Specifically, the response value set for each pixel is first extracted along the perturbation direction dimension. Then, the standard deviation and the difference between the maximum and minimum values ​​are calculated to reflect the degree of response fluctuation of the pixel under perturbation. A higher response volatility indicates inconsistent response of the pixel under different perturbations and poorer structural stability; conversely, a lower response volatility indicates a more stable edge structure. After calculating the response volatility of all pixels, the instability is further converted into a stability index by calculating its reciprocal, thus constructing an edge stability value map. Each pixel in the image corresponds to an edge stability value; a higher value indicates a more stable edge. Based on this, the stability values ​​of all edge regions are mapped onto the image plane to generate an edge stability distribution map, forming a stability visualization map for subsequent recognition processing. For example, in a region where the edges are blurred due to corrosion interference, the volatility of the response sequence will be significantly higher than in normal character regions, ultimately appearing as low-stability patches on the stability distribution map.

[0022] The edge response matrix is ​​a multi-channel tensor structure, where each channel represents the edge response value under a perturbation direction. The channel vector of each pixel in the matrix reflects its directional response consistency. By performing variance analysis on these channel vectors, the structural consistency of the pixel under different perturbations can be measured, while the mean difference is used to capture the extremes of response amplitude fluctuations. Combining these statistical indicators forms a composite parameter that measures the stability of edge pixels. To enhance the discriminative power of the stability indicator, an inverse function of response volatility is used for mapping, corresponding regions with larger fluctuations and poorer stability to smaller values, thus appearing as cool-colored areas in the stability distribution map, while edge regions with high stability correspond to high-value areas, i.e., a hot-colored distribution. This processing method has strong contrast and discriminability, which is beneficial for the subsequent recognition network to focus on stable regions during feature learning. Through the edge stability distribution map, the recognition system can accurately locate disturbed edge regions and provide a clear spatial reference for the discrimination of rust coverage.

[0023] The boundary response perturbation frequency is calculated based on the edge stability distribution map, and a joint feature vector is generated by combining the structural entropy change. Through cluster analysis and threshold discrimination of the joint feature vector, a highly correlated distribution pattern between the boundary response perturbation frequency and the structural entropy change is identified. When the boundary response perturbation frequency increases and the structural entropy change increases synchronously, it is determined that the edge of the box number character is covered by the rust pattern in the input image.

[0024] Based on the edge stability distribution map, the boundary response perturbation frequency can be calculated by traversing the edge regions of each character in the image and counting the number of abrupt changes in edge stability values ​​under adjacent perturbation directions. Specifically, a sliding window scans along the character edges, recording whether the stability value of each pixel exceeds a preset difference threshold under changing perturbation conditions; if it does, it is considered a boundary response perturbation. After mapping the perturbation frequency to an edge perturbation sensitivity map, a joint feature vector is generated by combining it with the structural entropy change within the image region. The structural entropy change is obtained by calculating the entropy difference in the image's gray-level co-occurrence matrix and is used to characterize the degree of disorder in the image texture. The perturbation frequency and structural entropy change are combined into a two-dimensional feature vector, and unsupervised clustering analysis is performed on the vector set of all edge regions, such as using density peak clustering algorithms, to determine which regions exhibit obvious co-occurrence patterns of high perturbation frequency and high entropy change. When the cluster center falls in the feature space with high perturbation frequency and increasing entropy value, the region can be identified as a typical rust-covered area. For example, when identifying a container number, if the edge area of ​​the character "C" frequently experiences abrupt changes in stability and the texture of that area exhibits highly chaotic variations, the system can determine that this is a location of rust interference.

[0025] Boundary response perturbation frequency measures the structural consistency of character edges under perturbation. It expresses the degree of edge structural damage by identifying the number of jumps in edge stability values ​​before and after perturbation. A higher frequency indicates a more unstable edge response, potentially affected by complex backgrounds or corrosion. Structural entropy change is a statistical feature reflecting the complexity of image texture; a higher entropy value indicates a more chaotic distribution of pixel grayscale within the region. Combining these two metrics allows for a joint description of whether character edges are covered by unstructured graphics: boundary response perturbation frequency reveals spatial structural instability, while structural entropy change reflects information disorder. Combining these two indicators to form a joint feature vector, and using cluster analysis to classify region patterns, can identify feature regions with shared high perturbation frequencies and high structural entropy, providing a basis for detecting corrosion coverage. By setting a threshold region for the joint feature vector, it is possible to determine whether character edges are covered by corrosion in an image, and further locate the interference location, providing a preliminary judgment signal for subsequent feature restoration and character reconstruction.

[0026] The fusion feature analysis module, based on the identified area where the character edges are covered by rusted graphics, uses multi-scale contrast map convolution to extract edge extension ambiguity, structural cohesion splitting rate and texture contour heterogeneity, and determines the morphological fusion features when the box number character edges are covered by rusted graphics. In this embodiment, the implementation of the fusion feature parsing module is as follows: Based on the region where the identified character edges are covered by the rusted graphics, a multi-scale contrast map convolutional structure is constructed. By simultaneously applying directional variable contrast map convolutional kernels at multiple scales, edge response intensity maps and local background contrast maps are generated, and the response results at each scale are uniformly encoded into an edge structure response sequence. To extract the difference between edge details and background interference in the region where the character edge is covered by the rusted texture, a multi-scale contrast map convolutional structure is first constructed in the identified rusted interference region. This structure deploys contrast map convolutional kernels with variable orientation in parallel across multiple receptive field scales. Image convolution operations are performed horizontally, vertically, and diagonally, extracting texture abrupt changes in different directions based on the character edge during the convolution process. At each scale, the convolutional kernel not only captures the edge response intensity of the character contour but also performs comparative analysis on the background grayscale fluctuations in the edge neighborhood, generating an edge response intensity map and a local background contrast map. By uniformly encoding the convolution results at different scales, an edge structure response sequence is formed for subsequent structural feature decomposition and fusion feature extraction. For example, when the character boundary and the rusted texture exhibit different response differences across multiple scales, this encoding sequence can capture its significant features in terms of edge continuity and contrast abrupt changes, assisting the subsequent network in effectively identifying the true character structure.

