A layered prompting method, device and equipment for jigsaw puzzle position assistance and a medium

Abstract

The present application relates to the field of artificial intelligence, and discloses a layered prompting method and device for jigsaw position help, equipment and medium, comprising: in response to a user's request for help with the position of a target jigsaw piece, obtaining the jigsaw piece characteristics of the target jigsaw piece; obtaining the whole picture structure information of the whole jigsaw and identifying the current completed area distribution information and constructing the jigsaw state heterogeneous graph, performing message passing and attention reasoning through the graph attention network to obtain the candidate area of the target jigsaw piece; according to the user's current cognitive level, the help round and the convergence degree of the candidate area, selecting the target prompt level from multiple preset prompt levels and outputting the position prompt information. The present application uses the graph attention network to realize intelligent reasoning of the candidate area and adaptively selects the prompt level combined with the user state, so as to help the user to gradually locate through multi-level progressive prompting, improve the flexibility and adaptability of jigsaw position prompting, and improve the education guiding effect.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a layered prompting method, apparatus, device, and medium for providing assistance with puzzle position. Background Technology

[0002] Currently, smart teaching tools such as reading pens can be used in conjunction with puzzles and other media. For example, after children correctly assemble the puzzle pieces, they can use the reading pen to lightly touch the patterns (such as animals, objects, people, etc.) in the puzzle pieces to hear their Chinese and English names, pronunciations, or related knowledge, thus achieving early childhood education through play.

[0003] Children often encounter problems in assembling puzzles on their own and need to seek puzzle assistance. However, traditional puzzle assistance mostly relies on human companionship or directly providing the complete answer, which makes it impossible for puzzle teaching aids to provide adaptive guidance and inspiration based on children's abilities, reducing the flexibility and adaptability of puzzle position prompts. Summary of the Invention

[0004] The main objective of this invention is to provide a layered prompting method, apparatus, device, and medium for puzzle position assistance, aiming to solve the technical problem of insufficient flexibility and adaptability of puzzle position prompts in the prior art.

[0005] The technical solution of the present invention is as follows: The first aspect of this invention provides a layered hint method for providing assistance with puzzle position, comprising: In response to a user's request for help regarding the location of a target puzzle piece, the puzzle piece features of the target puzzle piece are obtained; Obtain the overall structure information of the entire jigsaw puzzle and identify the distribution information of the currently completed areas; Based on the puzzle piece characteristics of the target puzzle piece and the distribution information of the currently completed area, a corresponding puzzle state heterogeneous graph is constructed; The heterogeneous graph of the puzzle state is processed by a graph attention network to perform message passing and attention reasoning, thereby obtaining the matching attention score between the target puzzle piece and each missing position node, and the candidate region of the target puzzle piece is determined based on the matching attention score. Based on the user's current cognitive level, the number of times they have sought help, and the convergence degree of the candidate region, the corresponding target prompt level is selected from multiple preset prompt levels. The location information of the corresponding target puzzle piece is output according to the target prompt hierarchy.

[0006] A second aspect of the present invention provides a layered prompting device for assisting with puzzle position, comprising: The feature extraction module is used to obtain the puzzle piece features of the target puzzle piece in response to a user's request for help in finding the location of the target puzzle piece. The analysis module is used to obtain the overall structure information of the entire jigsaw puzzle and identify the distribution information of the currently completed areas; The graph construction module is used to construct a corresponding heterogeneous graph of the puzzle state based on the puzzle piece characteristics of the target puzzle piece and the distribution information of the currently completed area; The candidate region determination module is used to perform message passing and attention reasoning on the heterogeneous graph of the puzzle state through a graph attention network, obtain the matching attention score between the target puzzle piece and each empty position node, and determine the candidate region of the target puzzle piece based on the matching attention score; The prompt level determination module is used to select the corresponding target prompt level from multiple preset prompt levels based on the user's current cognitive level, the number of times they have sought help, and the convergence degree of the candidate region. The hierarchical prompting module is used to output the position prompting information of the corresponding target puzzle piece according to the target prompting level.

[0007] A third aspect of the present invention provides a computer device including at least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which enables the at least one processor to perform the layered hint method for puzzle position assistance described above.

[0008] A fourth aspect of the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by one or more processors, cause the one or more processors to perform the aforementioned layered hint method for puzzle position assistance.

[0009] Beneficial Effects: This invention discloses a hierarchical prompting method, apparatus, device, and medium for jigsaw puzzle location assistance. Compared to existing technologies, this invention, in response to a user's request for help with the location of a target jigsaw puzzle piece, obtains the puzzle piece features; acquires the overall structure information of the entire puzzle and identifies the distribution information of currently completed areas; constructs a corresponding puzzle state heterogeneous graph based on the puzzle piece features and the distribution information of currently completed areas; performs message passing and attention reasoning on the puzzle state heterogeneous graph using a graph attention network to obtain the matching attention score between the target puzzle piece and each missing position node, and determines the candidate region of the target puzzle piece based on the matching attention score; selects the corresponding target prompting level from multiple preset prompting levels based on the user's current cognitive level, the number of times they have requested help, and the convergence degree of the candidate regions; and outputs the corresponding location prompting information of the target puzzle piece based on the target prompting level. This invention utilizes a graph attention network to achieve intelligent reasoning of candidate regions and adaptively selects the prompting level based on the user's state, thereby helping users gradually locate the puzzle piece through multi-level progressive prompts, improving the flexibility and adaptability of puzzle location prompts, and enhancing the educational guidance effect. Attached Figure Description

[0010] To more clearly illustrate the solutions in this invention, the accompanying drawings used in the description of the embodiments of this invention will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0011] Figure 1 A schematic diagram of an application environment for the layered prompting method for puzzle position assistance provided in an embodiment of the present invention; Figure 2 A flowchart of a layered prompting method for puzzle position assistance provided in an embodiment of the present invention; Figure 3 A schematic diagram of the functional modules of the layered prompting device for puzzle position assistance provided in an embodiment of the present invention; Figure 4 A schematic diagram of the hardware structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention is further described in detail below. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. The embodiments of the invention are described below in conjunction with the accompanying drawings.

[0013] The layered hint method for puzzle position assistance provided in this embodiment of the invention can be applied to, for example... Figure 1In the interactive scenario shown, where the smart reading pen is paired with the puzzle device, there are a terminal device 101, a smart reading pen 102, a network 103, and a server 104. The network 103 serves as a medium to provide a communication link between the terminal device 101, the smart reading pen 102, and the server 104. The network 103 can include various connection types, such as wired and / or wireless communication links (e.g., Bluetooth, Wi-Fi, NFC, etc.).

