Weakly-labeled based intelligent detection method for urban health examination problem

Abstract

The present disclosure provides a weakly annotated urban health examination problem intelligent detection method, which comprises the following steps: performing structure division on a high-density facility area of an urban health examination image to generate a plurality of independent facility unit area blocks; identifying a local candidate problem area in each independent facility unit area block based on an initialized area-level positioning model; establishing an association between the local candidate problem area and an image-level weakly annotated category label; calibrating the position of the local candidate problem area by using a small amount of strongly annotated samples, and optimizing the parameters of the area-level positioning model based on the calibration result; and performing facility-level problem detection on a real urban health examination image by using the optimized area-level positioning model, and outputting a problem positioning result. The method can effectively distinguish individual facilities in a high-density scene and achieve accurate problem positioning under weak annotation conditions.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of computer vision and artificial intelligence, in particular to an intelligent detection method for urban health examination problems based on weak annotation. BACKGROUND

[0002] At present, it has become an important trend to use a deep learning model to automatically analyze urban health examination images, and the goal is to efficiently identify defects of various facilities. In order to achieve high-precision detection, existing methods generally rely on a large amount of strongly annotated data, that is, the location and category of each problem in the image need to be accurately annotated, resulting in high data annotation cost and difficulty in wide application in practice. In order to overcome this bottleneck, weak annotation methods have begun to attract attention, which only require image-level category labels, greatly reducing the annotation burden. However, in the urban high-density facility scene, such as densely distributed power distribution boxes, valve covers, etc., existing weak annotation methods face serious challenges.

[0003] Because the model lacks location supervision of individual facilities, it tends to identify the entire dense region as a whole problem area, and cannot associate the image-level label with the specific facilities within the region, resulting in a category semantic diffusion problem. Although a small amount of strongly annotated samples can be introduced for auxiliary fine-tuning, under the condition that the appearance of the facilities is similar and the spatial distance is extremely small, the model is difficult to learn accurate facility-level positioning ability, ultimately leading to positioning confusion of the detection result, and failing to meet the demand of precise operation and maintenance.

[0004] Therefore, there is an urgent need for an intelligent detection method that can effectively distinguish individual facilities in high-density scenes and achieve accurate problem positioning under weak annotation conditions. SUMMARY

[0005] Therefore, in order to solve the problems brought by the prior art, the present application provides an intelligent detection method for urban health examination problems based on weak annotation.

[0006] In a first aspect, the present disclosure provides an intelligent detection method for urban health examination problems based on weak annotation, which comprises:

[0007] S1, performing structure division of high-density facility regions on urban health examination images to generate a plurality of independent facility unit region blocks;

[0008] S2, identifying local candidate problem regions inside each of the independent facility unit region blocks based on an initialized region-level positioning model;

[0009] S3, establishing an association relationship between the local candidate problem regions and image-level weak annotation category labels;

[0010] S4, calibrate the position of the local candidate problem region using a small amount of strongly labeled samples, and optimize the parameters of the region-level positioning model based on the calibration result;

[0011] S5, perform facility-level problem detection on the real urban medical image using the optimized region-level positioning model, and output the problem positioning result.

[0012] Optionally, the S1 comprises:

[0013] extracting the outer contour boundary features of the facility from the urban medical image;

[0014] identifying the interval features between adjacent facility mounting pedestals;

[0015] detecting the outer contour breakpoints based on the interval changes to generate structure separation lines;

[0016] dividing the image into a plurality of independent facility unit region blocks according to the structure separation lines.

[0017] Optionally, the S2 comprises:

[0018] extracting feature responses of the facility unit region blocks by a feature extraction module of the region-level positioning model, and screening out a set of suspected problem pixel positions based on the feature responses;

[0019] cropping the suspected regions screened out using the boundaries of the facility unit region blocks to which they belong to generate local candidate regions that do not cross regions, merging all the local candidate regions to form a set of local candidate problem regions.

[0020] Optionally, the S3 comprises:

[0021] inputting the features of the candidate problem regions in the set of local candidate problem regions into a classification module of the region-level positioning model to obtain a class confidence vector, and assigning the highest value class in the confidence vector to each candidate problem region in the set of local candidate problem regions to form an initial binding relationship;

[0022] performing class conflict resolution on the initial binding relationship to output a final region-level class binding set.

[0023] Optionally, the performing class conflict resolution on the initial binding relationship further comprises:

[0024] for a plurality of candidate regions bound to the same image-level class label, calculating the similarity between the region features and the standard features of the class;

[0025] based on the similarity comparison result, assigning the class label only to the candidate region with the highest similarity, and releasing the binding relationship of the remaining candidate regions with the class.

[0026] Optionally, the S4 comprises:

[0027] The candidate region in the association relationship is matched with the strong annotation sample in terms of category and space, and a spatial offset vector between boundary coordinates thereof is calculated;

[0028] The position of the candidate region that is successfully matched is calibrated based on the spatial offset vector;

[0029] The parameters of the region-level positioning model are updated by using the calibrated position and the position coordinates of the strong annotation sample.

[0030] Optionally, the updating of the parameters of the region-level positioning model by using the calibrated position and the position coordinates of the strong annotation sample comprises:

[0031] Region block image data is extracted based on the region-level category binding set and is input into the region-level positioning model to obtain a predicted position result;

[0032] A difference between the predicted position result and the calibrated position is calculated to construct a loss function;

[0033] The loss function is minimized to adjust internal parameters of the region-level positioning model.