[0027] Multi-scale contrast map convolution is a feature extraction structure designed for complex boundary environments in image recognition tasks. It simultaneously perceives local details and global structural features by deploying multiple convolution channels of different sizes in parallel. Each scale corresponds to a receptive field size of the convolution kernel, ensuring that character edges can be effectively detected regardless of whether they are compact or diffuse. Variable-orientation contrast map convolution kernels introduce angle adjustment parameters to traditional static convolution kernels, allowing the convolution operation to dynamically adjust response weights according to changes in edge direction, thereby more accurately identifying edge transition relationships in different directions. This design is beneficial for addressing the directional blurring problem of rust textures in edge regions, improving the robustness of edge detection. The essence of contrast map convolution lies not only in convolving the image itself but also in simultaneously modeling the contrast relationships between regions in the image, thereby enhancing the difference between character regions and background regions, and making the subsequently extracted edge structure response sequence more discriminative and structurally coherent in the spatial level.

[0028] Convolution response characteristic analysis is performed on the edge structure response sequence. Edge extension ambiguity is calculated by local gradient change trend, structural cohesion splitting rate is calculated by statistical analysis of character structure continuity breakpoints, and texture contour heterogeneity is calculated by texture subspace separation degree between contour region and adjacent background region. These characteristics represent character edge ambiguity, structural integrity and texture difference, respectively. A three-dimensional fusion feature space is formed by combining edge extension ambiguity, structural cohesion splitting rate and texture contour heterogeneity. The dense area of ​​fusion features is determined by density clustering and distribution hierarchical comparison. The fusion parameter at the center of the dense area of ​​fusion features is used as the morphological fusion feature, thereby determining the morphological fusion feature when the edge of the box number character is covered by the rusted graphic.

[0029] To accurately determine the morphological fusion features when the edges of box number characters are covered by rusted graphics, a three-dimensional fusion feature space needs to be constructed by combining three key feature parameters: edge extension ambiguity, structural cohesion splitting rate, and texture contour heterogeneity. In this space, each character region corresponds to a three-dimensional coordinate point, with its three dimensions representing the numerical values ​​of edge ambiguity, structural fragmentation, and texture difference, respectively. By applying density clustering in this space, highly concentrated and closely spaced point sets in feature distribution can be effectively identified. These points reflect a class of character regions with highly similar interference features. Combining a distribution hierarchical comparison strategy, the aggregation trend and boundary gradient of each dense point set are compared at multiple scales to further confirm the most representative dense fusion feature regions. Finally, the three-dimensional feature value of the center point of the dense fusion feature region is used as the most representative morphological fusion feature parameter for that region to guide subsequent character recognition optimization.

[0030] The 3D fusion feature space is a feature space with edge extension ambiguity, structural cohesion splitting rate, and texture contour heterogeneity as three coordinate axes. It is used to comprehensively reflect the boundary, structural, and texture features of character regions after interference. Density clustering is a clustering method based on point density. By evaluating the local density of each point in the feature space, it automatically identifies clusters with highly similar features, thereby avoiding the need to pre-determine the number of categories and improving the model's ability to perceive abnormal interference regions. Distribution hierarchical comparison is a technique that performs hierarchical analysis of the boundary gradient, feature mean, and variance of each cluster at multiple scales, used to further verify the stability and effectiveness of the clustering results. The fusion feature dense region refers to a region in the feature space where the feature value distribution is highly concentrated, representing the typicality of character region features under specific interference morphology. By extracting its central feature value, a highly representative fusion feature with low noise influence can be obtained as the final expression result of the morphological fusion feature.

[0031] In this embodiment, convolutional response characteristic analysis is performed on the edge structure response sequence. The edge extension ambiguity is calculated through the local gradient change trend, the structural cohesion splitting rate is calculated through the statistical analysis of character structure continuity breakpoints, and the texture contour heterogeneity is calculated through the texture subspace separation degree between the contour region and the adjacent background region. Specifically: The local gradient change trend of each pixel in the edge structure response sequence in the multi-scale convolutional layer is extracted, the gradient direction switching frequency and intensity fluctuation range between adjacent positions are calculated, and the gradient attenuation magnitude is obtained by comparative analysis between the continuous edge region and the blurred region, which is used to calculate the edge extension blur and represent the degree of transition of the character edge from clear to blurred. This process analyzes the local gradient change trends of each pixel in the edge structure response sequence within multi-scale convolutional layers to measure the transition characteristics of character edges from sharp to blurred. Specifically, it extracts the gradient direction and intensity information within the local neighborhood of each pixel from the convolutional output at multiple scales, calculates the gradient direction switching frequency at adjacent positions across consecutive scales (i.e., the number of times the main gradient direction changes within a unit area), and records the fluctuation range of the gradient amplitude to quantify the edge stability. When the character edge is sharp, the gradient direction tends to be consistent and the amplitude variation is small; however, in the blurred transition region, the direction switching is frequent and the amplitude fluctuation is significant. By comparing the gradient decay amplitude between the continuous edge region and the blurred region, an intensity curve showing gradient decay with scale is obtained. This curve reflects the degree of edge sharpness loss and serves as the basis for calculating edge extension blur, characterizing the blur diffusion trend of character boundaries in spatial structure. The edge structure response sequence in this technology is a structural representation formed by encoding multi-scale convolutional responses. The local gradient change trend is used to measure the continuity of the directional features of pixels. The gradient direction switching frequency reflects the degree of structural perturbation. The intensity fluctuation range measures the change in edge sharpness. The gradient decay amplitude reflects the spatial transition of the boundary from solid to blurry by combining various indicators. Finally, a quantitative indicator is formed to characterize the degree of edge blurring, which helps to identify the true structural direction of character edges in the environment of corrosion interference.

[0032] Based on the structural connectivity graph in the edge structure response sequence, the distribution density of breakpoints in the same structural path of the character graphic is statistically analyzed. Combining the projection position of the breakpoint on the structural principal axis and the distribution of continuous segment length, the continuity information of the character structure is extracted, the structural cohesion splitting rate is calculated, and the degree of fragmentation of the character outline in the spatial structure is represented. This process measures the continuity of character graphics using a structural connectivity graph in the edge structure response sequence, aiming to identify whether characters exhibit structural damage or stroke breaks when occluded by corrosion. Specifically, it works as follows: First, a connectivity algorithm extracts the path connections of the character graphics from the response sequence, constructing a structural connectivity graph to identify the character's pixel-level structural extension trajectory. Then, on each structural path, breakpoints—nodes with interrupted connections or significantly reduced responses—are detected, and their distribution density is statistically analyzed to measure the overall structural fragmentation. Next, the positions of these breakpoints on the structural principal axis are projected onto a one-dimensional axis to determine whether the fractures are concentrated in critical stroke areas or edge areas. Combined with the length distribution of continuous structural segments, the stability of the character outline and stroke integrity are evaluated. These quantitative features construct character structural continuity information, and the structural cohesion splitting rate is calculated, representing the balance between the integrity and fragmentation of the character's spatial topology. Structural connectivity diagrams are used to reveal the actual connections between character strokes. The density of fracture point distribution is a key indicator reflecting the degree of structural damage. Projection position is used to determine whether the damaged area is located in a critical structural position. The length of continuous segments reveals the local stability of the character. The structural cohesion splitting rate integrates these factors into a unified metric, which helps to accurately identify whether the character shape has been damaged by corrosion under complex background interference.