[0014] Users can use terminal device 101 and smart reading pen 102 to interact with server 104 via network 103 to receive or send messages, etc. A client supporting this interaction method can be installed on terminal device 101, allowing users to log in to view and edit the interaction settings of smart reading pen 102, view puzzle interaction records, etc. Terminal device 101 can be various electronic devices with a display screen and web browsing support, including but not limited to smartphones, tablets, and desktop computers.

[0015] The smart reading pen 102 is used in conjunction with the puzzle piece 105. When the user places the smart reading pen 102 on the puzzle piece 105, it can automatically recognize the identifiable labels on the surface of the puzzle piece, thereby triggering a recognition event and outputting corresponding interactive content in different interactive states. Simultaneously, the user can press the button on the smart reading pen 102 to input voice through its built-in microphone for intelligent voice dialogue, achieving accurate transmission of interactive needs. The smart reading pen 102 integrates motion sensing, near-field communication recognition, voice acquisition, and feedback output functions, enabling it to sense the user's touch, press, and other operations in real time and associate them as corresponding interactive events.

[0016] Server 104 can be a server providing various services, such as analyzing and processing data like puzzle information recognized by the smart reading pen 102 and user-inputted voice commands, generating AI interactive content, and feeding it back to the smart reading pen 102's backend server (this is just an example). Server 104 can parse and process the received puzzle recognition data and user interaction commands, generate adapted interactive feedback content, and then transmit it to the smart reading pen 102 via network 103, where the smart reading pen 102 provides feedback to the user in the form of voice, light, etc. Server 104 can be a cloud server, a distributed system server, or a server integrated with blockchain.

[0017] It should be understood that the number of terminal devices 101, smart reading pens 102, networks 103, and servers 104 mentioned above is merely illustrative. Depending on the implementation needs, there can be any number of terminal devices 101, smart reading pens 102, networks 103, and servers 104.

[0018] like Figure 2As shown, the layered hint method for puzzle position assistance provided in this embodiment of the invention specifically includes the following steps: S201. In response to the user's request for help regarding the location of the target puzzle piece, obtain the puzzle piece features of the target puzzle piece.

[0019] In this embodiment, the location assistance request is a help request initiated by the user through physical or voice interaction when encountering positioning difficulties during the puzzle-solving process, requesting the output of corresponding hierarchical prompts. Specifically, when a child picks up a puzzle piece whose position is uncertain during the puzzle-solving process, they can initiate a location assistance request by pressing and holding a preset button on the reading pen to touch the puzzle piece, saying "Where should this piece go?" after touching it with the reading pen, or clicking the assistance button on the APP interface, etc. After receiving the request, the system first identifies the target puzzle piece and then extracts its multi-dimensional feature vector, providing input conditions for subsequent heterogeneous graph construction and candidate region reasoning. Through a standardized assistance input interface, the user's vague assistance intention is transformed into a target piece feature query that the system can process, ensuring that all subsequent reasoning processes have clear features and avoiding prompt errors caused by unclear target piece identities.

[0020] For example, in a children's puzzle scenario, a child is completing a puzzle with a "forest animals" theme. The child picks up a puzzle piece with blue edges and white wing textures. Unsure of its location, the child presses and holds a preset button on the reading pen before touching the puzzle piece. Near-field communication (NFC) is used to read the puzzle piece's identifier as "P_023". The system then extracts the piece's image semantic features ("sky / bird wings"), color and texture features ("main color sky blue, fine texture granularity"), contour interface features ("top edge convex, right edge concave, bottom edge straight, left edge convex"), and orientation information ("standard upward 0°") from a pre-stored feature data table. These features are organized in a structured vector format, forming the node feature basis for subsequent graph neural network inference.

[0021] S202. Obtain the overall structure information of the entire puzzle and identify the distribution information of the currently completed areas.

[0022] In this embodiment, the overall image structure information is pre-established during the puzzle content creation stage, describing relevant information about the entire puzzle. It stores data such as the puzzle's region division, object relationships, background structure, and piece position mapping. Based on this overall image structure information, the static overall image structure template is aligned with the dynamic completion status through one or more methods, including a pressure sensor array on the puzzle base, a visual recognition module, or a position mapping table for real-time detection. This accurately identifies which positions are correctly occupied, which positions are still empty, and the adjacency relationships between placed pieces and empty positions, thus identifying the current completed area distribution information. By jointly acquiring the overall image structure information and the completion status distribution, a panoramic understanding from the global layout to local vacancies is established, providing a spatial topological basis for node initialization and edge relationship construction of the heterogeneous graph, ensuring that subsequent candidate region reasoning is performed under correct global constraints.

[0023] For example, in the "forest animals" jigsaw puzzle scenario described above, the entire puzzle is divided into 16 rows × 12 columns, totaling 192 position units. The background structure is divided into three semantic regions: a lower layer of grass, a middle layer of trees, and an upper layer of sky. Object relationships include "a deer in the center of the grass, a bird in the upper right corner of the sky, and a tree trunk running through the middle left," etc. Simultaneously, the pressure sensor array on the puzzle base detects in real time that 47 puzzle pieces have been correctly placed. By maintaining an occupancy state vector, completed areas are marked as 1, and empty areas are marked as 0. It also identifies that the theoretical position of the target piece P_023 is surrounded by a nearby puzzle piece P_022, thus identifying completed areas, empty areas, and nearby placed pieces, serving as an important basis for subsequent graph construction.

[0024] S203. Based on the puzzle piece characteristics of the target puzzle piece and the distribution information of the currently completed area, construct a corresponding puzzle state heterogeneous graph.

[0025] In this embodiment, the puzzle state heterogeneous graph is a representation of the actual completed state of the current puzzle, converted into graph structure data. It is constructed by abstracting puzzle pieces and missing positions in the physical space as graph nodes, and abstracting the multi-dimensional relationships between them as heterogeneous edges. Unlike traditional rule-based filtering methods, the puzzle state heterogeneous graph can simultaneously encode multiple relationship types such as spatial adjacency, semantic continuity, and color compatibility. This allows subsequent graph attention networks to perform message passing and relationship reasoning on a unified graph structure, transforming the complex puzzle location problem into a similarity reasoning problem between graph nodes. This overcomes the limitations of rule-based matching and allows the model to automatically learn the contribution weights of different relationship types to candidate region determination.