[0034] Optionally, the S5 comprises:

[0035] Facility region blocks are automatically divided from a real city body inspection image;

[0036] The optimized region-level positioning model is used to independently perform prediction on each region block to generate original detection results containing problem positions and categories;

[0037] The problem positions and categories in the original detection results are bound to the facility region blocks in which the problem positions and categories are located;

[0038] Based on the binding relationship, a standardized facility-level problem positioning result is generated and output, and the result at least contains facility identification, problem category, and problem position coordinate information.

[0039] In a second aspect, the disclosure provides an electronic device, which comprises a memory and at least one processor, the memory stores a computer program, and the processor is configured to execute the computer program to implement the method of the first aspect.

[0040] In a third aspect, the disclosure provides a computer storage medium, which stores a computer program, and the computer program is executed to implement the method of the first aspect.

[0041] The beneficial effects of the present disclosure are that, compared with the prior art, the present disclosure has the following advantages:

[0042] 1) In view of the problem that existing methods rely on a large amount of accurate labeled data, resulting in high cost and difficulty in large-scale application, the present method uses image-level weak labeling as the main supervision signal, greatly reducing the human and time cost of data labeling, and at the same time, through the construction of a physical structure division model and an independent region block generation mechanism, it realizes effective analysis of high-density facilities under weak supervision conditions, making the method economically feasible for practical application.

[0043] 2) In view of the problem that weak labeling methods in high-density scenes are prone to class semantic diffusion and cannot associate problems to specific facility units, the present method designs an independent association and conflict resolution mechanism for candidate regions and weak labeling classes, ensuring that each detected problem can be accurately bound to a unique facility unit, effectively blocking the erroneous propagation of class semantics between dense facilities, and significantly improving the accuracy of facility-level problem identification.

[0044] 3) In view of the problem that the positioning ability of the model is insufficient under the condition of similar appearance and extremely small distance of the facility, the detection result is easy to confuse, the present method combines the advantages of weak labeling and a small amount of strong labeling samples, through a region-level precise positioning calibration strategy, using the precise position information provided by the strong labeling samples to fine-tune the model, so that the model maintains the low-cost advantage of weak labeling while obtaining precise positioning ability close to strong supervision methods, and finally outputs reliable facility-level problem positioning results that can be used for precise operation and maintenance. BRIEF DESCRIPTION OF DRAWINGS

[0045] The accompanying drawings, which are incorporated into and form part of the specification, illustrate embodiments consistent with the present disclosure and, together with the specification, serve to explain the principles of the present disclosure.

[0046] Figure 1 A flowchart of a weak labeling-based urban physical examination problem intelligent detection method provided by an embodiment of the present disclosure is shown;

[0047] Figure 2 A flowchart of the association of candidate regions and weak labeling labels provided by an embodiment of the present disclosure is shown;

[0048] Figure 3 A flowchart of position calibration and model optimization provided by an embodiment of the present disclosure is shown.

[0049] Through the above drawings, the specific embodiments of the present disclosure have been shown, and more detailed descriptions will be given in the following. These drawings and written descriptions are not intended to limit the scope of the concept of the present disclosure in any way, but to illustrate the concept of the present disclosure to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0050] The present disclosure is further described below with reference to the accompanying drawings. The following examples are merely intended to more specifically illustrate the technical solutions of the present disclosure, and cannot be used to limit the protection scope of the present disclosure.

[0051] The components of the embodiments of the present application generally described and illustrated in the accompanying drawings can be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0052] Hereinafter, the terms "include", "have", and their synonymous words used in various embodiments of the present application are merely intended to denote a specific characteristic, number, step, operation, element, component, or combination of the foregoing, and should not be construed as excluding the presence or addition of one or more other characteristics, numbers, steps, operations, elements, components, or combinations of the foregoing.

[0053] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which various embodiments of the present application belong. Terms such as those defined in generally used dictionaries will be interpreted as having the same meaning as the contextual meaning in the relevant technical field and will not be interpreted as having an idealized or overly formal meaning, unless clearly defined in various embodiments of the present application.

[0054] To achieve high-precision urban inspection problem detection, existing methods generally rely on a large amount of accurate labeling of strong labeled data, resulting in high labeling costs and difficulty in large-scale application. While image-level weak labeling can significantly reduce costs, in high-density facility scenarios such as distribution boxes and valve covers, the model lacks location supervision of individual facilities, and is prone to class semantic diffusion and positioning confusion problems, i.e., tends to misjudge the entire dense area as a whole problem area, and cannot accurately associate the problem to the specific facility unit, making it difficult to meet the needs of precise operation and maintenance. To address this technical bottleneck, the embodiments of the present disclosure provide a weakly labeled urban inspection problem intelligent detection method as follows.

[0055] Figure 1 The flowchart of the weakly labeled urban inspection problem intelligent detection method provided by the embodiments of the present disclosure is shown as Figure 1 as follows:

[0056] S1: The structure of the high-density facility area of the city health examination image is divided, and a plurality of independent facility unit area blocks are generated.

[0057] The physical structure features of the dense facility area in the city health examination collected image are extracted, and the original image is divided into a plurality of region blocks with independent installation unit significance by identifying the physical boundary features between facilities, the installation base interval, and the entity structure information such as the outer contour breakpoint. These region blocks can reflect the real spatial layout of the equipment and provide a stable physical basis for the subsequent candidate region constraint. The specific implementation process includes the following sub-steps.