[0033] For the character contour region and the adjacent background region that have been divided in the edge structure response sequence, the texture feature vectors of the character contour region and the texture feature vectors of the adjacent background region are extracted within a fixed receptive field, respectively, and a high-dimensional texture feature distribution map is constructed. By calculating the difference between the mean vector Euclidean distance and the covariance matrix between the set of texture feature vectors of the character contour region and the set of texture feature vectors of the adjacent background region, the texture subspace separation degree is obtained, and the texture contour heterogeneity is quantified accordingly to characterize the difference in texture structure between the character boundary and the background.

[0034] To measure the degree of texture difference between a character outline region and its neighboring background region, the spatial boundaries of these regions are first delineated based on the edge structure response sequence, and a fixed receptive field is set for each type of region. The fixed receptive field refers to a predefined local area of ​​a certain size in the image feature map, used to unify the scale of texture feature extraction and ensure the comparability of feature extraction between different regions. Within the receptive field, a pre-trained texture feature extraction network extracts texture feature vectors from the character outline region and the background region, including features such as texture direction distribution, detail density, and texel arrangement statistics. Subsequently, the extracted texture feature vector sets of the two types of regions are mapped to a unified high-dimensional texture feature space, and their distribution in this space is plotted to observe the degree of separation in feature distribution between the two types of regions. The Euclidean distance between the mean vectors of the texture feature sets of the character outline region and the texture feature sets of the neighboring background region is calculated to evaluate the overall texture center difference between the two types of regions; simultaneously, the difference between the covariance matrices of the two sets is calculated to reflect the distribution difference of their internal texture structures. These two indicators are combined to construct the texture subspace separation degree. This separation metric is quantified as texture contour heterogeneity, used to characterize the strength of the distinction between character boundaries and the background in terms of texture characteristics. In this way, situations where the texture boundary between character boundaries and the background becomes blurred or uniform due to corrosion coverage can be effectively identified, providing a strong basis for subsequent character structure extraction and recognition.

[0035] The structure-guided reconstruction module, based on the determined morphological fusion features, performs channel reordering and spatial adaptive magnification on the image. It uses the fusion feature-guided reconstruction network to enhance the coherence of character structure and boundary closure, and extracts the true box number character structure features from the image. In this embodiment, the structure-guided reconstruction module is implemented as follows: Based on the determined morphological fusion features, channel reordering is performed on the input feature map of the image. The importance score of each dimension feature in the morphological fusion features to the image channel is used as the reordering criterion to construct a feature priority sequence and rearrange the order of each channel of the feature map. At the same time, a spatial adaptive amplification operator based on local response density is applied to the rearranged feature map to improve the response intensity of the fusion region in the spatial dimension and enhance the image's local perception ability of character structure. When performing channel reordering on the input feature map of an image based on defined morphological fusion features, a correlation matrix between the fusion features and each channel of the image can be constructed first. This matrix calculates the dot product similarity or mutual information index between each dimension of the fusion features and the feature map of each channel, yielding an importance score for each channel. This score result is used to establish a channel weight ranking list, forming a feature priority sequence, with high-priority channels prioritized to enhance the expression intensity of key character regions. Subsequently, to improve the spatial response of character structure regions, a spatial adaptive magnification operator based on local response density needs to be applied to the channel-reordered feature map. This operator dynamically adjusts the sampling stride of the convolution kernel and the magnification weights according to the response density of local regions in the image, thereby performing enhanced convolution in regions with strong character edge continuity, achieving perceptual weighting of locally salient regions. For example, the lower edge of the character "U" is significantly blurred under rust interference; the system will assign a higher magnification at this location to enhance the recognition salience of the character structure.

[0036] Channel reordering is a channel arrangement method based on feature importance evaluation. Its core lies in breaking the limitation of traditional fixed-order feature map transmission. By fusing the interaction between features and channels, it reconstructs the channel arrangement order, enhancing the specificity of feature expression. The feature priority sequence is an ordered list arranged according to the strength of correlation between channels and fused features. This sequence serves as the basis for channel reordering, prioritizing the transmission of channels that contribute the most to structural information. The spatial adaptive amplification operator based on local response density is a dynamic convolution control mechanism. It judges the feature activity level by statistically analyzing the density index of response points within local regions, applying high-weight amplification to regions with high response density. This achieves non-linear enhancement at the spatial scale during convolution, thereby improving the network's ability to capture the boundary coherence of character structures and local contour details. This mechanism is highly effective in handling character edge blurring caused by rust coverage.

[0037] The image feature map that has undergone channel reordering and spatial adaptive magnification is input into the fusion feature-guided reconstruction network. The morphological fusion features are encoded as structural enhancement factors through the fusion feature-guided path to participate in the feature convolution process, guiding the distribution of convolution weights to focus on the region of boundary continuity and structural consistency, thereby achieving an enhanced response of character structural coherence and boundary closure. After channel reordering and spatial adaptive magnification, the image feature map is input into the fusion feature-guided reconstruction network. Morphological fusion features are introduced as guiding information to regulate the convolution process. Specifically, the 3D morphological fusion features are encoded as a structure enhancement factor tensor. This tensor is embedded as a conditional input into the convolutional layers of the backbone feature pathway. Through element-wise weighting or channel-level adjustment with the weight tensor, the direction of the convolutional kernel response and the weight distribution are guided. In each layer, the receptive field concentration area is dynamically adjusted according to the excitation intensity of the structure enhancement factor, guiding the network to focus on the closed edge regions and continuous stroke regions of characters. For example, when recognizing a rusted and blurred "E" character, the structure enhancement factor will cause the convolutional kernel to enhance the response in the broken edge region on the right side of the character, thereby completing the character structure and achieving the coherence restoration and boundary closure reconstruction of the real character.

