Deep reinforcement learning multi-displacement processing method for planar note capping conflict

Abstract

The invention belongs to the technical field of data processing, and provides a planar note capping conflict-oriented deep reinforcement learning multi-displacement processing method, which comprises the following steps of: determining candidate planar notes with capping conflicts in a target area and the complexity of each capping conflict event, determining a target displacement mode of the candidate planar letters under the corresponding complexity; determining the number of glands, the area of the glands and the semantic importance of each candidate planar note, and calculating to obtain a conflict importance comprehensive score of each candidate planar note; according to the conflict importance comprehensive score, determining an intelligent agent selection strategy in a corresponding target displacement mode in each capping conflict event; and according to the target displacement mode and the agent selection strategy, in combination with a pre-established DQN model and a displacement constraint reward strategy, controlling the candidate planar annotation to execute a displacement action until the displacement action reaches an optimal position. According to the scheme, the high efficiency, accuracy, flexibility and reliability of the planar note capping conflict processing link are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a deep reinforcement learning multi-displacement processing method for planar annotation cover conflict. BACKGROUND

[0002] In the field of map annotation configuration, annotations, as the core elements of conveying geographic information, directly affect the readability and use value of the map. However, the annotation automatic configuration function of the current mainstream GIS (Geographic Information System) software still has significant limitations, especially in the processing of planar annotation cover conflict, which has become a key bottleneck of automatic geographic information mapping.

[0003] Existing annotation displacement schemes are usually limited to a single displacement method, which is difficult to meet the comprehensive needs of efficiency and flexibility in resolving cover conflicts under different cover conflict scenarios. With the development of deep learning technology, although some research has applied neural networks and deep reinforcement learning to annotation recognition and point annotation configuration, there is still a gap in the processing of planar annotation cover conflict. Existing deep reinforcement learning methods cannot dynamically match displacement methods according to the complexity of cover conflict, nor can they establish a perfect agent selection mechanism, resulting in a difficult balance between conflict resolution efficiency and key annotation position stability in complex scenarios, and the inability to effectively guide annotations to avoid new conflicts and maintain their association with the planar elements during the displacement process.

[0004] Therefore, the traditional annotation displacement processing scheme has the technical problems of insufficient efficiency and accuracy, as well as insufficient flexibility and reliability. SUMMARY

[0005] The present application provides a deep reinforcement learning multi-displacement processing method for planar annotation cover conflict to address the deficiencies of traditional annotation displacement processing schemes, such as insufficient efficiency and accuracy, as well as insufficient flexibility and reliability.

[0006] The present application provides a deep reinforcement learning multi-displacement processing method for planar annotation cover conflict, which includes: determining candidate planar annotations with cover conflict and the complexity of each cover conflict event in the target area, and determining the target displacement method for the candidate planar annotations under the corresponding complexity; determining the number of covers, cover area, and semantic importance of each candidate planar annotation, and calculating the conflict importance comprehensive score of each candidate planar annotation; determining the agent selection strategy under the corresponding target displacement method in each cover conflict event according to the conflict importance comprehensive score; and controlling the candidate planar annotations to perform displacement actions until the optimal position is reached according to the target displacement method and the agent selection strategy, combined with the pre-established DQN model and displacement constraint reward strategy.

[0007] According to the method provided by the application, the candidate planar annotation for determining whether there is a cap conflict in the target region comprises: constructing a respective annotation rectangle for each planar annotation in the target region; judging whether there is a spatial overlap between each annotation rectangle and other annotation rectangles to obtain a judgment result; and taking the planar annotation with the judgment result as the candidate planar annotation.

[0008] According to the method provided by the application, the complexity of each cap conflict event is determined, comprising: determining the number of annotations of the candidate planar annotations participating in the cap in each cap conflict event; if the number of annotations is less than three, the complexity of the cap conflict event is determined to be simple annotation cap; if the number of annotations is greater than three, the complexity of the cap conflict event is determined to be complex annotation cap.

[0009] According to the method provided by the application, the target displacement mode of the candidate planar annotation under the corresponding complexity is determined, comprising: if the complexity of the current cap conflict event is simple annotation cap, the target displacement mode of the candidate planar annotation is fixed displacement mode, wherein the moving direction of the candidate planar annotation in the fixed displacement mode is eight uniformly distributed directions with the center of the planar element as the origin; if the complexity of the current cap conflict event is complex annotation cap, the target displacement mode of the candidate planar annotation is free displacement mode, wherein the moving direction of the candidate planar annotation in the free displacement mode is freely adjusted in the specified buffer area.

[0010] According to the method provided by the application, the agent selection strategy in the corresponding target displacement mode in each cap conflict event is determined according to the conflict importance comprehensive score, comprising: in the fixed displacement mode, the agent selection strategy is: preferentially taking one or two candidate planar annotations with low conflict importance comprehensive score in the cap conflict event as movable agents; in the free displacement mode, the agent selection strategy is: taking one or two candidate planar annotations with high conflict importance comprehensive score in the cap conflict event as non-moving elements, and taking other candidate planar annotations except the non-moving elements as movable agents.

[0011] According to the method provided by the application, the number of coverings, the area of coverings and the semantic importance of each candidate planar annotation are determined, including: determining the number of coverings of each candidate planar annotation with other candidate planar annotations in each covering conflict event, performing normalization processing on the number of coverings to obtain the number of coverings; respectively determining the first total area of coverings of each candidate planar annotation with other candidate planar annotations in each covering conflict event and the second total area of coverings of each candidate planar annotation with planar elements other than the planar element to which the candidate planar annotation belongs, adding the first total area of coverings and the second total area of coverings and performing normalization processing to obtain the area of coverings; determining a target semantic level corresponding to the semantic information of each candidate planar annotation, and determining a semantic importance score corresponding to the target semantic level, and performing normalization processing on the semantic importance score to obtain the semantic importance.

[0012] According to the method provided by the application, the target semantic level corresponding to the semantic information of each candidate planar annotation is determined, including: determining a planning level, information transmission necessary content, industry standard information and element space influence level corresponding to the semantic information of each candidate planar annotation to obtain core information; determining the target semantic level corresponding to the semantic information of the candidate planar annotation from a pre-established core information and semantic level correspondence table according to the core information.

[0013] According to the method provided by the application, the conflict importance comprehensive score of each candidate planar annotation is calculated, including: determining the factor weight value corresponding to each of the number of coverings, the area of coverings and the semantic importance; according to the factor weight value, performing weighted summation on the number of coverings, the area of coverings and the semantic importance of each candidate planar annotation to obtain the conflict importance comprehensive score.