[0058] S1.1: Extract the outer contour boundary features of the facilities in the city health examination image to form a contour pixel set.

[0059] The system reads the original image collected by the city health examination, denoted as I, and performs contour recognition processing on the appearance structure of the facilities in the image. Contour recognition is achieved by using a contour extraction function C(·) such as a classic edge detection algorithm such as Canny, Sobel, or a combination of variants, which detects a continuous or approximately closed boundary pixel set based on image pixel gradient change, such as the visual contour of typical facilities such as power distribution box shell, valve manhole cover edge, or pressure gauge box body frame. The contour extraction process forms a facility outer contour boundary set B, which is mathematically expressed as B = C(I), where B represents the facility outer contour boundary set, and each element B(k) represents the pixel set of the kth contour boundary. By quantifying the contour pixel position, a geometric basis is provided for subsequent interval analysis to ensure that the visual boundary of different facility main bodies can be accurately represented, avoiding the problem of contour breakage or blur caused by image noise or light changes.

[0060] S1.2: Identify the interval features between adjacent facility installation bases and quantify the physical distance.

[0061] Based on the obtained set of facility outer contour boundaries B, the system identifies the spacing features of adjacent facility mounting pedestals. The mounting pedestal spacing refers to the horizontal or vertical physical distance between multiple facilities, such as the support pedestal distance between multiple distribution boxes or the embedded structure distance between pipe gallery manholes. The system achieves spacing feature extraction by calculating the minimum Euclidean distance between adjacent contour boundaries, assuming that the contour set B is arranged in spatial position order, and the specific formula is D = {d_k | d_k = min_distance[B(k), B(k+1)] | for k = 1 to |B| - 1}, where D represents the set of mounting pedestal spacings, each element d_k represents the kth adjacent facility spacing distance; the min_distance(·) function is used to calculate the shortest Euclidean distance between two contour boundary pixel sets. This step can quantify the fixed physical distance between facilities in high-density areas, providing distance variation basis for structure breakpoint analysis and region division, thereby enhancing the accuracy and robustness of division and ensuring that subsequent processing is based on real mounting layout rather than visual approximation.

[0062] S1.3: Detect outer contour breakpoints based on spacing changes and generate a set of structure separation lines.

[0063] Using the distance variation features in the set of mounting pedestal spacings D, the breakpoint positions in the facility arrangement are identified. Breakpoint positions usually correspond to natural boundaries in facility layout, such as positions where there are shape mutations or significant changes in spacing in multiple distribution box arrangements, indicating facility mounting unit boundaries. The system detects these breakpoints by analyzing the difference changes in the set of spacings D, first calculating the set of mounting spacing changes △D = {δ_k | δ_k = |d_{k+1} - d_k| for k = 1 to |D| - 1}, where δ_k represents the distance change of adjacent spacings. Then, based on the preset breakpoint detection threshold T_s dynamically calculated according to the overall distribution of the set of mounting pedestal spacings D (such as mean and standard deviation), the position index k that satisfies δ_k > T_s is identified, and these positions form the set of structure separation lines S.

[0064] Optionally, the system considers the spacing distribution pattern in the local area when detecting breakpoints, rather than judging the difference of a single spacing in isolation. For example, in areas where the facility arrangement is uniform, the system uses a relatively lenient threshold to tolerate minor fluctuations; while in areas where the facility arrangement itself is irregular, a more stringent threshold is used to ensure that real structure boundaries are not missed. This adaptive strategy based on local context significantly improves the generalization ability of the model in complex real-world scenarios.

[0065] This step identifies the logical boundaries of facility mounting by spacing mutations, making region division more consistent with physical reality and avoiding misdivision caused by relying solely on visual features, thereby improving the coherence and reliability of the overall structure.

[0066] S1.4: Divide the original image into multiple independent region blocks according to the structure separation lines, and generate a region block set.

[0067] According to the structure separation line set S, the original image I is divided into multiple independent region blocks, and a region block set R is generated. The division operation forms a rectangular or approximately rectangular block according to the separation line position, and each region block corresponds to a facility installation unit, ensuring spatial independence and integrity. Each element R(p) in the region block set R represents the pth region block image part; here, the division refers to determining the boundaries of each independent region block on the image I according to the separation lines identified in the set S, thereby obtaining the region block set R. This step obtains the region block set R, which provides a strict spatial constraint basis for the generation of local candidate problem regions in the subsequent steps, ensuring that each candidate region is generated only within one facility unit, thereby effectively avoiding the problem of cross-region aggregation and improving detection accuracy.

[0068] In the technical scheme of the embodiments of the present disclosure, by extracting the facility outer contour boundary, identifying the installation base interval, and detecting the structure breakpoint, the high-density facility region is divided into independent region blocks according to its real physical structure, providing a stable spatial constraint basis for subsequent processing, effectively avoiding the possible region misdivision caused by relying only on visual features, and ensuring the alignment of candidate region generation and real facility installation units.

[0069] S2: Identify local candidate problem regions within each independent facility unit region block based on the initialized region-level positioning model.

[0070] After completing the structure division of the high-density facility region block, the system initializes a region-level positioning model, which is used to identify potential problem regions within the region block based on visual features. Then, based on this model, local candidate problem regions are generated within each independent region block, avoiding cross-region aggregation and thus solving the defect of weak labeling mechanism considering multiple facilities as a unified problem source. The system ensures that the candidate regions correspond one-to-one to physical devices by calculating the feature response strength within the region block, performing preliminary screening of suspected problem regions, applying region boundary constraints, and merging to generate a region-level candidate set, providing an accurate positioning basis for subsequent class binding. The specific implementation process includes the following sub-steps.