[0038] Feature-guided reconstruction networks are deep neural architectures that combine fusion perception with structural completion mechanisms. Their core feature lies in using external fusion features as guiding factors to directly participate in the network's parameter regulation process. Morphological fusion features, mapped to structural enhancement factor tensors, are cross-fused with the main feature flow in either the spatial or channel dimensions to adjust the receptive center of gravity and response threshold of each convolutional operation. This structure not only retains the local extraction capabilities of traditional convolutions but also dynamically optimizes the response direction of convolutions through guiding paths, enabling the network to adaptively concentrate feature representation capabilities in regions where character structure is easily lost. Structural enhancement factors can be embedded into the convolutional kernels through a gating attention mechanism, making their response regions more closely match the structural connectivity paths and bounding box closure regions of the characters. This effectively restores the overall shape and edge clarity of characters when faced with complex background interference and rust coverage, improving recognition accuracy and robustness.

[0039] Extract the backbone structure contour of the connected regions from the image features output by the fusion feature-guided reconstruction network. Combine structural connectivity analysis and boundary closure discrimination algorithms to filter out interference regions that do not have structural closure properties, and retain the region features with strong connectivity and closed boundaries in the character contour region to complete the extraction of the true box number character structure features.

[0040] The backbone structure contours of connected regions are extracted from the image features output by the feature-guided reconstruction network. First, edge detection algorithms such as the multi-scale Canny operator are used for preliminary edge extraction of image features. Then, a connected graph model of pixels in the edge map is constructed using a structural connectivity analysis framework. In this model, continuous stroke paths are identified by setting a minimum connected length threshold and a structural path integrity scoring mechanism. Next, a boundary closure discrimination algorithm is used to identify whether a complete closed boundary structure is formed. Specifically, this involves calculating the Euclidean distance between the start and end points of the contour path, the area ratio of the closed contour region, and the boundary gradient difference between the closed contour region and the surrounding background region. When a contour region is determined to have significant structural connectivity and boundary closure, that region is retained as a true character structure region, while other regions that do not meet the conditions are filtered out. For example, when processing a "3" character whose strokes are partially covered by rust, only structurally connected and boundary-closed regions are filtered, effectively eliminating false structural textures generated by rust.

[0041] The backbone structure outline of a connected region refers to a connected structural region possessing backbone stroke characteristics in the character morphological features, typically manifested as a set of edge paths with directional continuity and a certain degree of thickness stability. Structural connectivity analysis is a connectivity evaluation method based on graph structure modeling. It constructs a connected subgraph by defining connection rules and jump tolerance between pixels, thereby determining whether a region belongs to a continuous character structure. Boundary closure discrimination algorithms are structural detection techniques used to verify shape integrity, typically including boundary tracking, closure determination, and topological integrity verification. Their purpose is to ensure that the identified character regions do not contain broken, hollow, or open structures. Through the joint determination of these two types of analyses, not only can the complete outline of the character be accurately identified, but also non-character connected regions caused by background textures, oxidation marks, or image compression artifacts can be effectively eliminated, ensuring that the final extracted box number character structural features are real and of high confidence.

[0042] The path evaluation and recognition module takes the extracted real container number character structure features as input to evaluate and guide the recognition path selection structure. It completes the recognition of real container number characters based on character continuity score and stroke closure score, and completes the character sequence output through jump decoding. In this embodiment, the path evaluation and identification module is implemented as follows: The extracted structural features of the real box number characters are input into the evaluation and guidance recognition path selection structure. By performing feature map channel compression and temporal feature expansion operations on the real box number character structural features, a spatiotemporal joint feature map containing stroke arrangement order, boundary closure attributes and local connection weights is formed. A path scoring unit is constructed in the evaluation and guidance recognition path selection structure to score the structural expressive ability of different path candidate sequences, which is used to initialize the recognition path. In the path initialization phase, the structural features of the actual box number characters are first input into the evaluation-guided path selection structure. Channel compression is performed on the input feature map to reduce redundant information and retain highly expressive features, while simultaneously unfolding the inherent temporal series features to reflect the temporal relationship of the character structure during stroke evolution. Based on this, a spatiotemporal joint feature map is constructed, incorporating stroke arrangement order, boundary closure attributes, and local connectivity weights. This fusion representation of high-dimensional joint features not only preserves the spatial structure of the character's form but also expresses the structural evolution logic of the character in visual order. This feature map is then fed into the evaluation-guided path selection structure, which embeds a path scoring unit. The path scoring unit comprehensively models the structural continuity, boundary closure, and stroke connectivity of candidate paths, calculates a score based on the activation pattern of each path on the feature map, and uses the scoring results to select the path sequence with the most complete structural expression and strongest visual rationality as the initialization basis for the recognition path.

[0043] The evaluation-guided recognition path selection structure is a path initialization framework that incorporates a multi-branch scoring mechanism. Its core function is to score the representation completeness of different paths based on the mapping relationship between the structural features of the actual box number characters and the candidate paths. This structure sets up a path scoring unit as a decision module, which contains multiple scoring branches used to model the consistency of stroke arrangement order, the completeness of boundary closure attributes, and the structural weights of local connections. The channel compression part is mainly used to reduce the channel dimension without losing important discriminative information, thereby improving the model's computational efficiency. The temporal feature expansion operation explicitly encodes the writing order of character strokes in the feature representation through an arrangement time window, enabling the network to capture the natural sequential logic of character composition. The spatiotemporal joint feature map serves as a unified data structure, supporting the path scoring unit in aligning and comparing different candidate paths, thus achieving recognition pre-selection in the path initialization stage and ensuring that subsequent recognition processes are based on structurally stable paths.

[0044] Based on the path candidate sequence output by the evaluation-guided identification path selection structure, a character continuity score and a stroke closure score model are constructed. The character continuity score is calculated by analyzing the stroke sequence connection stability and the cosine similarity of adjacent stroke directions in the structural features of the real box number characters. The stroke closure score is measured by the energy gradient intensity distribution and the rate of change of closure curvature in the identified contour closure path. The score results are fed back to the path selection structure as a weighting factor to complete the identification of real box number characters. The identified character probability distribution sequence is input into the jump decoding unit. Based on the probability of whitespace characters and the marking of repeated characters in the temporal features, the jump decoding unit performs a forward merging search by constructing a non-overlapping jump candidate point sequence, dynamically prunes the character sequence, removes redundant paths, and outputs the optimal character sequence to achieve the character sequence output.