[0026] For example, in the above "forest animal" puzzle scenario, the puzzle state heterogeneous graph constructed by the system contains three types of graph nodes: the first type is the nodes of the placed puzzle pieces (47 in total), the second type is the nodes of the missing positions (145 in total), and the third type is the target puzzle piece node P_023 (1). The node features of each placed puzzle piece node are spliced ​​together from its pre-labeled features. At the same time, heterogeneous edges are constructed based on the spatial adjacency relationship, semantic continuity relationship and color compatibility relationship between the graph nodes. For example, the empty position (3,6) is theoretically right-adjacent to the placed block P_022, so a spatial adjacency edge is created with a weight of 1.0. Meanwhile, the semantic label of P_022 is "bird / body," and the expected semantic label of the empty position (3,6) is "bird / wings." Both belong to the "bird" semantic object, so a semantically continuous edge is created with a weight of 0.8. Furthermore, the main color of P_022 is a blue-white gradient, and its spatial distance from the sky-blue target block P_023 is less than a preset threshold, so a color-compatible edge is created with a weight of 0.6, and so on. Through the construction of these multi-type edges, a heterogeneous graph of the puzzle state is obtained, thus fully displaying the multi-dimensional topological relationships of the current puzzle state.

[0027] S204. The heterogeneous graph of the puzzle state is processed by a graph attention network to perform message passing and attention reasoning, so as to obtain the matching attention score between the target puzzle piece and each empty position node, and the candidate region of the target puzzle piece is determined according to the matching attention score.

[0028] In this embodiment, the Graph Attention Network (GAT) is a graph neural network based on an attention mechanism, capable of adaptive message passing on heterogeneous graphs. Therefore, the GAT performs message passing and attention inference on the heterogeneous graph of the puzzle state. Specifically, through a multi-head attention mechanism, each empty position node dynamically calculates the contribution weights of different neighboring nodes based on the features of its neighboring placed puzzle pieces, thereby aggregating a high-order feature representation containing contextual semantics. Finally, candidate regions of the target puzzle piece are determined based on the matching attention scores between the target puzzle piece and each empty position node. Compared to candidate region determination based on matching rules, GAT can automatically learn the importance of each selection dimension through a data-driven approach and capture long-distance semantic associations (such as indirect associations between puzzle pieces in the sky region). Through multi-layer message passing, candidate region inference is completed within a unified probabilistic framework, outputting the matching probability of each empty position as the correct position of the target piece, effectively improving the accuracy and generalization ability of candidate region determination.

[0029] For example, in the "forest animals" puzzle scenario above, after GAT inference, the matching probability of the target block P_023 with each missing position is output. The matching probability of the missing position (3,6) is 0.85 (this position is expected to be a bird's wing, which highly matches the target block semantically, in color, and in terms of edge), the probability of the missing position (4,7) is 0.72 (it belongs to the sky area but the edge matching degree is slightly lower), the probability of the missing position (2,5) is 0.31 (the color matches but the semantics are cloud edges), and the probabilities of the remaining positions are all below 0.2. The top-k regions with the highest probabilities, such as the top 5, are selected as candidate regions for subsequent suggestion strategy decisions. Through GAT's attention inference, not only is the direct feature matching between the target block and the missing position considered, but the contextual constraints provided by the neighboring placed blocks are also fully utilized, making the convergence of candidate regions more accurate.

[0030] S205. Based on the user's current cognitive level, the number of times they have requested help, and the convergence degree of the candidate region, select the corresponding target prompt level from multiple preset prompt levels.

[0031] In this embodiment, the prompt output does not directly give the correct position of the target puzzle piece in the traditional way. Instead, it is selected based on a comprehensive consideration of the user's current cognitive level, the number of times the user has asked for help, and the convergence degree of the candidate regions. The user's current cognitive level can be set and dynamically adjusted according to historical performance. The convergence degree of the candidate regions can be the ratio of the probability of the Top-1 candidate to the number of candidate regions. For example, if the probability of the Top-1 candidate is 0.85 and the number of candidates is 5, the convergence degree is 0.17. Of course, other factors such as the number of candidate regions, historical success rate, and current puzzle completion degree can also be combined.

[0032] Based on various current factors and characteristics, a prompting strategy is selected, choosing the corresponding target prompting level from multiple preset prompting levels. Different prompting levels have different granularities, including global region prompts, local semantic prompts, neighbor block prompts, shape matching prompts, and direct location prompts. Specifically, when selecting a prompting level, a preset selection strategy can be followed. For example, if the cognitive level is high, it's the first time seeking help, and the candidate region has converged, a low-level prompting is started; if the cognitive level is low, the need for help is high, or multiple consecutive requests for help have occurred, a mid-level or high-level prompting can be started directly; if the number of candidate regions is greater than a threshold K1, a region or semantic prompt is executed first; if the number of candidate regions is less than a threshold K2 and the user has failed twice consecutively, a neighbor block prompt, shape prompt, or direct location prompt can be executed, etc. Of course, in other embodiments, other methods, such as policy networks, can also be used to determine the current target prompting level, thereby providing personalized prompting granularity for users with different cognitive levels and different stages of seeking help.

[0033] S206. Output the position prompt information of the corresponding target puzzle piece according to the target prompt hierarchy.

[0034] In this embodiment, after confirming the target prompt level, the position prompt information of the target puzzle piece corresponding to the current prompt level is output. That is, differentiated content rules, output granularity and presentation method are defined in advance for each prompt level to ensure that the prompt information not only meets the current user's cognitive load, but also effectively guides the user to move closer to the correct position.

[0035] For example, the location prompts output under global area prompts describe only a large area, such as "near the upper right corner," which is suitable for first-time requests and scenarios with a high level of cognitive ability. The location prompts output under local semantic prompts describe the target block as being in the background or among objects, such as "at the junction of the blue sky and treetops," which is suitable for first or second requests and scenarios where the candidate area is still large. The location prompts output under neighboring block prompts describe the relationship between the block and the adjacent block or object it should connect to, such as "it should be next to the left of that deer," which is suitable for scenarios where neighboring blocks have already been identified and the user has tried once without success. The location prompts output under shape matching prompts describe the matching method based on flat edges, concave / convex directions, and contours, which is suitable for scenarios where the user has failed multiple times and the system has identified a high-confidence candidate area. The location prompts output under direct positioning prompts directly provide an accurate grid, a highlighted location, or a unique candidate location, which is suitable for young users, users with high assistance needs, parents setting an efficiency-first mode, or scenarios where requests have failed multiple times consecutively. Specific output methods could include voice broadcasting from the reading pen, screen highlighting on a mobile terminal APP, or flashing guidance on the base plate, etc., which are not limited in this embodiment. By adapting the prompt level to the user's current understanding and puzzle-solving situation, and outputting corresponding location prompts, the system not only helps users locate the puzzle but also better cultivates their independent reasoning ability. This improves the flexibility and adaptability of puzzle location prompts and enhances the educational and guiding effect.