[0014] According to the method provided by the application, the displacement constraint reward strategy is established, including: if there is no covering conflict between the movable agent and other planar annotations during the execution of the displacement action, a positive value is set as the reward value, otherwise, the negative value of the absolute value of the covering area between the movable agent and other planar annotations is enlarged by a first set multiple to obtain a punishment value, to obtain a first covering constraint reward strategy; if there is no covering conflict between the movable agent and planar elements other than the planar element to which the movable agent belongs, a positive value is set as the reward value, otherwise, the negative value of the absolute value of the covering area between the movable agent and planar elements other than the planar element to which the movable agent belongs is enlarged by a second set multiple to obtain a punishment value, to obtain a second covering constraint reward strategy; if the target displacement mode is a free displacement mode, if the movable agent does not exceed the set movement range, a non-negative value is set as the reward value, if the movable agent exceeds the set movement range, a negative value is set as the punishment value, to obtain a movement range constraint reward strategy; the first covering constraint reward strategy, the second covering constraint reward strategy and the movement range constraint reward strategy are taken as the displacement constraint reward strategy.

[0015] According to the method provided by the application, the DQN model comprises: an online network configured to predict an expected return value corresponding to each candidate displacement action by using original fixed parameters, and output action prediction information according to the expected return value; and a target network configured to generate a target return value by using current fixed parameters, wherein the parameters of the target network are copied from the online network at a set period.

[0016] The deep reinforcement learning multi-displacement processing method for the planar annotation cap conflict provided by the application realizes efficient self-adaptive resolution of the planar annotation cap conflict, takes into account the processing efficiency and flexibility of the cap conflict, and guarantees the position stability of the high semantic level planar annotation by means of the comprehensive conflict importance score, thereby avoiding the core geographic information transmission deviation; meanwhile, the combination of the reinforcement learning mechanism of the DQN model and the displacement constraint reward strategy can effectively guide the annotation to avoid new cap conflicts and maintain the relevance with the associated planar element during the displacement process, thereby significantly improving the rationality of the annotation configuration and the readability of the map, solving the pain points of the traditional annotation displacement processing scheme, such as insufficient scene adaptability, low processing efficiency and insufficient reliability, and improving the efficiency, accuracy, flexibility and reliability of the planar annotation cap conflict processing link, thereby providing reliable technical support for high-quantity and high-quality map annotation automatic configuration in the big data era. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0018] Figure 1 is a flowchart of the deep reinforcement learning multi-displacement processing method for the planar annotation cap conflict provided by the embodiment of the application; Figure 2 is a principle diagram of constructing an annotation rectangle for a planar annotation; Figure 3 is a displacement principle diagram under a fixed displacement mode; Figure 4 is a displacement principle diagram under a free displacement mode; Figure 5 is a principle diagram of a full-type planar annotation conflict self-adaptive solution scheme; Figure 6 is an implementation principle diagram of an intelligent agent selection strategy; Figure 7 is an agent selection strategy schematic diagram in fixed displacement mode; Figure 8 is a principle schematic diagram of the moving range constraint in free displacement mode; Figure 9 is a loss function change schematic diagram of the DQN model; Figure 10 is a reward function change schematic diagram of the DQN model; Figure 11 is an adaptability curve schematic diagram of the ant colony algorithm. DETAILED DESCRIPTION

[0019] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below in combination with the drawings in the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in 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.

[0020] The details of the deep reinforcement learning multi-displacement processing method for facing flat annotation cap conflict provided by the embodiments of the present application will be described below. Figures 1 to 11 The details of the deep reinforcement learning multi-displacement processing method for facing flat annotation cap conflict provided by the embodiments of the present application will be described below.

[0021] As shown in Figure 1 The deep reinforcement learning multi-displacement processing method for facing flat annotation cap conflict provided by the embodiments of the present application mainly includes the following steps: Step 110: determining the candidate flat annotations with cap conflict existing in the target region and the complexity of each cap conflict event, and determining the target displacement mode of the candidate flat annotations under the corresponding complexity.

[0022] It can be understood that the displacement mode of the flat annotation can be divided according to the distribution quantity between the flat annotations. Specifically, if the candidate flat annotations involved in the conflict are more, the free displacement mode can be selected, and if the number of candidate flat annotations involved in the conflict is less, the fixed displacement mode can be selected. That is, the specific displacement mode can be determined according to the cap conflict distribution density (i.e. complexity) between the flat annotations.

[0023] This embodiment focuses on the spatial cap conflict problem between the flat annotations corresponding to the flat elements with small area. The solution to such a problem is closely related to the complexity of the actual cap scene, and then different displacement modes can be configured according to different complexity to improve the scene adaptability in the conflict resolution process.

[0024] Step 120: Determine the number of coverings, the area of coverings, and the semantic importance of each candidate planar annotation, and calculate the conflict importance comprehensive score of each candidate planar annotation.

[0025] In practical applications, the difficulty of solving the covering conflict of planar annotations will change with various factors, for example, the more planar annotations involved in the conflict, the more difficult it is to coordinate. Key planar annotations need to be prioritized to ensure their rationality in conflict processing. The differences in different factors will result in different importance of different conflicts. Therefore, by solving the conflict importance comprehensive score, the influence of each factor can be quantified to predict the importance level of each type of conflict, provide key reference for quickly identifying conflicts between annotations and taking displacement methods, and thus make the conflict resolution process more targeted. The embodiment analyzes the three factors of the number of coverings, the area of coverings, and the semantic importance that affect the degree of planar annotation covering, and uses them as influencing factors to participate in the solution of the conflict importance comprehensive score.

[0026] Step 130: According to the conflict importance comprehensive score, determine the agent selection strategy under the corresponding target displacement method in each covering conflict event.

[0027] It can be understood that the core operation logic of the DQN (Deep Q-Network, Deep Q-Network) model relies on the dynamic interaction between the agent and the environment. The agent is the core carrier of decision execution and learning feedback, and plays a key role in perceiving the state from the environment, outputting actions, receiving rewards, and iteratively optimizing strategies. By configuring the agent selection strategy in different conflict scenarios and different displacement methods, the embodiment can improve the scene adaptability of the conflict resolution process, and thus improve the conflict processing efficiency and accuracy.

[0028] Step 140: According to the target displacement method and the agent selection strategy, and combining the pre-established DQN model and the displacement constraint reward strategy, control the candidate planar annotation to execute displacement actions until the optimal position.