[0071] S2.1: Initialize the region-level positioning model and extract the region block feature response.

[0072] Based on the region block set R obtained in step S1, a region-level positioning model M is initialized. The model M has general image feature extraction and classification capabilities. Each region block R(p) is processed in turn using the initialized region-level positioning model M. The model M extracts local structural features from its internal pixels through its internal feature extraction module. Based on visual features such as texture changes, surface irregularities, and color mutations within the region, the module calculates an anomaly score for each pixel position and generates a feature response value matrix F(p). The specific formula is F(p) = M_feat[R(p)], where F(p) represents the feature response matrix of the pth region block, and each element F(p)(i,j) represents the response value of the ith row and jth column pixel in the region block; M_feat represents the feature extraction module of the positioning model, which can be implemented through one or a combination of methods such as pre-trained convolutional neural network (CNN) feature maps, local binary pattern (LBP) texture features, or histogram of oriented gradients (HOG). Its calculation process relies on the texture and surface continuity changes within the region, and quantifies the intensity of problem signs by analyzing the gradient statistics and color distribution differences of the pixel neighborhood. This step can identify the distribution of potential problem signs within the region, providing a quantitative basis for subsequent screening and ensuring that the candidate generation is based on objective features rather than subjective judgments.

[0073] S2.2: Threshold processing based on feature response matrix to screen out suspected problem pixel position set.

[0074] Based on the obtained feature response set F, the system performs threshold processing on the feature response matrix F(p) of each region block to filter out pixel positions that are significantly higher than the background response within the region. This processing is achieved by setting a response threshold T, which is dynamically determined based on the statistical characteristics of the feature response matrix F(p) of each region block. For example, the Otsu algorithm or a linear combination of the mean and standard deviation of the response value can be used, with a value range of generally 0.2 to 0.6, to filter out low response positions and retain potential problem areas. Specifically, the suspected region set G(p) is composed of all pixel positions (i,j) that satisfy the condition F(p)(i,j)>T, which can be mathematically expressed as G(p)={(i,j)|F(p)(i,j)>T}. In this set, each position (i,j) is determined as a suspected problem region. By dynamically adjusting the threshold to adapt to different feature changes in different regions, false positives caused by lighting or material differences can be avoided, thereby preliminarily locating the positions where problems may exist and providing a candidate range for subsequent accurate constraints.

[0075] S2.3: Apply region boundary constraints to crop suspected regions and generate a local candidate region set.

[0076] Based on the obtained suspected region set G and the region block set R obtained in step S1, a region boundary constraint is applied to ensure that the candidate region is generated only within a single region. For each continuous pixel region in the suspected region set G(p) of each region block, the system extracts its minimum circumscribed rectangle and checks whether the rectangle falls completely within the boundary range of the corresponding region block R(p). If the rectangle exceeds the region boundary, it is cropped to retain only the part coinciding with the region block. All the rectangles processed in this way constitute the local candidate region set Q(p). The spatial constraint eliminates cross-region misjudgments, ensuring that each candidate region corresponds strictly to a single facility unit, thereby improving positioning accuracy and reliability.

[0077] S2.4: Merge the candidate regions of all region blocks to form a region-level candidate set.

[0078] Based on the obtained local candidate region set Q, the candidate regions corresponding to all region blocks are merged according to the region number to form a region-level candidate set C. During the merging process, no cross-region fusion is performed to maintain the one-to-one correspondence between the candidate regions and the facility installation units. The specific formula is where P is the total number of region blocks, C represents the region-level candidate set, each element C_m (m = 1, 2,...) represents an independent candidate region, which originates from a specific region block; and represents the union operation, which is performed on the index p from 1 to the total number of regions P to merge the candidate region sets of each region block into a whole set. The final candidate region input for region-level category binding provides a direct data basis for subsequent step execution of independent category association, ensuring that the subsequent processing can be based on facility-level units for accurate analysis and output.

[0079] In the technical solution of the embodiments of the present disclosure, local candidate problem regions are generated within each independent region block based on feature response intensity, and strict region boundary constraints are used to ensure that the candidate regions do not cross regions, thereby solving the defect that multiple facilities are identified as a problem source as a whole under a weak labeling mechanism, and providing a facility-level positioning basis for subsequent accurate category binding.

[0080] S3: Establish the association relationship between the local candidate problem regions and the image-level weak labeling category labels.

[0081] After the local candidate problem regions are generated, an independent association relationship between the candidate regions and the weak labeling categories is established. Each candidate region is one-to-one bound with an image-level weak labeling category label, avoiding the spread of category semantics among multiple facilities, thereby solving the category confusion problem in high-density scenes. Figure 2 The candidate region and weak labeling label association flowchart provided by the embodiments of the present disclosure is as follows: Figure 2As shown, this step ensures that each candidate region is independently associated with the corresponding problem category by obtaining image-level weakly annotated class labels and generating a standardized class set, establishing a region numbering index system for candidate regions, establishing an initial class binding relationship based on region candidate features, performing region-level class conflict resolution, and generating a region-level class binding output set. The specific implementation process includes the following sub-steps.

[0082] S3.1: Obtain image-level weakly annotated class labels and perform standardization processing to generate a standard class set.