[0036] For example, in the "forest animals" puzzle scenario described above, an L1 target prompt level, i.e., a global area prompt, is determined based on the user's current cognition and puzzle-solving progress. This is then output via voice prompts using a reading pen: "This puzzle piece is located in the upper right part of the entire picture. You can try to find the sky first." Simultaneously, the upper half of the puzzle area is highlighted in a semi-transparent blue on the thumbnail of the full picture on the mobile app interface, helping children develop global spatial awareness. This prompt only describes a large area without mentioning specific objects or coordinates, thus preserving the space for children's independent exploration.

[0037] In the above embodiments, this invention discloses a hierarchical prompting method for jigsaw puzzle location assistance. In response to a user's request for help with the location of a target jigsaw puzzle piece, the method obtains the puzzle piece features; acquires the overall structure information of the entire puzzle and identifies the distribution information of currently completed areas; constructs a corresponding puzzle state heterogeneous graph based on the puzzle piece features and the distribution information of currently completed areas; performs message passing and attention reasoning on the puzzle state heterogeneous graph using a graph attention network to obtain the matching attention score between the target puzzle piece and each missing position node, and determines the candidate region of the target puzzle piece based on the matching attention score; selects the corresponding target prompting level from multiple preset prompting levels based on the user's current cognitive level, the number of times they have requested help, and the convergence degree of the candidate regions; and outputs the corresponding location prompting information of the target puzzle piece based on the target prompting level. This invention utilizes a graph attention network to achieve intelligent reasoning of candidate regions and adaptively selects the prompting level based on the user's state, thereby helping users gradually locate the puzzle piece through multi-level progressive prompts, improving the flexibility and adaptability of puzzle location prompts, and enhancing the educational guidance effect.

[0038] In one embodiment, step S201 includes: Receive a user's request for help regarding the location of a target puzzle piece, and extract the block identifier of the target puzzle piece from the location request; Based on the block identifier, the puzzle block features of the target puzzle block are extracted from a pre-established feature data table. The puzzle block features include image semantic features, color and texture features, contour interface features, and orientation features.

[0039] In this embodiment, upon receiving a user's request for assistance regarding the location of a target puzzle piece, the block identifier of the target puzzle piece is obtained from the request to confirm its unique identity. Furthermore, feature extraction is performed on the target puzzle piece. The specific feature extraction method can be one or a combination of pre-annotation, image recognition, and edge matching. The pre-annotation method involves establishing a feature data table for each puzzle piece during the puzzle content creation stage. This table records the block identifier, overall image coordinates, associated semantic object, main color distribution, number of straight edges, concave / convex interface encoding, and standard orientation information. Based on the block identifier, the puzzle piece features of the target puzzle piece can be directly extracted from the pre-established feature data table, including image semantic features, color and texture features, contour interface features, and orientation features, achieving fast and stable puzzle feature acquisition.

[0040] If image recognition is used, the target block image can be captured by the camera built into the mobile terminal or reading pen. The color histogram, texture features, semantic labels and orientation features are extracted using a pre-trained convolutional neural network (such as ResNet-18), and then compared with a pre-stored template for confirmation. This method is suitable for scenarios where traditional paper puzzles or labels are damaged.

[0041] If edge matching is used, the target block contour is sampled, and boundary curves, convex / groove sequences, flat edge information and edge direction information are extracted. Then, it is matched with the standard block contour library, so that robust recognition can be achieved even when the image angle is deflected or partially occluded.

[0042] This embodiment uses a multi-source feature extraction combination mechanism to ensure stable acquisition of multi-dimensional features of the target puzzle piece under various hardware conditions and environmental interference.

[0043] In one embodiment, step S202 includes: The target puzzle piece is used to identify the entire puzzle to which it belongs, and the overall structure information of the entire puzzle is obtained. The entire jigsaw puzzle is divided into M×N position units based on the overall image structure information, and the occupancy status of each position unit is maintained. If the location unit has detected a valid puzzle piece and it matches the expected block identifier or the set of compatible blocks, it is identified as a completed area; If the location unit does not detect a puzzle piece, or the detected puzzle piece does not meet the orientation and / or adjacency matching conditions, it is identified as a missing area; If a positional unit in the completed area is located around the correct position of the target puzzle piece, it is identified as a neighboring puzzle piece.

[0044] In this embodiment, when identifying the distribution information of the currently completed area, the entire puzzle to which the target puzzle piece belongs is first confirmed, and the overall structure information of the entire puzzle is called. That is, the overall structure information is generated during the puzzle making stage, including the size parameters of the entire puzzle, the M×N grid division scheme, the theoretical block identifier of each position unit, the regional semantic label, and the object relationship diagram (such as "the deer is located in the center of the grass, adjacent to the flower on the left").

[0045] Then, based on the overall image structure information, the entire puzzle is divided into M×N positional units, and an M×N dimensional occupancy state matrix is ​​maintained, where each element... Oi , j ∈{0,1, 1}, Oi , j =1 indicates that the correct tile has been placed in the cell at that location. Oi , j=0 indicates that the cell at that location is empty. Oi , j =-1 indicates that the cell at this location has an incorrect tile (the tile identifier does not match or the orientation / adjacency condition is not met).

[0046] If a certain location unit has detected a valid puzzle piece and it matches the expected block identifier or the set of compatible blocks, it is identified as a completed area. The identification of a completed area not only checks whether the block identifier is consistent, but also checks whether the actual orientation of the puzzle piece is consistent with the standard orientation (error less than 15°), and whether the edge interface of the puzzle piece and the identified neighboring block is physically compatible (convex-concave matching). If a certain location cell does not detect a puzzle piece, or the detected puzzle piece does not meet the orientation and / or adjacency matching conditions, it is identified as a missing area or an area to be corrected. If a positional unit in the completed area is near the correct position of the target puzzle piece, it is identified as a neighboring puzzle piece. That is, for the identification of neighboring puzzle pieces, the system uses the theoretical position of the target puzzle piece (…). x , y Centered on ), examine its eight neighboring (up, down, left, right, and four diagonal) positional cells. Oi , j For puzzle pieces with a value of 1, mark them as having been placed next to other puzzle pieces.