[0029] In the embodiment, by adapting the displacement constraint reward strategy for planar annotation conflict processing scenarios, combining the DON model and the pre-determined target displacement method and agent selection strategy, the effective processing of planar annotation covering conflicts can be more efficiently and accurately achieved.

[0030] In an embodiment, the candidate planar annotations with covering conflicts in the target region are determined, specifically including: First, a respective annotation rectangle corresponding to each planar annotation in the target region is constructed.

[0031] In the embodiment, the planar elements can be modeled by annotation rectangles. By constructing a standardized geometric rectangle as a spatial proxy model of the planar annotation, the original planar annotation such asFigure 2 The obtained note rectangle is shown in the A1 subgraph of FIG. 10. Figure 2 The obtained note rectangle is shown in the B1 subgraph of FIG. 10. Figure 2 The C1 subgraph of FIG. 10 shows that the original note can be replaced by the note rectangle in subsequent displacement coordination and conflict resolution processes.

[0032] Then, it is determined whether there is spatial overlap between each note rectangle and other note rectangles, and a determination result is obtained.

[0033] Finally, the planar note with the determination result of yes is taken as a candidate planar note.

[0034] In this embodiment, the candidate planar note with a grommet conflict needs to consider the overlap between planar notes in the selection process to accurately screen the candidate planar note with a grommet abnormality.

[0035] In an embodiment, the complexity of each grommet conflict event is determined, specifically including: First, the number of notes of the candidate planar note participating in grommet in each grommet conflict event is determined.

[0036] In one case, if the number of notes is less than or equal to three, the complexity of the grommet conflict event is determined to be note grommet simple.

[0037] In this case, the grommet conflict event only contains three or fewer candidate planar note grommet conflicts, the conflict relationship is clear, and the influence range is limited, so it can be divided into note grommet simple.

[0038] In another case, if the number of notes is greater than three, the complexity of the grommet conflict event is determined to be note grommet complex.

[0039] In this case, the grommet conflict event involves complex grommet conflicts between more than three candidate planar notes, the conflict relationship is complex, and the influence range is large, so it can be divided into note grommet complex. By dynamically optimizing the positional relationship of all conflict notes, it is ensured that the spatial grommet conflict between planar notes is effectively solved, and the corresponding association between the planar note and the geographic target can be preserved.

[0040] In this embodiment, the number of notes participating in grommet can be used to divide the complexity, specifically into two cases of note grommet simple and note grommet complex.

[0041] In an embodiment, the target displacement mode of the candidate planar note under the corresponding complexity is determined, specifically including: If the complexity of the current overlay conflict event is simple (address overlay is simple), then the target displacement method for the candidate area annotation is a fixed displacement method. Under the fixed displacement method, the movement direction of the candidate area annotation is one of eight evenly distributed directions with the center of the area feature as the origin.

[0042] In scenarios with simple annotation overlay, where only a small number of annotations overlap, a fixed displacement method can effectively resolve conflicts. The advantage of the fixed displacement method is that it can limit the movement range of area annotations to the greatest extent, thereby ensuring a high degree of correlation between area annotations and their corresponding area features.

[0043] In the fixed displacement mode, an 8-directional movement strategy is specifically adopted, such as... Figure 3 As shown in sub-figure A2, the movement range of the planar annotations is limited to eight evenly distributed directions with the center of the planar element as the origin. That is, the planar annotations can only move in the up, down, left, right, and four diagonal directions, and stop at these eight preset positions. By selectively moving the planar annotations to these eight preset positions, the overlapping conflict between planar annotations can be effectively resolved. The specific effect is as follows: Figure 3 As shown in subgraph B2.

[0044] If the complexity of the current overlay conflict event is annotation overlay complexity, then the target shifting method for the candidate surface annotation is the free shifting method. In the free shifting method, the movement direction of the candidate surface annotation is freely adjusted within the specified buffer.

[0045] In complex scenarios where multiple planar annotations overlap each other, the limited movement range allowed by fixed displacement methods makes it difficult to completely eliminate overlap conflicts. To address this, this embodiment introduces a free displacement method. Specifically, using the center point of the planar element as a reference point and adding guide lines, the movement range of the planar annotations is expanded, allowing them to explore suitable placement positions within a wider space, thus effectively solving the problem of multiple annotations overlapping in complex scenarios.

[0046] In the free displacement mode, the overlap conflict between multiple planar annotations is quite complex, for example... Figure 4 As shown in sub-figure A3, a better solution needs to be found over a wider range using free displacement. Therefore, area annotations allow for free adjustment within a specific spatial range. To clarify the relationship between area annotations and their corresponding area features, guide lines are configured for the area annotations to indicate their target. Ultimately, by optimizing the spatial position of the area annotations, the overlap conflict is resolved, as detailed below. Figure 4 As shown in sub-figure B3.

[0047] The free-shifting method grants planar annotations the ability to flexibly adjust within the permitted space, freeing them from the constraints of fixed candidate positions. It also allows agents to perform more refined searches and explorations of potential conflict-free areas. Therefore, the free-shifting method can effectively adapt to challenging scenarios such as multiple annotation overlays and complex building distributions, significantly improving the efficiency of resolving spatial overlap conflicts. Specifically, for example… Figure 4 The C3 and D3 subgraphs are shown in the diagram.

[0048] By flexibly configuring appropriate displacement methods based on the complexity of different conflict scenarios, the spatial overlap conflict of area annotations can be resolved in a case-by-case and targeted manner. For simple annotation overlap scenarios, fixed displacement is preferred; for complex overlap scenarios, free displacement is selected. Through the organic integration and synergistic application of the two displacement methods, the method's adaptability to diverse spatial distribution scenarios of annotations is significantly improved, ensuring that overlap conflicts of various area annotations in a single map can be comprehensively and efficiently resolved, thereby guaranteeing the clarity of map visualization and the accuracy of information transmission.

[0049] In one embodiment, determining the number of overlays, the overlay area, and the semantic importance of each candidate surface annotation specifically includes: On the one hand, determine the number of times each candidate face annotation overlaps with other candidate face annotations in each overlap conflict event, and normalize the number of overlaps to obtain the number of overlaps.

[0050] The number of overlapping conflicts caused by areal annotations determines the difficulty of resolving these conflicts. Spatial overlapping conflicts between multiple areal annotations are often more difficult to resolve than those between a few. For example, cluster conflicts formed by multiple overlapping areal annotations are far more difficult to resolve than scattered conflicts between two or three areal annotations. Therefore, we use the number of overlapping candidate areal annotations as an influencing factor to identify the importance of their conflicts and construct a quantitative function to achieve objective evaluation. First, we construct a statistical function for the number of overlapping areal annotations. m Any overlay conflict event of a candidate facet annotation S , can be represented as follows: (1) annotation of any candidate surface The number of times it overlaps with other candidate face annotations in this conflict is counted, i.e.: (2) in, For the first i Candidate facet annotations Number of times the cap is pressed, For indicator functions, if the first... i Candidate facet annotations With thej Candidate facet annotations If a space cap exists, then ,on the contrary .