[0045] The skip decoding unit processes the recognized character probability distribution sequence to select the optimal output path from multiple possible character sequences. First, the unit receives a time-ordered character probability distribution sequence, with each time step containing the predicted probabilities of all character categories (including whitespace). During analysis, the skip decoding unit detects positions with high whitespace probabilities, which typically represent transitional states between characters or invalid output. Simultaneously, the skip decoding unit identifies repeating character markers to determine if the same character appears repeatedly in consecutive time steps. Based on whitespace and repeating character positions, the system constructs a set of non-overlapping skip candidate point sequences. Subsequently, a forward merging search algorithm dynamically searches for paths on these candidate points, performing path merging based on probability accumulation and structural consistency. A dynamic pruning strategy is incorporated to remove candidate paths that do not meet probability thresholds or lack structural coherence in real time, effectively reducing redundant recognition paths. Finally, a character sequence with a reasonable structure and the highest probability is output, achieving accurate generation of the character sequence output. For example, when recognizing "C1234", the model may output multiple redundant characters, such as "CC112334". By skipping decoding, the redundant characters and whitespace can be skipped, and "C1234" can be accurately restored.

[0046] The skip decoding unit is a decoding module that performs post-processing based on temporal recognition results. Its core functions are path selection and sequence deduplication. The probability of whitespace characters in the temporal features reflects whether the current position is within an invalid segment between characters on the time axis. Repeated character markers are used to identify redundant outputs of the same character at consecutive time points. These two types of information provide the basis for constructing skip candidate points. The non-overlapping skip candidate point sequence is constructed by selecting discontinuous but structurally compatible positions in the time dimension, ensuring that the decoding path does not repeat or cross-interfere. Forward merging search is a dynamic search strategy along the time axis, aiming to construct the path with the optimal cumulative probability among skip candidate points and achieve efficient fusion of path branches. The dynamic pruning mechanism sets a real-time scoring threshold to immediately delete paths with low scores or that do not meet structural constraints, improving path decision-making efficiency and avoiding error accumulation. The synergistic work of these mechanisms improves computational efficiency and recognition robustness while maintaining accuracy, making it particularly suitable for handling complex recognition tasks with blurred character structures and frequent repetitive predictions against a background of corrosion interference.

[0047] In this embodiment, based on the path candidate sequence output by the evaluation-guided path selection structure, a character continuity score and a stroke closure score model are constructed. The character continuity score is calculated by analyzing the stroke sequence connection stability and the cosine similarity of adjacent stroke directions in the structural features of the actual box number characters. The stroke closure score is measured by identifying the energy gradient intensity distribution and the rate of change of closure curvature in the closed contour path. Specifically: Based on the path candidate sequence output by the evaluation-guided path selection structure, the stroke connection trajectory in each candidate path is extracted, the direction vector of each stroke connection trajectory is modeled, and the direction cosine similarity and connection point response stability between adjacent strokes are calculated. By statistically analyzing the connection continuity coefficient and structural jump probability between strokes, a character continuity scoring model is constructed to measure the overall arrangement coherence of character strokes. When analyzing candidate path sequences, the first step is to extract stroke connection trajectories from each candidate path, i.e., spatial connection clues between consecutive strokes. These stroke connection trajectories are then modeled using direction vectors, where the geometric direction of each stroke is represented by its principal axis direction vector. Next, cosine similarity is calculated between the direction vectors of adjacent strokes to determine the continuity of the strokes in direction; the closer the cosine value is to one, the more consistent the stroke directions. Furthermore, response stability is calculated at each stroke connection point, evaluating the stability of the connection point by analyzing the change in activation intensity of the feature map at that location. Then, the connection continuity coefficients throughout the entire stroke sequence are statistically analyzed, including the average direction similarity and stability score. This is combined with the detection of structural jump points in the path, such as abrupt changes in direction or regions of sudden drops in connection strength, and the probability of structural jumps is calculated accordingly. Finally, these values ​​are synthesized into a character continuity scoring model to evaluate whether the strokes in the entire candidate path exhibit a highly coherent structural expression. The goal of this model is to distinguish between true character strokes and incorrectly connected strokes, selecting structurally stable character recognition paths by accurately measuring the directional consistency and connection reliability of stroke arrangement. For example, in the character "H", if the middle vertical stroke is stably connected to the two vertical strokes on the left and right and their directions are similar, the score will be higher than if the stroke direction changes abruptly or the connection is unclear, thus increasing the path priority and enhancing the final recognition accuracy.

[0048] Based on the character continuity scoring model, the contour closure region in the path candidate sequence is identified, the pixel gradient response map and local curvature distribution map of the closed path are extracted, the average value of the energy gradient intensity and the standard deviation of the curvature change rate on the closed path are calculated, and the stroke closure scoring model is constructed to evaluate the degree of contour closure and edge closure consistency in each character candidate. When constructing a stroke closure scoring model, the first step is to identify character contour regions with a closing tendency in the candidate path sequence. Closed paths can be located using contour tracking algorithms, for example, extracting pixel trajectories that form a closed loop around the central region in characters like "O", "D", or "8". Next, the pixel gradients in the closed path are analyzed to generate a pixel gradient response map, which measures the intensity change of the edge. Simultaneously, the local curvature of each point on the closed path is extracted to form a curvature distribution map, used to characterize the smoothness and curvature of the path. Then, the average intensity value of the energy gradient on the closed path is statistically analyzed to reflect the salience of the edge contour; the standard deviation of the rate of change of curvature reflects the geometric fluctuation of the contour. A higher energy gradient indicates a clearer boundary, while a smaller rate of change of curvature indicates a smoother and more closed contour. Combining these two indicators, a closure scoring model is established to quantitatively evaluate the closure integrity of the contour in each candidate character. This scoring result can be used to determine whether a character possesses closed edge attributes, thereby excluding candidate paths with broken structures or incomplete strokes. For example, when recognizing the characters "C" and "G", if the model detects insufficient closure of the right edge and significant fluctuations in curvature, the system will assign a lower closure score, thereby increasing the likelihood of accurately recognizing "C". This method can improve the accuracy of closed character recognition, and is particularly suitable for scenarios where localized corrosion interference in the image causes edge defects.

[0049] The character continuity score and stroke closure score results are normalized and fused to generate a unified path score vector, which is fed back as a weighting factor to the evaluation guide recognition path selection structure. The recognition priority of the path candidate sequence is reordered, and the path with the most complete structure and the highest stroke coherence is selected based on the comprehensive score to complete the recognition output of the real box number character.