[0083] As shown in Figure 2 Based on the region-level candidate set C output by step S2, the weakly annotated class label L_img corresponding to the city health examination collected image is read at the same time. The weakly annotated class label is usually described in text form, such as box damage, door panel rust, or pipeline loosening, indicating the problem category existing in the whole image, but the specific location is not specified. The system performs standardization processing on L_img, converts the original label to a unified class system through a class mapping function M_map(·), which is realized based on a pre-defined class dictionary or through semantic normalization by a natural language processing (NLP) model, forming a standardized class set L. The specific formula is L=M_map(L_img), where L represents the standardized class set, each element L(k) represents the kth standard class; M_map(·) represents the class mapping function, its processing process includes text analysis, synonym merging and class coding, ensuring that labels from different sources can be mapped to a consistent semantic system. This step provides a class source basis for subsequent region-level association, ensuring the consistency and operability of class information, and avoiding binding errors caused by label isomerization.

[0084] S3.2: Establish a unique numbering index system for the candidate region set to form a region numbering set.

[0085] Based on the region-level candidate set C, a unique number is assigned to each candidate region to establish a region block numbering index system. The number generation is realized through an index function, and the system iterates through each candidate region in the candidate set C, assigns a unique identifier in order, and forms a region numbering set N. The specific formula is N=index(C), where N represents the region numbering set, each element N(m) represents the mth candidate region index; index(·) represents the numbering generation function, its execution process is based on the spatial position or generation order of the candidate region, ensuring that each number has a stable and unique correspondence with the candidate region. This step enables subsequent class binding to be performed region by region, providing structural support for fine processing and avoiding region confusion or repeated processing.

[0086] S3.3: Establish initial class binding relationship based on region candidate features through model classification module.

[0087] Based on the region number set N, the standardized class set L, and the region-level candidate set C, the local feature performance of the candidate region C(m) corresponding to each region number N(m) is analyzed to establish an initial class binding relationship. The local feature analysis includes surface damage area ratio, color abnormality degree, structure contour change degree, and other visual attributes. The system processes the candidate region C(m) through the feature extraction module M feat of the region-level positioning model M to generate its high-level feature representation, and takes this feature representation as the region feature description value vector T(m). Subsequently, the system matches the region feature description with the class semantic features, which is realized through a class binding function bind(·). The bind function is the classification module of the region-level positioning model M, which selects the most likely corresponding class label for the region. The function calculates the region feature T(m) and the class confidence vector output by the classification model pre-trained based on the weakly annotated dataset. The system assigns the class with the highest value in the confidence vector to the current region, thereby realizing binding. The specific formula is A(m) = bind[T(m), L], where A(m) represents the initial class binding result of the mth region. The weakly annotated class information is reduced from the whole image level to the region level, so that the class semantics is attached to specific facilities, laying a structural foundation for solving the class diffusion problem in high-density facility scenes.

[0088] Optionally, to further improve the reliability of binding, the system will use the region block physical boundary divided in step S1 to perform consistency verification on the initial binding result: if the typical problem form corresponding to the binding class of a candidate region is significantly contradictory to the overall physical structure of the region block where it is located, for example, the box body damage, which is a typical problem, is usually manifested as local concave cracks, while the region block where it is located is a complete rectangular distribution box surface, in this case, the system will reduce the confidence of the binding result, and make a comprehensive judgment by referring to the binding information of other candidate regions. This mechanism effectively utilizes prior physical structure knowledge and avoids class misbinding caused by local image noise or light and shadow changes.

[0089] S3.4: Perform region-level class conflict resolution to eliminate repeated binding and improve correlation accuracy.

[0090] Based on the initial region class binding set A and the region number set N, region-level class conflict resolution is performed. In a high-density facility scenario, multiple candidate regions can be bound to the same class, forming a class conflict, resulting in semantic overlap and inaccurate positioning. When multiple regions are detected to be bound to the same class L(k), the system re-computes the similarity between the region features T(m) of these conflict regions and the standard feature template of the class L(k), and retains the binding relationship between the candidate region with the highest similarity and the class L(k). The class labels of the remaining conflict regions are set to the null class label. Finally, a final class set is obtained after conflict resolution, in which each candidate region is bound to at most one class, denoted as conflict resolution class set U. By eliminating repeated binding, the accuracy and reliability of class association are improved, and misjudgment caused by multiple facilities sharing the same label is avoided.

[0091] S3.5: Generate a binding output set of candidate regions and classes to form a one-to-one correspondence relationship.

[0092] Based on the conflict resolution class set U and the region-level candidate set C, a region-level class binding output set is generated. The system merges the region set C and the class set U according to their indices m to form a one-to-one correspondence, so that each candidate region C(m) is bound to its final class U(m) to form a binding pair, generating a region-level class binding set B_bind. Each element B_bind(m) in the set B_bind is a two-tuple, denoted as (C(m), U(m)), where C(m) is the mth candidate region and U(m) is its bound class. The implementation of the foregoing binding operation includes data structure merging and association mapping, such as encapsulating C(m) and U(m) as a key-value pair or a structure body containing location and class attributes, to ensure that the candidate region and the class form a stable one-to-one correspondence relationship. The final region-level class binding result is output in this step, providing complete input data for subsequent region-level fine positioning calibration, ensuring that subsequent processing can be optimized based on accurate class and location information.

[0093] In the technical scheme of the embodiments of the present disclosure, by establishing an independent one-to-one binding relationship between the candidate region and the image-level weakly labeled class label, and performing region-level class conflict resolution, the semantic diffusion and confusion of classes among high-density facilities are effectively prevented, so that the weak supervision information at the whole image level can be accurately reduced and associated to specific facility units, and the accuracy of class discrimination is improved.