[0051] Then, the number of capping cycles is normalized to obtain a standardized number of capping cycles, as shown in the formula: (3) (4) in, For the first i Candidate facet annotations The number of caps, The maximum original number of caps under the current conflict event, if This indicates no conflict. .

[0052] On the other hand, the sum of the first overlapping area of ​​each candidate face annotation with other candidate face annotations and the sum of the second overlapping area with face elements that do not belong to it are determined in each overlapping conflict event. The sum of the first overlapping area and the sum of the second overlapping area are added together and normalized to obtain the overlapping area.

[0053] When spatial overlap conflicts occur between areal annotations, the difficulty of resolving the conflict increases with the size of the overlap area. It's important to note that the overlap area must encompass two core scenarios: first, the area where conflicting annotations overlap with each other; and second, the area where a conflicting annotation overlaps with an areal element that it doesn't belong to. The sum of these two areas directly determines the severity of the conflict in the overlap dimension. The larger the sum, the more significant the damage to annotation readability and element relevance, and the higher the importance of the conflict. Therefore, a quantification function needs to be constructed to calculate the combined impact of these two areas. First, the sum of the first overlap areas of the candidate areal annotation and other candidate areal annotations is calculated using the following formula: (5) in, Indicates the first i Candidate facet annotations The sum of the areas, Indicates the first i Candidate facet annotations The sum of the first overlay areas between the annotations and other candidate areal annotations. Indicates the first i Candidate facet annotations The sum of the second overlapping areas between itself and other non-self-owned areal features.

[0054] The sum of the areas is normalized as follows: (6) (7) wherein, is the i cap area of the i-th candidate planar label, is the total cap area maximum of all candidate planar labels in the current conflict event. If represents no cap, then . .

[0055] In another aspect, a target semantic level corresponding to semantic information of each candidate planar label is determined, and a semantic importance score corresponding to the target semantic level is determined. The semantic importance score is normalized to obtain a semantic importance degree.

[0056] In one specific implementation, the target semantic level corresponding to the semantic information of each candidate planar label is determined, specifically including: First, a planning level, information transmission essential content, industry standard information, and element space influence level corresponding to the semantic information of each candidate planar label are determined to obtain core information.

[0057] Then, according to the core information, the target semantic level corresponding to the semantic information of the candidate planar label is determined from a pre-established core information and semantic level correspondence table.

[0058] The semantic information of the planar element label is a key indicator that defines the core attribute of the planar element, and directly determines the information value of the label in spatial labeling. For example, government buildings and nature reserves, which are high semantic level planar labels, have much higher information transmission priority than low semantic level planar labels such as residential areas and temporary construction sites. When planar labels of different semantic levels overlap, the importance of the conflict also varies with the difference in semantic value of the labels: the overlap conflict of high semantic labels directly affects the transmission of core spatial information, and may even lead to misjudgment of the element attribute by the user, so the importance of the conflict is much higher than that of the overlap conflict of low semantic labels.

[0059] Based on this, when dealing with label overlapping problems, the semantic information of the labels in conflict needs to be classified and judged first. In this embodiment, the semantic information of the planar label is divided into three semantic levels of core level, important level, and ordinary level according to the importance of the element. In actual application, the label semantic classification is mainly realized according to four types of basis, i.e., planning level of planar element, information transmission essential content, industry standard information, and element space influence level.

[0060] Specifically, if the planning level is a national level, a provincial administrative unit or a core planning area, the information transmission necessary content involves unique identification or security control non-substitutable information, the industry standard information involves compliance with the highest priority labeling requirements of the state or industry, and the element space influence level is wide area radiation, it can be divided into the core level; if the planning level is a district, a county administrative unit or a planning area, the information transmission necessary content involves supplementary key information such as function or ownership, the industry standard information involves compliance with the priority labeling requirements of the industry, and the element space influence level is regional radiation, it can be divided into the important level; if the planning level is a community, a street unit or ordinary land, the information transmission necessary content involves auxiliary information such as characteristics or temporary attributes, the industry standard information involves compliance with the basic labeling requirements of the industry, and the element space influence level is local radiation, it can be divided into the ordinary level.

[0061] After determining the target semantic level, the semantic importance score can be determined according to the following formula: (8) wherein, is the semantic importance score of the i-th candidate planar annotation. i is the semantic importance score of the i-th candidate planar annotation.

[0062] Further, the semantic importance score is normalized, specifically as follows: (9) wherein, is the semantic importance score of the i-th candidate planar annotation. i is the maximum value of the semantic original score of all candidate planar annotations in the current conflict event, is the minimum value of the semantic original score of all candidate planar annotations in the current conflict event, and if there is a single level annotation in the conflict event, . In an embodiment, the conflict importance comprehensive score of each candidate planar annotation is calculated, specifically including: First, the factor weight values corresponding to the number of gaskets, the area of gaskets and the semantic importance are determined.

[0063] Then, according to the factor weight values, the number of gaskets, the area of gaskets and the semantic importance of each candidate planar annotation are weighted and summed to obtain the conflict importance comprehensive score.

[0064]

[0065] ​​​To comprehensively evaluate the importance of the candidate planimetric label cap conflict, the cap number, cap area, and semantic importance are integrated as three core influencing factors, and the multi-dimensional information is quantitatively fused by weighted summation. The weights of the three factors are determined according to the influence degree of each factor on the priority of conflict resolution in practical application. Specifically, the factor weight value of the cap number is set to 0.4, the factor weight value of the cap area is set to 0.3, and the factor weight value of the semantic importance is set to 0.3. Based on this, the calculation formula of the comprehensive score of the conflict importance is as follows: (10) wherein, is the comprehensive score of the conflict importance of the i-th candidate planimetric label, i i is the factor weight value of the cap number, s norm is the factor weight value of the cap area, norm and is the factor weight value of the semantic importance.