[0050] When normalizing and fusing the character continuity score and stroke closure score, the two scoring indicators first need to be standardized on a numerical scale to ensure they fall within the same evaluation range. Common methods include min-max normalization or Z-score standardization, thus ensuring comparability in the subsequent weighting process. The normalized continuity score reflects the coherence between strokes, while the closure score reflects the integrity of character structural boundaries. These two standardized scores are weighted and synthesized using set weighting coefficients to form a unified path score vector. This path score vector is input as a weighting factor into the evaluation-guided recognition path selection structure, driving the path scoring mechanism to prioritize all candidate paths. Path ranking strategies can employ weighted average or soft maximum fusion functions to highlight paths with better scores. Finally, based on the fused score results, the path with the most complete structure and the highest stroke coherence is selected for real character output recognition. For example, among multiple character candidates, if a path has a continuous stroke order but poor boundary closure, it will be placed in a lower position due to a lower overall score. Conversely, if a path scores highly on both metrics, its recognition priority will be increased, thereby enhancing the system's robust recognition capability for characters affected by corrosion. This scoring fusion strategy effectively improves the accuracy and robustness of path selection while ensuring the continuity of character structure and the integrity of boundaries.

[0051] The dynamic path control module divides the recognition results into three types based on the fusion degree vector and the extracted confidence score, and executes dynamic control of the recognition path for direct output, structural enhancement recognition, or character relocation and re-recognition respectively to complete the generation of the final recognition result.

[0052] In this embodiment, the dynamic path control module is implemented as follows: Based on the fusion degree vector and the extracted confidence score, a fusion confidence mapping model is constructed to generate a fusion score value for each recognition result. The fusion score value is obtained by weighted summation of the features of each dimension of the fusion degree vector and the fusion of the extracted confidence score. Based on the preset fusion score threshold range, the recognition results are divided into three recognition path categories: direct output path, structure enhancement recognition path, and character relocation and re-recognition path. The fusion degree vector and the extracted confidence score are used to jointly measure the structural feature integrity of the recognition result and the model's recognition confidence level. By constructing a fusion confidence mapping model, these two indicators are weighted and fused to generate a unified fusion score. The fusion degree vector includes fusion features across multiple dimensions, such as character edge blurring, structural fragmentation rate, and texture heterogeneity, reflecting the degree of interference with the character. The extracted confidence score originates from the recognition model's probability output of the current recognition result, measuring the model's credibility with the current recognition path. In the fusion confidence mapping model, the fusion degree vector is first normalized in dimensions. Then, a weighting function is introduced to assign weights to each dimension feature according to its recognition relevance. This weighted result is then fused with the confidence score to calculate a fusion score. Taking the recognition of a single character in an image as an example, when the character boundary blurring feature value is high, the structural fragmentation rate is high, and the model confidence is low, the fusion score will decrease, indicating significant uncertainty in the recognition result.

[0053] The fusion confidence mapping model is a scoring function model that integrates structural fusion features with recognition confidence indicators. It is typically constructed using a weighted linear mapping or a small feedforward neural network. This model receives a fusion degree vector and an extracted confidence score as input, and outputs a single fusion score value. This score value is used to determine the quality of the recognition result and guide the classification of subsequent recognition paths. A preset fusion score threshold range defines the criteria for dividing three recognition paths, with two critical values: when the fusion score value is higher than the first threshold, it indicates that the character structure is complete and the recognition confidence is high, allowing direct output; when the fusion score value is between the two thresholds, it indicates that the character structure is partially damaged and the confidence is moderate, requiring structural enhancement recognition; when the fusion score value is lower than the second threshold, it indicates that the character information is severely missing or the probability of recognition error is high, requiring character relocation and re-recognition processing. This scoring mechanism enables the recognition process to have dynamic adjustment capabilities, adaptively switching recognition paths under different recognition quality conditions, thereby improving the overall robustness and accuracy of recognition.

[0054] Recognition results with fusion score values ​​higher than the first threshold in the fusion score threshold range are included in the direct output path. Confidence values ​​are confirmed for the character recognition results in the direct output path and they are directly output as the final recognition results, ending the subsequent processing flow and ensuring the rapid completion and stable generation of high-confidence paths. When the fusion score exceeds the first threshold within the fusion score threshold range, the character recognition result can be determined to have reached a high level in both structural integrity and model confidence, making direct output feasible. Therefore, the recognition result is assigned to the direct output path, and a confidence value confirmation operation is performed within this path. This involves verifying the category probability distribution of the currently recognized character to ensure that the highest probability output item is highly consistent with the actual character category and that no significant ambiguity interference occurs. During this process, a one-dimensional maximum probability decision strategy can be used to make the final judgment on each character position, while confidence detection is performed on low-probability output items located in adjacent character categories to eliminate potential interference. For example, in a container number image, if the fusion score of the 5th character is 0.93, higher than the first threshold of 0.85, and the category confidence of this character reaches 97%, this character can be immediately used as the recognition output without further enhancement or relocation processing, thereby accelerating the processing flow and ensuring result stability.

[0055] The first threshold in the fusion scoring threshold range is a key parameter in the dynamic path control mechanism, used to define the judgment boundary of the direct output path. Its value is usually determined by combining the fault tolerance requirements of the recognition scene and the recognition confidence distribution obtained statistically during model training. This threshold represents the upper bound of the joint score of structural integrity and recognition confidence. A score value higher than this threshold usually means that there is no significant corrosion, occlusion, or deformation in the recognition area, and that the model has sufficient confidence in the current result. In implementation, the first threshold can be selected by performing fusion scoring statistics on multiple labeled samples and selecting the scoring position corresponding to a recognition accuracy of over 95%. This ensures that high-quality recognition results are quickly assigned to the direct path in practical applications, reducing redundant computational pressure and improving the overall system processing efficiency and response speed.

[0056] Recognition results with fusion scores between the first and second thresholds are assigned to the structure enhancement recognition path. The image feature map is subjected to structure enhancement processing to improve the coherence of character structure and boundary closure before the recognition result is regenerated. Recognition results with fusion scores below the second threshold are assigned to the character relocation and re-recognition path. The character image region is relocated and image features are extracted. The structure reconstruction and confidence enhancement recognition processing are performed to generate new recognition results. The fusion score corresponding to the new recognition result is input into the fusion confidence mapping model to update the path classification criteria. The path division and recognition are performed iteratively until the final recognition result that meets the output conditions is generated.