[0047] This embodiment maintains the occupancy status of independent regions and determines the multi-condition matching, enabling accurate differentiation between three states: correct placement, vacancy, and incorrect placement, thus providing a reliable basis for node classification in subsequent heterogeneous graph construction.

[0048] In one embodiment, step S203 includes: Based on the current completed area distribution information, obtain the completed areas and the vacant areas; Each placed puzzle piece in the completed area, each empty position in the empty area, and the target puzzle piece are taken as graph nodes. The node features of each graph node include semantic embedding vector, color embedding vector, edge encoding vector, and orientation encoding vector. Heterogeneous edge relationships are constructed based on the spatial adjacency, semantic continuity, and color compatibility relationships between the graph nodes to obtain the corresponding heterogeneous graph of the puzzle state.

[0049] In this embodiment, the construction of the heterogeneous graph of the puzzle state is a process of transforming the physical puzzle state into structured data that can be processed by a graph neural network. Since heterogeneous graphs allow for multiple types of nodes and edges, they can more finely depict the complex relationships between entities in the real world. Therefore, in this embodiment, completed regions and empty regions are obtained based on the distribution information of the currently completed regions. Each placed puzzle piece in the completed region, each empty position in the empty region, and the target puzzle piece are used as graph nodes. The node features of each graph node include semantic embedding vectors, color embedding vectors, edge encoding vectors, and orientation encoding vectors. Among them, the placed puzzle piece nodes carry rich physical features, the empty position nodes carry expected features, and the target puzzle piece nodes serve as query anchors. The three types of nodes are interconnected through multi-dimensional heterogeneous edges, enabling the subsequent GAT to perform reasoning in a relationship-aware environment.

[0050] When constructing heterogeneous edge relationships, corresponding edges are established based on the spatial adjacency, semantic continuity, and color compatibility relationships between graph nodes. Spatial adjacency edges are established based on the physical adjacency of positional units. For example, if the positional units corresponding to two nodes share an edge (up, down, left, right) in the grid, a strong spatial adjacency edge is established with a weight of 1.0; if they only share a vertex (diagonally), a weak spatial adjacency edge is established with a weight of 0.5. Semantic continuity edges are established based on hierarchical matching of semantic labels. If the semantic labels of two nodes share the same parent concept (e.g., "bird / wings" and "bird / body" share "bird"), a semantic continuity edge is established with a weight of 0.8; if they belong to the same background region (e.g., "sky / clouds" and "sky / sun"), the weight is 0.6, etc. Color compatibility edges can be established based on the Euclidean distance in the Lab color space. If the Lab distance between the primary colors of two nodes is less than a preset threshold, such as 30, a color compatible edge is established, with higher weights for colors that are more similar.

[0051] This embodiment models the puzzle state as a heterogeneous graph, thereby transforming the puzzle location problem into a similarity reasoning problem between graph nodes. By constructing graph nodes and various types of edges between them, the multi-dimensional topological relationships of the current puzzle state are fully represented, providing a reliable basis for subsequent candidate region reasoning and determination.

[0052] In one embodiment, step S204 includes: The heterogeneous graph of the puzzle state is processed by multi-head attention mechanism of graph attention network to perform multi-layer message passing, and the features of the neighborhood puzzle piece nodes are aggregated for each empty position node, and the attention weight between each empty position node and the target puzzle piece node is calculated. The matching attention score is calculated based on the attention weight, and the probability value of each missing position as a candidate region of the target puzzle piece is output. Sort the probability values ​​from largest to smallest, and select a preset number of empty positions at the top of the sort as candidate regions for the target puzzle piece.

[0053] In this embodiment, adaptive message passing on a heterogeneous graph is performed using a graph attention network. Specifically, a multi-head attention mechanism is used, allowing each vacant node to dynamically calculate the contribution weights of different neighboring nodes based on the features of its neighboring nodes with placed puzzle pieces, thereby aggregating a high-order feature representation containing contextual semantics. Specifically, a two-layer GATv2 network can be used for inference, with 4 or 6 attention heads. In the first layer of message passing, each vacant node aggregates the features of its neighboring nodes with placed puzzle pieces through four independent attention heads. The formula for calculating the attention coefficient for each attention head is as follows:

[0054] in, l Indicates the number of network layers. h Indicates the attention head number. and They represent the first l Layer nodes i and nodes j eigenvectors, It is a learnable linear transformation matrix. For the attention weight vector, [ ∥ The ] symbol represents a vector concatenation operation, which, after softmax normalization, yields the attention weights. ,node i In the h The updated features under the size are:

[0055] The outputs from the four heads are concatenated and then linearly projected to obtain the updated node features for the first layer. The second layer further propagates based on this, ultimately calculating the target puzzle piece nodes. t With each vacant location node v Matching attention score:

[0056] in, and The projection matrix of the query and key is transformed into a probability distribution after softmax normalization. The Top-K empty positions with the highest probability are selected as candidate regions for the target puzzle piece.

[0057] This embodiment utilizes message propagation and reasoning through graph attention networks. Based on parallel computation of multi-head attention and cascaded aggregation of multi-layer message passing, it not only considers the direct feature matching between the target block and the missing position, but also makes full use of the context constraints provided by the neighboring placed blocks, realizing multi-factor joint reasoning for candidate regions, making candidate region convergence more accurate.

[0058] In one embodiment, step S205 includes: Based on the user's current cognitive level, the number of times they have requested help, and the convergence degree of the candidate region, a corresponding request state vector is constructed. The help request state vector is input into a policy network trained by reinforcement learning, and the policy network outputs the expected cumulative reward for each preset prompt level based on the help request state vector. Select the preset prompt level with the highest expected cumulative reward as the target prompt level.

[0059] In this embodiment, the selection of the target prompt level is no longer based on fixed rule matching judgment, but is modeled as a sequential decision problem and dynamically optimized through a policy network trained by reinforcement learning. The advantage of reinforcement learning is that it can consider long-term cumulative benefits rather than local greed. For example, if a lower-level prompt is selected at present, it may increase the probability of the user asking for help again, but in the long run, it helps to cultivate the user's independent reasoning ability. Conversely, directly giving a higher-level prompt can solve the problem immediately, but it may bring "learning loss". The selection of adaptive prompt strategy is realized through a policy network trained by reinforcement learning, which enables the handling of complex situations that are difficult to cover by conventional rule matching selection.