[0066] In this embodiment, a hierarchical progressive agent selection mechanism is designed, and the movable agent is accurately determined by combining the label priority and the spatial topological relationship. In view of the diversity of conflict resolution scenarios, a fixed and free dual-mode shift framework is proposed, the strategy optimization is realized through the DQN model, and an action decision system is constructed, which includes the constraints of label cap between labels, label cap between labels and other planimetric elements, and label moving range. Finally, a full-type planimetric label conflict adaptive solution covering simple to complex scenarios is formed, and the specific process is shown in Figure 5

[0067] Referring to Figure 5 , in the steps of cap conflict element extraction and importance sorting, the conflict of the planimetric label in the target area can be quantified first, and the cap number N norm , the cap area A norm and the semantic importance W norm are counted. The comprehensive score of the conflict importance of each candidate planimetric label is obtained by weighting, and the priority of the conflict label is sorted. At the same time, based on the complexity of the conflict event, the fixed shift mode (i.e. the fixed model in Figure 5 ) or the free shift mode (i.e. the free model in Figure 5 ) is matched, and the corresponding movable agent is constructed.

[0068] In the training and experience pool interaction link of the DQN model, the interaction data of the agent and the environment is stored in the experience pool as the core, which involves the state s t and the action a ​t , reward r t and other parameters, double-network training of the current network and the target network. The current network can output the Q value of the action to guide the agent to select the shift action; the target network can calculate the target Q value to optimize the current network parameters through the loss function; the target network parameters are kept stable through soft update to avoid training fluctuations.

[0069] In the constraint reward condition judgment link, the agent can be guided to reasonably shift through dynamic rewards and static rewards. Dynamic rewards limit the cover between inscriptions and non-associated planar elements, and static rewards limit the movement range. The combination of the two types of rewards can ensure that the agent shift not only resolves the cover, but also maintains the relevance to the associated planar elements.

[0070] In the reasonable scheme output link, after the DNQ model training is completed, the final configuration scheme without cover and stable inscription position is output according to the scheme layout arrangement and the change of the reward function.

[0071] In actual application, whether the agent selects reasonably or not is related to the correctness of the map inscription position, and also greatly affects the conflict processing efficiency. For example, for a simple scenario of inscription cover that can be solved by fixed shift, one of the candidate planar inscriptions that causes spatial cover conflict can be selected as a movable agent to solve the spatial cover conflict. However, if all candidate planar inscriptions are selected as movable agents to participate in model calculation, too many agents will cause unnecessary redundant calculation, thereby affecting the training convergence speed of the model, the final strategy performance and the overall running efficiency, and also affecting the correctness of the inscription position.

[0072] Therefore, only the party involved in the conflict needs to be reasonably selected as a movable agent to participate in the shift to solve the conflict, thereby improving the calculation efficiency. Similarly, for the free shift mode, the candidate planar inscription that causes the cover needs to be reasonably selected as a movable agent to participate in the shift, so as to ensure the correctness of the map inscription and improve the algorithm running efficiency. The agent selection strategy in the embodiment can be referred to in Figure 6 .

[0073] In an embodiment, the agent selection strategy under the corresponding target shift mode in each cover conflict event is determined according to the conflict importance comprehensive score, and specifically includes: In the fixed shift mode, the agent selection strategy is: preferentially selecting one or two candidate planar inscriptions with low conflict importance comprehensive score in the cover conflict event as a movable agent.

[0074] In combination with Figure 6As shown, in the agent selection strategy, for a binary annotation conflict event containing only two candidate planar annotations, the lower-scored lower-level annotation can be preferentially selected as the movable agent for displacement adjustment based on the conflict importance comprehensive score, as shown in Figure 7 As shown in the A4 subgraph, if the conflict is not resolved, the higher-scored higher-level annotation is used to replace the displacement, as shown in Figure 7 As shown in the B4 subgraph, for a ternary annotation conflict event containing three candidate planar annotations, the two candidate planar annotations with lower scores are preferentially moved, as shown in Figure 7 As shown in the C4 subgraph, if the above-mentioned selection for binary and ternary annotations cannot achieve conflict resolution, a free displacement method is needed to select the agent. Figure 7 As shown in the D4 subgraph, if the above-mentioned selection for binary and ternary annotations cannot achieve conflict resolution, a free displacement method is needed to select the agent.

[0075] In the free displacement method, the agent selection strategy is to select one or two candidate planar annotations with high conflict importance comprehensive scores as non-movable elements, and select the other candidate planar annotations as movable agents.

[0076] As shown in Figure 6 For complex grommet scenes composed of more than three annotations, the free displacement method needs to be started, and one or two candidate planar annotations with the highest scores need to be fixed as non-movable elements, and the remaining candidate planar annotations need to be included in the movable agent set for global optimization configuration. This hierarchical strategy realizes systematic conflict resolution from simple to complex scenes by combining annotation importance level evaluation and spatial relationship analysis.

[0077] In an embodiment, the DQN model specifically includes: An online network for predicting the expected return value corresponding to each candidate displacement action through the original fixed parameters, and outputting action prediction information according to the expected return value.

[0078] A target network for generating a target return value through the current fixed parameters, and the parameters of the target network are copied from the online network at a set period.

[0079] In this embodiment, as shown in Figure 5 The DQN model adopts a double-network structure design, including a current network and a target network. The online network is a neural network actually used to generate action selection in the training process, which mainly calculates the Q value of each action according to the current state, and the agent selects the corresponding action through the size of the Q value. Its formula is: (11) Wherein, represents environment state at time t, represents action selected at time t, represents online network parameters, represents network forward propagation process, output dimension equals to the number of all possible actions, takes action corresponding dimension value, i.e., the value of this action, .

[0080] The target network is a copy of the DQN model introduced to calculate the target Q value. The role of the target network is to maintain stability during training. Its calculation formula is: (12) where, represents target value at time t, represents reward environment obtained after performing action at time t, is a discount factor, representing the weight of future rewards. The larger the value, the more emphasis on long-term rewards, represents t next state at time t+1, represents the next state action output by the target network, value of the next action, represents the maximum value obtained for all possible actions on the next state, represents the t +1 step action that can be selected, represents the parameters of the target network.

[0081] The DQN model internally sets an experience replay mechanism. Each experience (i.e., state, action, reward, next state) obtained during the interaction between the agent and the environment is stored in the experience pool. When the experience pool is full, new experiences will overwrite the oldest experiences, ensuring that the latest interaction data is stored in the replay pool. During training, the model continuously selects a batch of experiences from the experience pool at random to calculate the target Q value. Then, the difference between the Q value calculated by the current network and the target Q value is calculated, i.e., the loss function calculation. Finally, the loss function is minimized through gradient descent to update the parameters of the current network. The specific formula is: (13) where, is the loss function value, An experience pool is configured to store the action experience of the agent and participate in random sampling in experience replay to reduce sample correlation and improve training stability. The expectation of the state-action-reward-next state sample sampled from the experience pool is represented, that is, the average loss of the batch sample represents the expected loss of the whole.