[0057] When the fusion score of the recognition result falls between the first and second thresholds, it indicates that the recognition result is in an intermediate state in terms of structural clarity and recognition confidence. It neither meets the high confidence requirement for direct output nor is it low enough to require relocation and re-recognition. Therefore, it is categorized into the structure enhancement recognition path. In this path, a structure enhancement processing module is introduced onto the original image feature map to strengthen the contour continuity and boundary closure of characters, improving the overall geometric stability of the characters. For example, morphological fusion features are used to guide regions in the feature map with structural breaks or blurred boundaries to perform enhanced convolution operations, allowing blurred connection points to obtain higher boundary responses during convolution, thus regenerating a character image with stronger structural integrity. After processing, recognition is performed again, and the fusion score is updated by combining the confidence value of the enhanced output. If the fusion score rises above the first threshold, it is directly output; if it falls below the second threshold, the path switches to character relocation and re-recognition.

[0058] When the fusion score is below the second threshold, it indicates severe interference or missing character structures, and the model lacks basic confidence in the recognition result. This situation is categorized into the character relocation and re-recognition path. In this path, the character image region is first re-detected and relocated, potential interference areas are removed, and new image features are extracted. Then, a structural reconstruction operation is performed to rebuild the occluded character structure, and a confidence enhancement mechanism is used to improve the model's confidence level in the recognition result. This enhancement mechanism may include feature channel enhancement, attention allocation optimization, and confidence supervision feedback. Finally, the fusion score of the new recognition result is re-input into the fusion confidence mapping model to determine if it meets the output standard. If the new score still does not meet the output requirements, the path division and recognition process is repeated, forming a dynamic self-looping recognition mechanism until the final character recognition result that meets structural integrity and confidence reliability is output. The first and second thresholds represent the score boundaries for direct output and mandatory re-recognition, respectively. The intermediate range is used to identify ambiguous areas. This three-value division ensures that the system maintains recognition accuracy while possessing flexible path adjustment capabilities.

[0059] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0060] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0061] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0062] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0063] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0064] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0065] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A train car number, train container number, and train type identification device, characterized in that, It includes an edge disturbance perception module, a fusion feature parsing module, a structure-guided reconstruction module, a path evaluation and identification module, and a dynamic path control module; The edge perturbation sensing module calculates the edge stability of each character edge region in the input image by constructing a local perturbation sensing network, and identifies whether the edge of the box number character is covered by the rust pattern based on the boundary response perturbation frequency and structural entropy change. The fusion feature analysis module, based on the identified area where the character edges are covered by rusted graphics, uses multi-scale contrast map convolution to extract edge extension ambiguity, structural cohesion splitting rate and texture contour heterogeneity, and determines the morphological fusion features when the box number character edges are covered by rusted graphics. The structure-guided reconstruction module, based on the determined morphological fusion features, performs channel reordering and spatial adaptive magnification on the image. It uses the fusion feature-guided reconstruction network to enhance the coherence of character structure and boundary closure, and extracts the true box number character structure features from the image. The path evaluation and recognition module takes the extracted real container number character structure features as input to evaluate and guide the recognition path selection structure. It completes the recognition of real container number characters based on character continuity score and stroke closure score, and completes the character sequence output through jump decoding. The dynamic path control module divides the recognition results into three types based on the fusion degree vector and the extracted confidence score, and executes dynamic control of the recognition path for direct output, structural enhancement recognition, or character relocation and re-recognition respectively to complete the generation of the final recognition result.

2. The train car number, train container number, and train model identification device according to claim 1, characterized in that, The implementation of the edge disturbance sensing module is as follows: A local perturbation sensing network is constructed by establishing a hierarchical sensing structure containing multi-scale convolution kernels and multi-directional perturbation kernels for each character edge region in the input image. During the convolution process, perturbation parameters with variable amplitudes are introduced to perform perturbation response mapping on the pixels in the character edge region, generating an edge response matrix containing multi-directional perturbation results for edge stability calculation. Statistical calculations are performed on the perturbation results of each character edge region based on the edge response matrix. The variance and mean difference of pixel response intensity under different perturbation conditions are calculated, and the edge stability value is defined by the inverse function of response volatility. An edge stability distribution map is constructed on the image plane to represent the structural stability of character edges under local perturbation. The boundary response perturbation frequency is calculated based on the edge stability distribution map, and a joint feature vector is generated by combining the structural entropy change. Through cluster analysis and threshold discrimination of the joint feature vector, a highly correlated distribution pattern between the boundary response perturbation frequency and the structural entropy change is identified. When the boundary response perturbation frequency increases and the structural entropy change increases synchronously, it is determined that the edge of the box number character is covered by the rust pattern in the input image.

3. The train car number, train container number, and train model identification device according to claim 1, characterized in that, The implementation of the fusion feature parsing module is as follows: Based on the region where the identified character edges are covered by the rusted graphics, a multi-scale contrast map convolutional structure is constructed. By simultaneously applying directional variable contrast map convolutional kernels at multiple scales, edge response intensity maps and local background contrast maps are generated, and the response results at each scale are uniformly encoded into an edge structure response sequence. Convolution response characteristic analysis is performed on the edge structure response sequence. Edge extension ambiguity is calculated by local gradient change trend, structural cohesion splitting rate is calculated by statistical analysis of character structure continuity breakpoints, and texture contour heterogeneity is calculated by texture subspace separation degree between contour region and adjacent background region. These characteristics represent character edge ambiguity, structural integrity and texture difference, respectively. A three-dimensional fusion feature space is formed by combining edge extension ambiguity, structural cohesion splitting rate and texture contour heterogeneity. The dense area of ​​fusion features is determined by density clustering and distribution hierarchical comparison. The fusion parameter at the center of the dense area of ​​fusion features is used as the morphological fusion feature, thereby determining the morphological fusion feature when the edge of the box number character is covered by the rusted graphic.