[0060] Specifically, first, a 5-dimensional help request state vector, s, is constructed. t =[ c , r , γ , ρ , [The cognitive level] c The system dynamically evaluates users based on their historical performance. Initial values ​​can be set by parents in the app (levels 1-5, where 1 represents younger / lower cognitive development and 5 represents older / higher cognitive development). Each successful independent completion of a puzzle piece... c A slight increase of 0.1, for every two consecutive failures c Slightly decreased by 0.1, with an upper limit of 5.0 and a lower limit of 1.0, etc.; number of requests for help. r This is the count in the current continuous request sequence; it is reset to 0 if the user successfully splices the sequence within the observation window; candidate region convergence. γ The probability of Top-1 candidate divided by the total number of candidate regions K ,Right now γ=ptop1 / KThe larger this value, the more convergent the candidate region; historical success rate ρ The percentage of successful puzzle completions in the user's most recent N (e.g., 10) requests for assistance; the current puzzle completion rate. This is the ratio of the number of correctly placed blocks to the total number of blocks.

[0061] The constructed help-seeking state vector is input into a policy network trained using reinforcement learning. Specifically, this policy network can employ a three-layer multilayer perceptron structure, including an input layer, hidden layers, and an output layer. It is pre-trained in a jigsaw puzzle simulation environment using the Proximal Policy Optimization (PPO) algorithm, simulating placement behaviors of users with different cognitive levels and providing corresponding puzzle-solving rewards. For example, a user receives a +10 reward for successfully completing the puzzle after receiving a hint, with higher rewards for lower hint levels; a -2 reward for asking for help again within the observation window; a -5 reward for placing the puzzle incorrectly; a -20 reward for abandoning the puzzle; and a +50 reward for completing the entire puzzle, etc. These specific settings can be flexibly adjusted according to training needs, with the training objective being to maximize long-term cumulative rewards. After training, each time a user asks for help, the current state vector is input into the policy network. The network forward propagates and outputs the expected cumulative reward for each hint level, selecting the level with the highest expected cumulative reward as the target hint level.

[0062] In this embodiment, the long-term learning benefits of the user are weighed through a policy network during the selection of prompt level, so as to achieve adaptive prompt level selection. For example, although the highest prompt level L5 can solve the problem immediately (high immediate reward), it will sacrifice the user's subsequent opportunity for independent exploration (long-term reward loss), while the lowest prompt level L1 may cause the user to ask for help again (immediate risk), but in the long run, it helps to maintain the user's learning participation. Therefore, through global optimization of reinforcement learning, the rigidity of decision-making in non-boundary situations of fixed rules can be effectively avoided, and the optimal balance between teaching effect and task efficiency can be achieved.

[0063] In one embodiment, after step S206, the method further includes: Enter a preset duration observation window, and monitor whether the user has correctly assembled the target puzzle piece within the observation window; The actual reward is calculated based on the user's splicing results and help requests within the observation window, and the parameters of the policy network are updated based on the actual reward. If the splicing is not completed correctly within the observation window, the current help round and the convergence degree of the candidate region in the state vector are updated, and the updated state vector is re-input into the policy network. The policy network upgrades the target prompting level based on the updated state vector and outputs corresponding location prompting information according to the upgraded target prompting level.

[0064] In this embodiment, a closed-loop feedback mechanism for reinforcement learning is implemented through an observation window, which associates a single prompt output with the user's actual behavior, thus forming a complete closed loop of output, observation, reward, and update. Specifically, the preset observation window can be set to a specified duration, such as 30 seconds, or until a new puzzle piece placement behavior by the user is detected. During the window period, the pressure sensor array on the puzzle base and the NFC recognition module monitor in real time whether the target puzzle piece is placed in the correct position. If the user correctly completes the puzzle within the window period, a positive reward is obtained and used to update the policy network, making the network more inclined to output the same level of prompt when encountering similar situations in the future. If the user fails, a negative reward is obtained and the prompt level is upgraded. At the same time, the policy network is fine-tuned online using the interaction data to continuously optimize the policy.

[0065] Furthermore, if the user fails to complete the stitching correctly within the observation window, the current request round and the convergence degree of the candidate region in the state vector are updated, where the request round... r Increase by 1, candidate region convergence γ The system recalculates based on the current number of candidate regions, for example, by reducing the number of candidate regions to further improve the convergence of candidate regions. The updated state vector is then re-inputted into the policy network, which re-infers the expected cumulative reward for each level, thereby selecting a new target cue level. This new target cue level must be no lower than the previous level (upgrade rule) to ensure that the granularity of the cue increases monotonically and to avoid cyclical cueing.

[0066] This embodiment achieves continuous evolution of the prompting strategy through closed-loop feedback of the observation window and online updates of the strategy network. This allows the same user to obtain the most suitable prompting granularity at different learning stages and in different emotional states, balancing puzzle completion efficiency and long-term cognitive development effects.

[0067] It should be noted that there is no necessary order between the above steps. Those skilled in the art will understand from the description of the embodiments of the present invention that the above steps may have different execution orders in different embodiments, that is, they may be executed in parallel or in turn, etc.

[0068] Further reference Figure 3 As a response to the above Figure 2 The present invention provides an embodiment of a layered prompting device for puzzle position assistance, which, according to the method shown, provides a solution for puzzle position assistance. Figure 2 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0069] like Figure 3 As shown, the layered prompting device 30 for puzzle position assistance described in this embodiment includes: The feature extraction module 301 is used to obtain the puzzle piece features of the target puzzle piece in response to a user's request for help in finding the location of the target puzzle piece. The analysis module 302 is used to obtain the overall structure information of the entire jigsaw puzzle and identify the distribution information of the currently completed areas; The graph construction module 303 is used to construct a corresponding puzzle state heterogeneous graph based on the puzzle piece characteristics of the target puzzle piece and the distribution information of the currently completed area; The candidate region determination module 304 is used to perform message passing and attention reasoning on the heterogeneous graph of the puzzle state through a graph attention network, obtain the matching attention score between the target puzzle piece and each empty position node, and determine the candidate region of the target puzzle piece based on the matching attention score; The prompt level determination module 305 is used to select the corresponding target prompt level from multiple preset prompt levels based on the user's current cognitive level, the number of times the help request is received, and the convergence degree of the candidate region. The hierarchical prompting module 306 is used to output the position prompting information of the corresponding target puzzle piece according to the target prompting level.

[0070] The module referred to in this invention is a series of computer program instruction segments that can perform specific functions. It is more suitable than a program for describing the hierarchical prompt execution process of puzzle position assistance. For the specific implementation of each module, please refer to the corresponding method embodiments above, which will not be repeated here.