[0082] In an embodiment, a displacement constraint reward strategy is established, specifically including: During the execution of the displacement action, if there is no overhanging conflict between the movable agent and other planar annotations, a positive value is set as the reward value, otherwise, the negative value of the absolute value of the overhanging area between the movable agent and other planar annotations is enlarged by a first set multiple to obtain a first overhanging constraint reward strategy.

[0083] It can be understood that although the two displacement methods are different, they both ensure that the agent does not overhang with other annotations during movement, which is a core goal to be achieved to solve the conflict. Therefore, an annotation overhanging constraint needs to be constructed for the two displacement methods. Through the construction of this constraint condition, the agent annotation can be guided not to overhang with other annotations during movement. If overhanging occurs, the overhanging area between the annotation and other annotations is calculated , and the negative value of the absolute value of the overhanging area is taken as the basis for punishment, and multiplied by to increase the significance of the punishment. If no overhanging occurs, a reward of 1 is given to indicate that no punitive constraint is imposed on the action. The calculation formula of the reward value in the first overhanging constraint reward strategy is specifically: (14) If there is no overhanging conflict between the movable agent and the planar elements not belonging to itself, a positive value is set as the reward value, otherwise, the negative value of the absolute value of the overhanging area between the movable agent and the planar elements not belonging to itself is enlarged by a second set multiple to obtain a second overhanging constraint reward strategy.

[0084] When the candidate planar annotation as the movable agent participates in movement, this embodiment regards the overhanging between the candidate planar annotation and the planar elements belonging to itself as reasonable overhanging. In addition to ensuring that the final movement position of the annotation does not overhang with other annotations, it also needs to ensure that the annotation does not overhang with other planar elements not belonging to itself. When the candidate planar annotation overhangs with the planar elements not belonging to itself during action, the overhanging area is calculated, and the negative value of the absolute value of the overhanging area is taken as the basis for punishment, and multiplied by The amplification coefficient increases the significance of the penalty. Compared to the annotation-overlay constraint, its amplification coefficient is an order of magnitude smaller, distinguishing the levels of constraints between two different constraints. This indicates that resolving the annotation-overlay constraint is the primary focus when resolving spatial overlay conflicts between annotations. If no overlay occurs, a reward value of 1 is assigned, indicating no penalty constraint on its action. Therefore, the reward value in the second overlay constraint reward strategy is... The specific calculation formula is as follows: (15) If the target displacement method is free displacement, if the movable agent does not exceed the set displacement range, a non-negative value will be set as the reward value; if the movable agent exceeds the set displacement range, a negative value will be set as the penalty value, resulting in a displacement range-constrained reward policy.

[0085] The core difference between free-movement and fixed-movement methods lies in their approaches. Fixed-movement methods iterate through preset discrete positions to find a suitable target location, while free-movement methods achieve this through continuous movement. While this continuous movement enhances exploration flexibility, it also carries the risk of excessive deviation. Without constraints on its movement range, the agent might explore areas unrelated to the isometric features to which the annotation belongs, significantly weakening the correlation between the annotation and the feature, thus violating the core principles of annotation.

[0086] Therefore, in the free-movement method, it is necessary to construct clear movement range constraints. By defining reasonable exploration boundaries, the advantages of continuous movement in complex conflict scenarios can be retained, while ensuring that the annotation always explores within a space strongly related to the elements. Ultimately, a balance is found between eliminating overlay and maintaining relevance. In this embodiment, the movement range is defined by extending a certain range outward from the conflict annotation itself, such as... Figure 8 As shown in subgraph A5, after processing by the DQN model, conflicting annotations find suitable solutions within their respective movement ranges, as follows: Figure 8 As shown in subgraphs B5 and C5, the conflict between annotations is ultimately resolved, as follows: Figure 8 As shown in the D5 sub-figure.

[0087] By constructing a conflict annotation movement range constraint for the free-shifting method, if a candidate face annotation shifted using the free-shifting method exceeds the constraint range, a penalty value of -10 is assigned; conversely, a value of 0 is assigned as a reward value. The reward value in the movement range constraint reward strategy... The specific formula is: (16) The first capping constraint reward strategy, the second capping constraint reward strategy, and the movement range constraint reward strategy are used as the displacement constraint reward strategy.

[0088] When the agent performs an action, the above constraint reward is judged, and the total reward value can be obtained by adding the reward values obtained by judging each constraint condition , the specific formula is: (17) The total reward value is calculated whether it is greater than or equal to 0 to determine whether it meets all the above constraints, if it is 0, the current round of movement is ended, and the next round continues to train, if it is less than 0, it continues to move until the constraint condition is met or the agent displacement limit is reached, then the training of this round can be ended.

[0089] In order to verify the effect that the above method provided by the embodiment can achieve, this time the data of the research area containing 201 planar elements is selected to verify the effectiveness of the above method. The research area contains 201 planar annotations, and the conflict importance comprehensive score of the annotation data in the research area is calculated to extract the candidate planar annotations involved in the overlay conflict and the importance ranking between the candidate planar annotations. After analysis, there are a total of 53 candidate planar annotations with spatial overlay conflicts in the research area. After sorting according to the conflict importance comprehensive score, the specific information of the top 10 candidate planar annotations can be seen in Table 1.

[0090] Table 1 Specific information of part of the candidate planar annotations in the research area

[0091] After completing the calculation of the conflict importance comprehensive score, a series of fine processing procedures are performed on the annotations in the research area. First, the annotation rectangle corresponding to each planar annotation is extracted, and these highlighted annotation rectangles are used as substitutes for the annotations. Then, the displacement mode is determined according to the difference in the number of overlays, and the movable agent is selected according to the agent selection strategy.

[0092] When processing the annotation spatial overlay conflict, the DQN model parameters are set as follows: the experience pool is set to 10000 (sufficient storage experience to avoid overfitting), and the training batch is 64; the initial exploration rate is 0.8, and the decay rate is 0.995 (initially explore fully, and later tend to the optimal strategy), the total training batch is 600 times, and the target network is updated every 10 rounds (balance learning effect and calculation efficiency); the learning rate is 0.001 (to ensure learning efficiency), and the discount factor is 0.99 (emphasis on long-term rewards); the action step of each agent is 0.00001, and the action upper limit of fixed displacement and free displacement is set to 15 times and 20 times respectively (end when the upper limit is met or the constraint condition is met, and all agents are traversed in turn). For the overlay conflict scenario, each agent iterates for 600 rounds.