4. The train car number, train container number, and train model identification device according to claim 3, characterized in that, Convolutional response characteristic analysis is performed on the edge structure response sequence. Edge extension blurriness is calculated through local gradient change trends, structural cohesion splitting rate is calculated through character structure continuity breakpoints, and texture contour heterogeneity is calculated through texture subspace separation between the contour region and the adjacent background region. Specifically: The local gradient change trend of each pixel in the edge structure response sequence in the multi-scale convolutional layer is extracted, the gradient direction switching frequency and intensity fluctuation range between adjacent positions are calculated, and the gradient attenuation magnitude is obtained by comparative analysis between the continuous edge region and the blurred region, which is used to calculate the edge extension blur and represent the degree of transition of the character edge from clear to blurred. Based on the structural connectivity graph in the edge structure response sequence, the distribution density of breakpoints in the same structural path of the character graphic is statistically analyzed. Combining the projection position of the breakpoint on the structural principal axis and the distribution of continuous segment length, the continuity information of the character structure is extracted, the structural cohesion splitting rate is calculated, and the degree of fragmentation of the character outline in the spatial structure is represented. For the character contour region and the adjacent background region that have been divided in the edge structure response sequence, the texture feature vectors of the character contour region and the texture feature vectors of the adjacent background region are extracted within a fixed receptive field, respectively, and a high-dimensional texture feature distribution map is constructed. By calculating the difference between the mean vector Euclidean distance and the covariance matrix between the set of texture feature vectors of the character contour region and the set of texture feature vectors of the adjacent background region, the texture subspace separation degree is obtained, and the texture contour heterogeneity is quantified accordingly to characterize the difference in texture structure between the character boundary and the background.

5. The train car number, train container number, and train model identification device according to claim 1, characterized in that, The implementation of the structure-guided refactoring module is as follows: Based on the determined morphological fusion features, channel reordering is performed on the image input feature map. The importance score of each dimension feature in the morphological fusion features to the image channel is used as the reordering criterion to construct a feature priority sequence and rearrange the order of each channel of the feature map. At the same time, a spatial adaptive amplification operator based on local response density is applied to the rearranged feature map. The image feature map that has undergone channel reordering and spatial adaptive magnification is input into the fusion feature-guided reconstruction network. The morphological fusion features are encoded as structural enhancement factors through the fusion feature-guided pathway to participate in the feature convolution process, guiding the distribution of convolution weights to focus on regions with boundary continuity and structural consistency. Extract the backbone structure contour of the connected regions from the image features output by the fusion feature-guided reconstruction network. Combine structural connectivity analysis and boundary closure discrimination algorithms to filter out interference regions that do not have structural closure properties, and retain the region features with strong connectivity and closed boundaries in the character contour region to complete the extraction of the true box number character structure features.

6. The train car number, train container number, and train model identification device according to claim 1, characterized in that, The path evaluation and identification module is implemented as follows: The extracted structural features of the real box number characters are input into the evaluation and guidance recognition path selection structure. By performing feature map channel compression and temporal feature expansion operations on the real box number character structural features, a spatiotemporal joint feature map containing stroke arrangement order, boundary closure attributes and local connection weights is formed. A path scoring unit is constructed in the evaluation and guidance recognition path selection structure to score the structural expressive ability of different path candidate sequences, which is used to initialize the recognition path. Based on the path candidate sequence output by the evaluation-guided identification path selection structure, a character continuity score and a stroke closure score model are constructed. The character continuity score is calculated by analyzing the stroke sequence connection stability and the cosine similarity of adjacent stroke directions in the structural features of the real box number characters. The stroke closure score is measured by the energy gradient intensity distribution and the rate of change of closure curvature in the identified contour closure path. The score results are fed back to the path selection structure as a weighting factor to complete the identification of real box number characters. The identified character probability distribution sequence is input into the jump decoding unit. Based on the probability of whitespace characters and the marking of repeated characters in the temporal features, the jump decoding unit performs a forward merging search by constructing a non-overlapping jump candidate point sequence, dynamically prunes the character sequence, removes redundant paths, and outputs the optimal character sequence to achieve the character sequence output.

7. The train car number, train container number, and train model identification device according to claim 6, characterized in that, Based on the path candidate sequence output by the evaluation-guided path selection structure, character continuity scoring and stroke closure scoring models are constructed. Character continuity scoring is calculated by analyzing the stroke sequence connection stability and cosine similarity of adjacent stroke directions in the structural features of real box number characters. Stroke closure scoring is measured by identifying the energy gradient intensity distribution and the rate of change of closure curvature in the closed contour path. Specifically: Based on the path candidate sequence output by the evaluation-guided path selection structure, the stroke connection trajectory in each candidate path is extracted, the direction vector of each stroke connection trajectory is modeled, and the direction cosine similarity and connection point response stability between adjacent strokes are calculated. By statistically analyzing the connection continuity coefficient and structural jump probability between strokes, a character continuity scoring model is constructed to measure the overall arrangement coherence of character strokes. Based on the character continuity scoring model, the contour closure region in the path candidate sequence is identified, the pixel gradient response map and local curvature distribution map of the closed path are extracted, the average value of the energy gradient intensity and the standard deviation of the curvature change rate on the closed path are calculated, and the stroke closure scoring model is constructed to evaluate the degree of contour closure and edge closure consistency in each character candidate. The character continuity score and stroke closure score results are normalized and fused to generate a unified path score vector, which is fed back as a weighting factor to the evaluation guide recognition path selection structure. The recognition priority of the path candidate sequence is reordered, and the path with the most complete structure and the highest stroke coherence is selected based on the comprehensive score to complete the recognition output of the real box number character.

8. The train car number, train container number, and train model identification device according to claim 1, characterized in that, The implementation of the dynamic path control module is as follows: Based on the fusion degree vector and the extracted confidence score, a fusion confidence mapping model is constructed to generate a fusion score value for each recognition result. The fusion score value is obtained by weighted summation of the features of each dimension of the fusion degree vector and the fusion of the extracted confidence score. Based on the preset fusion score threshold range, the recognition results are divided into three recognition path categories: direct output path, structure enhancement recognition path, and character relocation and re-recognition path. Recognition results with fusion score values ​​higher than the first threshold in the fusion score threshold range are included in the direct output path. Confidence values ​​are confirmed for the character recognition results in the direct output path and they are directly output as the final recognition results, ending the subsequent processing flow and ensuring the rapid completion and stable generation of high-confidence paths. Recognition results with fusion scores between the first and second thresholds are assigned to the structure enhancement recognition path. The image feature map is subjected to structure enhancement processing to improve the coherence of character structure and boundary closure before the recognition result is regenerated. Recognition results with fusion scores below the second threshold are assigned to the character relocation and re-recognition path. The character image region is relocated and image features are extracted. The structure reconstruction and confidence enhancement recognition processing are performed to generate new recognition results. The fusion score corresponding to the new recognition result is input into the fusion confidence mapping model to update the path classification criteria. The path division and recognition are performed iteratively until the final recognition result that meets the output conditions is generated.

Images