[0094] S4: Use a small amount of strongly labeled samples to calibrate the position of the local candidate problem region, and optimize the parameters of the region-level positioning model based on the calibration result.

[0095] After the independent association of the candidate region and the weakly labeled category is completed, a small amount of strongly labeled samples are used for region-level fine positioning calibration. By combining the strongly labeled samples with accurate problem location labels, the problem location is fine-tuned within each region block, thereby improving the positioning accuracy of the system under highly dense facility conditions. Figure 3 The position calibration and model optimization flowchart provided by the embodiments of the present disclosure is as shown in Figure 3 The step extracts strongly labeled position samples and establishes a region correspondence relationship, calculates the offset between the region candidate position and the strongly labeled position, performs region-level position correction based on the offset, and updates the region-level positioning model parameters, so that the model forms accurate positioning capability for a single facility. The specific implementation process includes the following sub-steps.

[0096] S4.1: Extract strongly labeled samples and establish category and spatial correspondence relationship with candidate regions.

[0097] As shown in Figure 3 Based on the region-level category binding set B_bind obtained in step S3, the strongly labeled sample set S_strong annotated by humans is read. The strongly labeled sample contains the rectangular box coordinate information of the true problem location, such as the explicitly labeled box damage area or door panel rust location. The system first performs screening based on complete category label consistency through the category matching function match(·). On this basis, for the candidate with consistent category, a loose spatial overlap condition is used for verification, such as checking whether there is any overlap of the boundary box or whether the center point falls within the other region, thereby establishing the correspondence relationship between the candidate region and the strongly labeled sample. The specific formula is H = match(B_bind, S_strong), where H represents the candidate region corresponding to the strongly labeled set, and each element H(m) represents the strongly labeled position bound to the mth region; The match(·) function performs optimal matching based on category semantic similarity and spatial overlap degree by traversing all candidate regions and strongly labeled samples.

[0098] In the high-density facility scene, to solve the positioning confusion problem, the matching process follows a key principle: when multiple candidate regions overlap with the same strongly labeled sample box in space, the system preferentially selects the candidate region whose physical boundary of the region block is most consistent with the region block where the center point of the strongly labeled box is located. This strategy ensures that the calibration process can strengthen the model's awareness that one region block corresponds to one facility entity, thereby accurately feeding the strongly labeled signal to the correct facility unit.

[0099] This step ensures that each region block can obtain the corresponding true position reference, providing accurate benchmark data for subsequent position calibration, and effectively establishing a bridge between weakly labeled regions and strongly labeled positions.

[0100] S4.2: Calculate the spatial offset between the candidate region position and the corresponding strong annotation position.

[0101] Based on the obtained candidate region corresponding strong annotation set H and the region-level class binding set B_bind, the spatial offset between the candidate position of each candidate region and the corresponding strong annotation position is calculated. The offset is calculated by the coordinate difference, and the system extracts the position coordinates in the candidate region B_bind(m) and the boundary coordinates of the strong annotation region H(m), and calculates the pixel distance in the horizontal and vertical directions. The specific formula is O(m)=H(m)-B_bind(m).position, where O(m) represents the position offset of the mth region, each element contains horizontal offset Δx and vertical offset Δy; H(m) represents the strong annotation position coordinates of the mth region; B_bind(m).position represents the candidate position coordinates of the mth region, which is extracted from the B_bind set. By quantifying the position difference, accurate adjustment parameters are provided for subsequent correction, and the position calibration is based on objective spatial deviation rather than subjective estimation.

[0102] S4.3: Fine correction of candidate region position based on offset, optimization of positioning accuracy.

[0103] Based on the obtained candidate region offset set O and the region-level class binding set B_bind, the candidate position of each candidate region is fine corrected. The correction process is realized by a calibration function F_calibrate(·), which takes the original candidate position B_bind(m).position and the calculated position offset O(m) as input, and outputs the optimized calibration position K(m). The purpose is to correct the initial prediction according to the supervision signal of strong annotation, while maintaining the smoothness of the prediction result, avoiding overfitting to the single strong annotation sample H(m) that may exist in the annotation noise. The specific formula is K(m)=F_calibrate[B_bind(m).position,O(m)]. Through position optimization, the candidate region is more consistent with the real problem position, which significantly improves the positioning accuracy and provides more accurate supervision signal for model training.

[0104] S4.4: Update the region-level positioning model parameters using the calibration position and the strong annotation position.

[0105] Based on the obtained calibrated candidate position set K, the region-level class binding set B_bind, the region-level positioning model M is updated. The model updating process is: based on the position information in B_bind, the region block image data is extracted from the original training image, input into the current model M to obtain the predicted position result, then the difference between the predicted position and the calibrated more accurate target position K is calculated, and the gradient descent or other bias minimization strategy is used to adjust the internal parameters of the model M. The specific formula is M_optimized=update(M,B_bind,K), wherein M_optimized represents the updated region-level positioning model; M represents the region-level positioning model to be optimized; B_bind represents the training data used for this model update, which provides region block position information to extract image data; the update(·) function encapsulates the entire process of the foregoing forward prediction, loss calculation and backward parameter update. This step enables the model M to learn the calibrated and accurate position distribution, so as to evolve into a more stable facility-level positioning model M_optimized. This step enables the model to automatically learn the distribution of real problem positions in the region, form a stable facility-level positioning capability, and ensure that the system can independently and accurately perform problem positioning in subsequent applications.