[0093] In simple overlay conflict scenarios, due to the relatively simple conflict relationship and the relatively simple interaction logic between annotations, the model took a total of 2508.152 seconds after a series of calculations and adjustments. However, in complex overlay conflict scenarios, due to the existence of multi-level overlay relationships and the intricate mutual influence between annotations, the model needs to consume more resources for conflict analysis and position optimization, and the total time reached 5849.29 seconds.

[0094] Further calculations show that the average time cost for a single agent to complete a single round of calculation in a simple overlay conflict scenario is 0.21 seconds. This means that the agent can quickly adjust its position to resolve the conflict in this scenario. However, in a complex overlay conflict scenario, the average time cost for a single agent to complete a single round of calculation is 0.98 seconds, reflecting that the process of position optimization by the agent is more time-consuming in complex scenarios. From the overall time consumption perspective, the model still maintains high processing efficiency for handling overlay problems between annotations. Table 2 shows the statistical data of the processing results.

[0095] Table 2 Statistical data of processing results

[0096] As shown in Table 2, the agents using fixed shift and free shift methods achieved success rates of 100% and 78.6% respectively, demonstrating highly efficient processing capabilities. The change in the loss function of the DQN model during conflict resolution is as follows: Figure 9 As shown.

[0097] The change in the loss function under a simple capping environment shows that... Figure 9 The loss function curves of the first twenty agents show a high degree of similarity during the training of the DQN model, all exhibiting a significant decrease and a tendency to plateau. This indicates that the model performance is stabilizing. This phenomenon verifies the effectiveness of the DQN algorithm in adaptive parameter adjustment; that is, by continuously optimizing the neural network weights, the model can gradually approach the optimal solution. Overall, the changes in its loss function demonstrate that the DQN model exhibits strong robustness in spatial occlusion conflict resolution tasks, achieving efficient conflict resolution even when network parameters are not completely stable through dynamic adjustment.

[0098] Analyzing the changes in the loss function under complex capping environments, and observing... Figure 9The loss function change curve corresponding to the last ten agents shows that, due to the existence of multiple complex cap conflicts between annotations in the complex cap environment, and the complex distribution of planar elements, the difficulty of conflict resolution in this environment is much higher than that of the simple cap environment involving only cap conflicts between annotations. Specifically, the fluctuation range of the loss function in the complex cap environment is significantly larger than that in the simple cap environment, and the downward trend is also weaker. This phenomenon reflects that the DQN model needs to continuously adjust parameters and strategies to find the optimal solution to the spatial cap conflict between annotations when dealing with such complex conflicts.

[0099] On the other hand, the reward function of the DQN model is as shown in Figure 10 From the change of the reward function in the simple cap environment, Figure 10 The reward values of the first twenty agents all show a sustained upward trend, which intuitively reflects the gradual adaptation of the agents to the environment and the optimization of their performance during the training process. And under the guidance of their respective reward functions, different agents have successfully eliminated the spatial cap conflict between annotations, indicating that the training state of the current agent has met the core requirements of the task. In the complex cap environment, a deep analysis of the dynamic evolution process of the last ten reward functions shows that the reward functions of the overall agents all show a certain upward trend. Further combined with the intuitive data and trends presented in Figure 10 The change trajectory of the current reward function clearly indicates that the system has successfully achieved the optimal conflict solution, effectively verifying the effectiveness and reliability of the DQN model with the free displacement mechanism in solving complex spatial cap conflicts between annotations.

[0100] The following group of comparative tests further verifies the advantages of the above-mentioned method provided by the embodiment. In the first group of comparative tests, on the one hand, the traditional ant colony algorithm is used to solve the spatial cap conflict between annotations, and on the other hand, the optimization effect of different single displacement methods is tested simultaneously, and the advantages and disadvantages of each method are clearly determined through multi-dimensional comparison. Among them, the core design of the ant colony algorithm is as follows: The pheromone concentration marks the quality of the agent's action; the heuristic function evaluates the position adaptability according to the cap area. The action selection combines greedy selection and roulette selection to balance convergence efficiency and search breadth; the pheromone update dynamically guides the search by old pheromone evaporation and new pheromone release to approach the global optimal solution. Combined with the characteristics of the annotation conflict scene, the key parameters are set as follows: pheromone evaporation coefficient 0.15, pheromone importance factor 1.2, heuristic function importance factor 1.7, greedy selection probability 0.7, and pheromone increment constant 150.

[0101] For the ant colony algorithm with fixed displacement and free displacement, 600 training rounds are set, the upper limit of each round is consistent with the method provided in this embodiment, which is set to 15 times. The total grommet area of each ant with other marks and buildings after each training round is calculated, and the global fitness is calculated, and the specific formula is as follows: (18) wherein, is the fitness value, is the total grommet area of the mark.

[0102] In the comparative test, the conflicting mark elements are compared to the ants in the ant colony algorithm, and the movement of the mark to a position that does not overlap with any other element is compared to the food source in the ant colony algorithm. Through this mapping, the logical conversion from the ant optimization path to the continuous movement of the conflicting mark to find a conflict-free position is realized, and the goal of resolving the spatial overlap conflict between marks is ultimately achieved. The statistical data of the processing results of the ant colony algorithm can be seen in Table 3.

[0103] Table 3 Statistical data of the processing results of the ant colony algorithm

[0104] As can be seen from Table 3, the success rate in handling the overlap conflict between marks is 25.6% and 0% respectively, which is significantly different from the success rate of 94.3% of the method provided in this embodiment. This sharp contrast highlights the superiority of the method provided in this embodiment in handling such problems compared to the ant colony algorithm.

[0105] The fitness curve of the ant colony algorithm is shown in Figure 11 According to the analysis of the dynamic changes of the curve, the curve shows a very significant oscillation characteristic. Specifically, during the algorithm running process, the fitness value fluctuates frequently in a large range, lacking a stable downward or upward trend. In sharp contrast, the reward function used in the method provided in this embodiment exhibits a highly convergent state, and under the same number of iterations or running computation, the reward function value can quickly and stably approach a relatively fixed optimal value. This sharp difference indicates that the ant colony algorithm cannot achieve efficient convergence under the same running computation setting. Further observation of the data in the curve shows that when different displacement models are used, the fitness curve of the ant colony algorithm shows a similar trend. Specifically, regardless of how the displacement mode is adjusted, the fitness curve always fails to show a clear convergence trend, but continues to fluctuate within a certain range. This phenomenon confirms that when dealing with such map mark overlap problems, even if the displacement mode is changed as a key parameter, the algorithm cannot overcome its learning bottleneck and effectively optimize the search process to find an ideal solution.