[0071] In one embodiment, the feature extraction module 301 includes: The receiving unit is used to receive a user's request for help regarding the location of a target puzzle piece, and to extract the block identifier of the target puzzle piece from the request for help regarding its location; The feature extraction unit is used to extract the puzzle block features of the target puzzle block from a pre-established feature data table based on the block identifier. The puzzle block features include image semantic features, color and texture features, contour interface features, and orientation features.

[0072] In one embodiment, the analysis completion module 302 includes: The whole image acquisition unit is used to identify the whole puzzle to which the target puzzle piece belongs based on the target puzzle piece, and to acquire the whole image structure information of the whole puzzle piece; The division and maintenance unit is used to divide the entire puzzle into M×N position units according to the overall image structure information, and maintain the occupancy status of each position unit; The identification unit is configured to identify a completed region if the location unit has detected a valid puzzle piece and it matches the expected block identifier or a set of compatible blocks; identify a missing region if the location unit has not detected a puzzle piece, or if the detected puzzle piece does not meet the orientation and / or adjacency matching conditions; and identify a neighboring puzzle piece if the location unit in the completed region is around the correct position of the target puzzle piece.

[0073] In one embodiment, the graph construction module 303 includes: The distribution acquisition unit is used to acquire completed areas and vacant areas based on the current completed area distribution information; The node determination unit is used to take each placed puzzle piece in the completed area, each empty position in the empty area, and the target puzzle piece as graph nodes. The node features of each graph node include semantic embedding vector, color embedding vector, edge encoding vector, and direction encoding vector. Edge building units are used to construct heterogeneous edge relationships based on the spatial adjacency, semantic continuity, and color compatibility relationships between graph nodes, thereby obtaining the corresponding heterogeneous graph of the jigsaw puzzle state.

[0074] In one embodiment, the candidate region determination module 304 includes: The message passing unit is used to perform multi-layer message passing on the heterogeneous graph of the puzzle state through the multi-head attention mechanism of the graph attention network, aggregate the features of the neighborhood puzzle piece nodes that have been placed for each empty position node, and calculate the attention weight between each empty position node and the target puzzle piece node. The probability calculation unit is used to calculate the matching attention score based on the attention weight and output the probability value of each missing position as a candidate region of the target puzzle piece. The candidate region determination unit is used to sort the probability values ​​from largest to smallest and select a preset number of empty positions at the top of the sort as candidate regions for the target puzzle piece.

[0075] In one embodiment, the prompt level determination module 305 includes: The state construction unit is used to construct a corresponding help-seeking state vector based on the user's current cognitive level, the number of help-seeking rounds, and the convergence degree of the candidate region. The input unit is used to input the help request state vector into the policy network trained by reinforcement learning, and the policy network outputs the expected cumulative reward for each preset prompt level based on the help request state vector. The level selection unit is used to select the preset prompt level with the highest expected cumulative reward as the target prompt level.

[0076] In one embodiment, the device 30 further includes: An observation unit is used to enter an observation window of a preset duration and monitor whether the user has correctly assembled the target puzzle piece within the observation window. The parameter update unit is used to calculate the actual reward based on the user's splicing results and help-seeking behavior in the observation window, and update the parameters of the policy network based on the actual reward. The state update unit is used to update the current help round and candidate region convergence degree in the state vector if the splicing is not completed correctly within the observation window, and re-input the updated state vector into the policy network. The prompt upgrade unit is used to upgrade the target prompt level by the policy network based on the updated state vector, and output the corresponding location prompt information according to the upgraded target prompt level.

[0077] In the above embodiments, this invention discloses a hierarchical prompting device for jigsaw puzzle location assistance. Responding to a user's request for help with the location of a target jigsaw puzzle piece, the device obtains the puzzle piece features; acquires the overall structure information of the entire puzzle and identifies the distribution information of currently completed areas; constructs a corresponding puzzle state heterogeneous graph based on the puzzle piece features and the distribution information of currently completed areas; performs message passing and attention reasoning on the puzzle state heterogeneous graph using a graph attention network to obtain matching attention scores between the target puzzle piece and each missing position node, and determines candidate regions for the target puzzle piece based on the matching attention scores; selects a corresponding target prompting level from multiple preset prompting levels based on the user's current cognitive level, the number of times they have requested help, and the convergence degree of the candidate regions; and outputs the corresponding location prompting information for the target puzzle piece based on the target prompting level. This invention utilizes a graph attention network to achieve intelligent reasoning of candidate regions and adaptively selects prompting levels based on the user's state, thereby helping users gradually locate the puzzle piece through multi-level progressive prompts, improving the flexibility and adaptability of puzzle location prompts, and enhancing the educational guidance effect.

[0078] Specific limitations regarding the layered prompting device for puzzle position assistance can be found in the limitations of the layered prompting method for puzzle position assistance described above, and will not be repeated here. Each module in the aforementioned layered prompting device for puzzle position assistance can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0079] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0080] Another embodiment of the present invention provides a computer device, such as... Figure 4 As shown, the computer device 40 includes: One or more processors 401 and memory 402, Figure 4 The following section uses a processor 401 as an example. The processor 401 and the memory 402 can be connected via a bus or other means. Figure 4 Taking the example of a connection between China and Israel via a bus.

[0081] The processor 401 is used to perform various control logics of the computer device 40. It can be any conventional processor, microprocessor, state machine, general-purpose processor, digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), microcontroller, ARM (Acorn RISC Machine) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination of these components.

[0082] The memory 402, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions corresponding to the layered prompting method for puzzle position assistance in the embodiments of the present invention. The processor 401 executes various functional applications and data processing of the computer device 40 by running the non-volatile software programs, instructions, and units stored in the memory 402, thereby implementing the layered prompting method for puzzle position assistance in the above method embodiments.

[0083] Another embodiment of the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by one or more processors, perform the steps of the layered hint method for puzzle position assistance in any of the above method embodiments.

[0084] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0085] Based on the above description of the embodiments, those skilled in the art will understand that the methods described in the embodiments can be implemented using software plus necessary general-purpose hardware platforms. Of course, they can also be implemented using hardware, but in many cases, the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0086] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The computer program can be stored in a non-volatile, computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The storage medium can be a memory, magnetic disk, floppy disk, flash memory, optical storage, etc.