[0106] In the technical solution of the embodiment of the present disclosure, a small amount of strong labeled samples with accurate position information is used to calculate the position offset of the candidate region with bound classes and to perform fine calibration, thereby making up for the deficiency of pure weakly labeled supervision in spatial positioning, significantly improving the positioning accuracy of the model for the problem position in a single facility, and enhancing the practicality and reliability of the system in complex dense scenes.

[0107] S5: Perform facility-level problem detection on real city inspection images using the optimized region-level positioning model, and output the problem positioning result.

[0108] After completing the region-level fine positioning calibration, perform facility-level city inspection problem intelligent detection and output the final positioning result. Apply the optimized detection model to real city inspection images to realize accurate identification and positioning of specific problem facilities in high-density facility scenes, and solve the positioning confusion problem caused by class semantic aggregation under weak labeling conditions. By extracting facility region blocks from real images, performing model prediction on the region blocks to generate original detection results, binding the predicted positions to the corresponding facility regions, and generating final facility-level positioning output results, a problem facility positioning result that can be directly used in a city management system is formed. The specific implementation process includes the following sub-steps.

[0109] S5.1: Automatically divide the real city inspection image into facility region blocks to generate a region block set.

[0110] Based on the optimized regional positioning model M_optimized obtained in step S4, the real city survey image I_real is read, and automatic division of the facility region block is performed. The region division is realized by a region division function split(·), which identifies independent facility units in the image and generates a region block set based on the facility outer contour features and installation base spacing information using the same physical structure analysis method as step S1. The specific formula is R_real = split(I_real), where R_real represents the region block set divided from the real image, and each element R_real(p) represents the pth facility region block; the split(·) function ensures that the division result accurately reflects the actual facility layout through a series of image processing techniques such as edge detection, contour analysis and spacing calculation. This step provides a structured input unit for subsequent model prediction, ensures that the detection process is based on the real facility physical boundary, and avoids detection deviation caused by inaccurate region division.

[0111] S5.2: Independently perform model prediction for each region block to generate original detection results containing positions and categories.

[0112] Based on the obtained region block set R_real and the optimized regional positioning model M_optimized, problem detection and classification are independently performed for each region block. The system inputs each region block R_real(p) into the optimized regional positioning model M_optimized, which analyzes visual features and predicts problem positions and categories within the region, and outputs detection results containing position coordinates and category labels. The specific formula is Y(p) = M_optimized[R_real(p)], where Y(p) represents the original detection result of the pth region, each element containing a predicted position coordinate and a category label; M_optimized represents the optimized regional positioning model with region block standardization capability, which realizes feature extraction and target positioning through convolutional neural network and region proposal network. This step ensures that each facility unit is independently analyzed, avoids mutual interference between regions, and provides accurate original detection data for subsequent result binding.

[0113] S5.3: Bind the predicted positions in the original detection results to the corresponding facility region blocks.

[0114] Based on the obtained original detection result set Y and the region block set R real, each predicted position is bound to the facility region to which it belongs. The binding operation is achieved through spatial position association. The system determines the region block R real(p) to which the predicted position belongs according to the predicted position coordinates, and establishes a mapping relationship between the detection result Y(p) and the corresponding region block to form a bound result set Z. This step establishes a clear correspondence between the detection result and the physical facility, providing a structured data basis for the final output.

[0115] S5.4: Generate standardized facility-level problem positioning output results, including facility number, category and location information.

[0116] Based on the obtained region binding result set Z, the final facility-level positioning output result is generated. The system converts each binding entry into a standardized output format through a structured output function format(·), including facility number, problem category, problem location coordinates and other complete information. The specific formula is F output=format(Z), where F output represents the final facility-level positioning result set, and each element F output(p) represents the problem description, classification and location information of the pth facility. The format(·) function standardizes and organizes the data to ensure that the output result meets the data interface requirements of the urban management system. Before generating the final output, the system can perform a simple posteriori verification logic, for example, if a region block is identified as missing a box door but the physical outline integrity score of the region where it is located is higher than a preset empirical value, the system will mark the confidence of the result as to be verified for manual intensive review. This step forms a detection output that can be directly used in practical applications, providing reliable technical support for precise maintenance and management decisions of urban facilities.

[0117] In the technical scheme of the embodiments of the present disclosure, the optimized model after structure division, candidate generation, category binding and fine positioning calibration is applied to real urban body examination images, and finally the facility-level problem positioning and classification results are output, forming a complete closed loop from image input to precise operation and maintenance decision, solving the positioning confusion problem of high-density facility detection under weak labeling conditions, and meeting the actual needs of urban fine management.

[0118] In conclusion, the present application aims at the problem that the existing method relies on a large amount of accurate labeled data, resulting in high cost and difficulty in large-scale application. The present application uses image-level weak labeling as the main supervision signal, greatly reducing the labor and time cost of data labeling. At the same time, by constructing a physical structure division model and an independent region block generation mechanism, effective analysis of high-density facilities under weak supervision conditions is realized, making the method economically feasible for practical application. In view of the problem that weak labeling methods in high-density scenes are prone to class semantic diffusion and cannot associate problems to specific facility units, the present application designs an independent association and conflict resolution mechanism for candidate regions and weak labeling classes, ensuring that each detected problem can be accurately bound to a unique facility unit, effectively blocking the error propagation of class semantics between dense facilities, and significantly improving the accuracy of facility-level problem identification. In view of the problem that the positioning ability of the model is insufficient under the condition of similar appearance and extremely small spacing of facilities, and the detection result is easy to confuse, the present application combines the advantages of weak labeling and a small amount of strongly labeled samples, and uses a region-level fine positioning calibration strategy to fine-tune the model using the accurate position information provided by the strongly labeled samples, so that the model can maintain the low-cost advantage of weak labeling while obtaining the precise positioning ability close to the strong supervision method, and finally output reliable facility-level problem positioning results that can be used for precise operation and maintenance.