[0106] The above comparative test results fully verify that the method provided in the embodiment has significant advantages in conflict processing efficiency, algorithm robustness and engineering applicability, etc.

[0107] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A deep reinforcement learning multi-displacement processing method for surface annotation overlay conflicts, characterized in that, include: Identify candidate surface annotations with overlay conflicts within the target area and the complexity of each overlay conflict event, and determine the target displacement mode of the candidate surface annotations under the corresponding complexity level; The number of overlays, the overlay area, and the semantic importance of each candidate face annotation are determined, and the comprehensive conflict importance score of each candidate face annotation is calculated. Based on the comprehensive score of conflict importance, the agent selection strategy for the corresponding target shifting method in each overlay conflict event is determined; Based on the target shifting method and the agent selection strategy, and combined with the pre-established DQN model and shift constraint reward strategy, the candidate surface annotation is controlled to perform shifting actions until the optimal position is reached.

2. The deep reinforcement learning multi-displacement processing method for surface-oriented annotation overlay conflicts according to claim 1, characterized in that, Identify candidate areal annotations within the target area that indicate overlap conflicts, including: Construct a corresponding annotation rectangle for each faceted annotation within the target area; Determine whether there is spatial overlap between each annotation rectangle and other annotation rectangles, and obtain the determination result; The face annotations that are determined to be true are selected as candidate face annotations.

3. The deep reinforcement learning multi-displacement processing method for surface-oriented annotation overlay conflicts according to claim 1, characterized in that, Determine the complexity of each capping conflict event, including: Determine the number of candidate facet annotations involved in each capping conflict event; If the number of annotations is three or less, the complexity of the overprinting conflict event is determined to be simple annotation overprinting. If the number of annotations is greater than three, then the complexity of the overprint conflict event is determined to be annotation overprint complexity.

4. The deep reinforcement learning multi-displacement processing method for surface-oriented annotation overlay conflict according to claim 3, characterized in that, Determining the target shifting method for the candidate planar annotations at the corresponding complexity level includes: If the complexity of the current overlay conflict event is simple (overlay of annotation), then the target displacement method of the candidate surface annotation is a fixed displacement method, wherein, under the fixed displacement method, the movement direction of the candidate surface annotation is eight evenly distributed directions with the center of the surface element as the origin; If the complexity of the current overprint conflict event is complex (overprinting of annotations), then the target shifting method of the candidate surface annotation is a free shifting method, wherein, under the free shifting method, the moving direction of the candidate surface annotation is freely adjusted within a specified buffer.

5. The deep reinforcement learning multi-displacement processing method for surface-oriented annotation overlay conflicts according to claim 4, characterized in that, Based on the comprehensive score of conflict importance, the agent selection strategy for each overlay conflict event under the corresponding target shifting method is determined, including: In the fixed displacement mode, the agent selection strategy is as follows: prioritize one or two candidate face annotations with low overall conflict importance scores in the overlay conflict event as movable agents; In the free-shifting mode, the agent selection strategy is as follows: one or two candidate facets with high comprehensive scores of conflict importance in the overlay conflict event are selected as non-movable elements, and all other candidate facets besides the non-movable elements are selected as movable agents.

6. The deep reinforcement learning multi-displacement processing method for surface-oriented annotation overlay conflicts according to claim 1, characterized in that, Determine the number of overlays, the overlay area, and the semantic importance of each candidate facet annotation, including: Determine the number of times each candidate face annotation overlaps with other candidate face annotations in each overlap conflict event, and normalize the number of overlaps to obtain the number of overlaps. For each candidate facet annotation, determine the sum of the first overlap area with other candidate facet annotations and the sum of the second overlap area with facet elements that do not belong to it in each overlap conflict event. Add the sum of the first overlap area and the sum of the second overlap area and normalize them to obtain the overlap area. The target semantic level corresponding to the semantic information of each candidate face annotation is determined, and the semantic importance score corresponding to the target semantic level is determined. The semantic importance score is normalized to obtain the semantic importance.

7. The deep reinforcement learning multi-displacement processing method for surface-oriented annotation overlay conflicts according to claim 6, characterized in that, Determine the target semantic level corresponding to the semantic information of each candidate facet annotation, including: Determine the planning level, necessary information transmission content, industry standard information, and spatial influence level of each candidate surface annotation to obtain core information; Based on the core information, the target semantic level corresponding to the semantic information of the candidate surface annotation is determined from the pre-established core information and semantic level comparison table.

8. The deep reinforcement learning multi-displacement processing method for surface-oriented annotation overlay conflicts according to claim 1, characterized in that, The overall conflict importance score for each candidate facet annotation is calculated, including: Determine the factor weight values ​​corresponding to the number of caps, the area of ​​the caps, and the semantic importance; Based on the aforementioned factor weight values, the number of overlays, the overlay area, and the semantic importance of each candidate surface annotation are weighted and summed to obtain a comprehensive score for conflict importance.

9. The deep reinforcement learning multi-displacement processing method for surface-oriented annotation overlay conflicts according to claim 1, characterized in that, Establish a shift-constrained reward strategy, including: During the execution of the displacement action, if there is no overlap conflict between the movable intelligent agent and other surface annotations, a positive value will be set as the reward value; otherwise, the negative value of the absolute value of the overlap area between the movable intelligent agent and other surface annotations will be multiplied by a first set multiple as the penalty value, thus obtaining the first overlap constraint reward strategy. If there is no overlap conflict between the mobile agent and the non-dominant surface feature, a positive value will be set as the reward value; otherwise, the negative value of the absolute value of the overlap area between the mobile agent and the non-dominant surface feature will be multiplied by a second set factor as the penalty value, thus obtaining the second overlap constraint reward strategy. If the target displacement method is a free displacement method, if the movable agent does not exceed the set displacement range, a non-negative value will be set as the reward value; if the movable agent exceeds the set displacement range, a negative value will be set as the penalty value, thus obtaining a displacement range constraint reward policy. The first capping constraint reward strategy, the second capping constraint reward strategy, and the movement range constraint reward strategy are used as the displacement constraint reward strategy.

10. The deep reinforcement learning multi-displacement processing method for surface-oriented annotation overlay conflicts according to claim 1, characterized in that, The DQN model includes: An online network is used to predict the expected reward value corresponding to each alternative shift action using existing fixed parameters, and output action prediction information based on the expected reward value. A target network is used to generate a target return value using current fixed parameters, the parameters of which are copied from the online network at set intervals.

Images