[0087] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0088] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A layered prompting method for jigsaw puzzle position assistance, characterized by, include: In response to a user's request for help regarding the location of a target puzzle piece, the puzzle piece features of the target puzzle piece are obtained; Obtain the overall structure information of the entire jigsaw puzzle and identify the distribution information of the currently completed areas; Based on the puzzle piece characteristics of the target puzzle piece and the distribution information of the currently completed area, a corresponding puzzle state heterogeneous graph is constructed; The heterogeneous graph of the puzzle state is processed by a graph attention network to perform message passing and attention reasoning, thereby obtaining the matching attention score between the target puzzle piece and each missing position node, and the candidate region of the target puzzle piece is determined based on the matching attention score. Based on the user's current cognitive level, the number of times they have sought help, and the convergence degree of the candidate region, the corresponding target prompt level is selected from multiple preset prompt levels. Output the position prompt information of the corresponding target puzzle piece according to the target prompt hierarchy; The step of selecting a target prompt level from multiple preset prompt levels based on the user's current cognitive level, the number of times they have sought help, and the convergence degree of the candidate region includes: Based on the user's current cognitive level, the number of times they have requested help, and the convergence degree of the candidate region, a corresponding request state vector is constructed. The help request state vector is input into a policy network trained by reinforcement learning, and the policy network outputs the expected cumulative reward for each preset prompt level based on the help request state vector. Select the preset prompt level with the highest expected cumulative reward as the target prompt level.

2. The tiered hinting method of seeking assistance for a jigsaw puzzle position of claim 1, wherein, The step of responding to a user's request for help regarding the location of a target puzzle piece, and obtaining the puzzle piece features of the target puzzle piece, includes: Receive a user's request for help regarding the location of a target puzzle piece, and extract the block identifier of the target puzzle piece from the location request; Based on the block identifier, the puzzle block features of the target puzzle block are extracted from a pre-established feature data table. The puzzle block features include image semantic features, color and texture features, contour interface features, and orientation features.

3. The tiered hinting method of seeking assistance for a puzzle position of claim 1, wherein, The process of obtaining the overall structure information of the entire jigsaw puzzle and identifying the distribution information of the currently completed areas includes: The target puzzle piece is used to identify the entire puzzle to which it belongs, and the overall structure information of the entire puzzle is obtained. The entire jigsaw puzzle is divided into M×N position units based on the overall image structure information, and the occupancy status of each position unit is maintained. If the location unit has detected a valid puzzle piece and it matches the expected block identifier or the set of compatible blocks, it is identified as a completed area; If the location unit does not detect a puzzle piece, or the detected puzzle piece does not meet the orientation and / or adjacency matching conditions, it is identified as a missing area; If a positional unit in the completed area is located around the correct position of the target puzzle piece, it is identified as a neighboring puzzle piece.

4. The layered prompting method for puzzle position assistance according to claim 1, characterized in that, The step of constructing a corresponding puzzle state heterogeneous graph based on the puzzle piece features of the target puzzle piece and the distribution information of the currently completed areas includes: Based on the current completed area distribution information, obtain the completed areas and the vacant areas; Each placed puzzle piece in the completed area, each empty position in the empty area, and the target puzzle piece are taken as graph nodes. The node features of each graph node include semantic embedding vector, color embedding vector, edge encoding vector, and orientation encoding vector. Heterogeneous edge relationships are constructed based on the spatial adjacency, semantic continuity, and color compatibility relationships between the graph nodes to obtain the corresponding heterogeneous graph of the puzzle state.

5. The layered prompting method for puzzle position assistance according to claim 4, characterized in that, The step of performing message passing and attention reasoning on the heterogeneous graph of the puzzle state using a graph attention network to obtain the matching attention score between the target puzzle piece and each missing position node, and determining the candidate region of the target puzzle piece based on the matching attention score, includes: The heterogeneous graph of the puzzle state is processed by multi-head attention mechanism of graph attention network to perform multi-layer message passing, and the features of the neighborhood puzzle piece nodes are aggregated for each empty position node, and the attention weight between each empty position node and the target puzzle piece node is calculated. The matching attention score is calculated based on the attention weight, and the probability value of each missing position as a candidate region of the target puzzle piece is output. Sort the probability values ​​from largest to smallest, and select a preset number of empty positions at the top of the sort as candidate regions for the target puzzle piece.

6. The layered prompting method for puzzle position assistance according to claim 1, characterized in that, After outputting the position hint information of the corresponding target puzzle piece according to the target hint hierarchy, the method further includes: Enter a preset duration observation window, and monitor whether the user has correctly assembled the target puzzle piece within the observation window; The actual reward is calculated based on the user's splicing results and help requests within the observation window, and the parameters of the policy network are updated based on the actual reward. If the splicing is not completed correctly within the observation window, the current help round and the convergence degree of the candidate region in the state vector are updated, and the updated state vector is re-input into the policy network. The policy network upgrades the target prompting level based on the updated state vector and outputs corresponding location prompting information according to the upgraded target prompting level.

7. A layered prompting device for assisting with puzzle position, characterized in that, include: The feature extraction module is used to obtain the puzzle piece features of the target puzzle piece in response to a user's request for help in finding the location of the target puzzle piece. The analysis module is used to obtain the overall structure information of the entire jigsaw puzzle and identify the distribution information of the currently completed areas; The graph construction module is used to construct a corresponding heterogeneous graph of the puzzle state based on the puzzle piece characteristics of the target puzzle piece and the distribution information of the currently completed area; The candidate region determination module is used to perform message passing and attention reasoning on the heterogeneous graph of the puzzle state through a graph attention network, obtain the matching attention score between the target puzzle piece and each empty position node, and determine the candidate region of the target puzzle piece based on the matching attention score; The prompt level determination module is used to select the corresponding target prompt level from multiple preset prompt levels based on the user's current cognitive level, the number of times they have sought help, and the convergence degree of the candidate region. The hierarchical prompting module is used to output the position prompting information of the corresponding target puzzle piece according to the target prompting hierarchy; The prompt level determination module includes: The state construction unit is used to construct a corresponding help-seeking state vector based on the user's current cognitive level, the number of help-seeking rounds, and the convergence degree of the candidate region. The input unit is used to input the help request state vector into the policy network trained by reinforcement learning, and the policy network outputs the expected cumulative reward for each preset prompt level based on the help request state vector. The level selection unit is used to select the preset prompt level with the highest expected cumulative reward as the target prompt level.

8. A computer device, characterized in that, Includes at least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the layered hint method for puzzle location assistance as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions that, when executed by one or more processors, cause the one or more processors to perform the layered hint method for puzzle position assistance as described in any one of claims 1-6.

Images