[0119] According to embodiments of the present disclosure, the present disclosure also provides an electronic device, which can include a processor, a communications interface, a memory and a communications bus, wherein the processor, the communications interface and the memory complete mutual communication through the communications bus. The processor can invoke a logical instruction in the memory to execute the method provided by each method.

[0120] In addition, the logical instruction in the memory described above can be implemented in the form of a software functional unit and sold or used as an independent product, which can be stored in a computer-readable storage medium. Based on such understanding, the technical solutions of the present disclosure essentially or in part or part of the technical solutions can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in the embodiments of the present disclosure. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0121] In another aspect, the present disclosure also provides a non-transitory computer readable storage medium having stored thereon a computer program, which, when executed by a processor, implements the method provided by the above method.

[0122] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the present embodiment according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0123] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software plus necessary universal hardware platforms, and of course can also be realized by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part of the prior art that makes a contribution, can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0124] It should be understood that the above embodiments are only used to illustrate the technical solutions of the present disclosure, rather than limit them; although the present disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features therein; and these modifications or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure.

Claims

1. A method for intelligent detection of urban health checkup problems based on weak labeling, characterized in that, The method includes: S1. Divide the urban physical examination image into high-density facility areas to generate multiple independent facility unit area blocks; S2. Based on the initialized regional-level positioning model, identify local candidate problem regions within each of the independent facility unit regional blocks; S3. Establish the association between the local candidate problem regions and the image-level weak annotation category labels; S4. Use a small number of strongly labeled samples to perform position calibration on the local candidate problem region, and optimize the parameters of the region-level localization model based on the calibration results; S5. Use the optimized regional-level localization model to perform facility-level problem detection on real urban physical examination images and output the problem localization results; S2 includes: The feature extraction module of the regional positioning model extracts the feature response of the facility unit area block, and filters out a set of suspected problem pixel locations based on the feature response. For the selected suspected areas, the boundaries of the facility unit area blocks to which they belong are used for trimming to generate local candidate areas that do not cross regions. All local candidate areas are merged to form a set of local candidate problem areas. S3 includes: The features of the candidate problem regions in the local candidate problem region set are input into the classification module of the regional localization model to obtain the category confidence vector. The category with the highest value in its confidence vector is assigned to each candidate problem region in the local candidate problem region set to form an initial binding relationship. Perform category conflict resolution on the initial binding relationship and output the final region-level category binding set.

2. The intelligent detection method for urban health checkup problems based on weak labeling as described in claim 1, characterized in that, S1 includes: Extract the outer contour boundary features of the facilities from the urban physical examination images; Identify the spacing characteristics between adjacent facility installation bases; Detection of outer contour breakpoints based on interval variation to generate structural separation lines; The image is divided into multiple independent facility unit area blocks based on the structural dividing lines.

3. The intelligent detection method for urban health checkup problems based on weak labeling according to claim 1, characterized in that, The process of performing category conflict resolution on the initial binding relationship also includes: For multiple candidate regions that are bound to the same image-level category label, calculate the similarity between the features of each region and the standard features of that category; Based on the similarity comparison results, only the category label is assigned to the candidate region with the highest similarity, and the binding relationship between the remaining candidate regions and the category is removed.

4. The intelligent detection method for urban health checkup problems based on weak labeling according to claim 1, characterized in that, S4 includes: The candidate regions in the association relationship are matched with strongly labeled samples by category and space, and the spatial offset vector between their boundary coordinates is calculated. The positions of successfully matched candidate regions are calibrated based on the spatial offset vector. The parameters of the regional positioning model are updated using the calibrated location and the location coordinates of the strongly labeled sample.

5. The intelligent detection method for urban health checkup problems based on weak labeling according to claim 4, characterized in that, The step of updating the parameters of the regional localization model using the calibrated location and the location coordinates of the strongly labeled samples includes: Based on the region-level category binding set, region block image data is extracted and input into the region-level localization model to obtain the predicted location result; Calculate the difference between the predicted location and the calibrated location to construct a loss function; Minimize the loss function to adjust the internal parameters of the regional localization model.

6. The intelligent detection method for urban health checkup problems based on weak labeling according to claim 1, characterized in that, S5 includes: Automatically divide real urban health check images into facility area blocks; The optimized regional localization model is used to perform predictions independently on each region block, generating raw detection results that include the location and category of the problem. Bind the location and category of the problem in the original detection results to the facility area block where it is located; Based on the binding relationship, standardized facility-level problem location results are generated and output, and the results include at least facility identifier, problem category and problem location coordinates.

7. An electronic device, characterized in that, The electronic device includes a memory and at least one processor, the memory storing a computer program, and the processor executing the computer program to implement the intelligent detection method for urban physical examination problems based on weak labeling as described in any one of claims 1-6.

8. A computer storage medium, characterized in that, It stores a computer program, which, when executed, implements the intelligent detection method for urban health checkup problems based on weak annotation, as described in any one of claims 1-6.

Images