Asset evaluation method and device and computer program product

Abstract

The invention discloses an asset evaluation method, an asset evaluation device and a computer program product. The method relates to the field of artificial intelligence, and comprises the following steps: receiving a point cloud data set of a target collateral, and extracting N groups of point cloud clusters from the point cloud data set; for each group of point cloud clusters, determining point cloud centers of the point cloud clusters, extracting local features of the point cloud clusters, determining point cloud agents of the point cloud clusters based on the point cloud centers and the local features, and obtaining N point cloud agents; inputting the N point cloud agents into an encoder of the target model to obtain an encoding vector of each point cloud agent, and inputting the N encoding vectors into a decoder of the target model to obtain missing point agents; and complementing the missing point cloud in the point cloud data set based on the missing point agent to obtain a target point cloud data set, and performing asset assessment on the target guarantee through the target point cloud data set to obtain an asset assessment result. Through the method and the device, the problem of inaccurate asset assessment caused by missing of the point cloud data in related technologies is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, in particular to an asset evaluation method and device and computer program product. BACKGROUND

[0002] The point cloud completion method in the related art is to realize the completion of the point cloud by transformation and filling of the grid form data, but limitations are often encountered when processing complex geometric structures or large-scale data sets, especially when processing point cloud data with high irregularity and local details, the completion effect may not be satisfactory.

[0003] The learning-based point cloud completion technology is a new way to solve the point cloud completion problem. The learning-based point cloud completion method mainly relies on machine learning and deep learning models, and restores the missing part from the incomplete point cloud by training the model. However, the learning-based point cloud completion method has the problem that some details that are crucial to the reconstruction of geometric shapes may be simplified when the input point cloud data is encoded into low-level feature representations, only retaining features that can summarize global information. In the decoding stage, even if the prediction of points can be restored from the global features, the distribution of these predicted points often lacks structure, resulting in significant differences in geometric shapes between the generated point cloud and the real point cloud.

[0004] In view of the problem that the asset evaluation is inaccurate due to the missing of point cloud data in the related art, no effective solution has been proposed so far. SUMMARY

[0005] The main purpose of the present application is to provide an asset evaluation method, device and computer program product to solve the problem that the asset evaluation is inaccurate due to the missing of point cloud data in the related art.

[0006] In order to achieve the above-mentioned purpose, according to one aspect of the present application, an asset evaluation method is provided. The method comprises: receiving a point cloud data set of a target collateral, extracting N groups of point cloud clusters from the point cloud data set, wherein N is a positive integer; for each group of point cloud clusters, determining a point cloud center of the point cloud cluster, extracting a local feature of the point cloud cluster, determining a point cloud agent of the point cloud cluster based on the point cloud center and the local feature, and obtaining N point cloud agents; inputting the N point cloud agents into an encoder of a target model to obtain an encoding vector of each point cloud agent, inputting the N encoding vectors into a decoder of the target model to obtain a missing point agent; completing the missing point cloud in the point cloud data set based on the missing point agent to obtain a target point cloud data set, and performing asset evaluation on the target collateral through the target point cloud data set to obtain an asset evaluation result.

[0007] Optionally, the inputting the N point cloud proxies into an encoder of the target model to obtain an encoding vector of each point cloud proxy comprises: converting each point cloud proxy into a feature vector, performing linear projection and max-pooling operation on each feature vector to obtain a global feature of each point cloud proxy; performing linear transformation on the global feature of each point cloud proxy to obtain a coordinate of the point cloud proxy; determining a concatenated representation of the global feature and the coordinate of each point cloud proxy, and inputting each concatenated representation into a multi-layer perception to obtain the encoding vector of the point cloud proxy.

[0008] Optionally, the converting each point cloud proxy into a feature vector comprises: performing feature extraction on the point cloud proxy through a multi-level feature learning block of the encoder to obtain M feature maps of different resolution scales, wherein the multi-level feature learning block comprises M branch networks, each branch network is configured to extract a feature map of a resolution scale from the point cloud proxy, and M is a positive integer; capturing a topological feature of the point cloud proxy through a topological perception block of the encoder, wherein the topological feature is configured to represent a feature of a local geometric structure of the point cloud proxy; extracting features in the M feature maps of different resolution scales through a hierarchical feature fusion block of the encoder to obtain M features of different resolution scales, performing feature fusion on the M features of different resolution scales to obtain a fusion feature; and integrating the fusion feature and the topological feature to obtain the feature vector of the point cloud proxy, wherein the integration processing comprises at least one of the following: feature splicing, weighted summation, and network layer mapping.

[0009] Optionally, the capturing the topological feature of the point cloud proxy through the topological perception block of the encoder comprises: determining a query coordinate and a key-value coordinate of the point cloud proxy, wherein the query coordinate is configured to focus on a local region of the point cloud proxy, and the key-value coordinate is configured to represent context information of the local region; retrieving K points near the query coordinate through a preset machine learning model and the key-value coordinate, wherein K is a positive integer, and the K points are determined as neighbor points; extracting features of the K neighbor points to obtain neighbor features, and screening the topological feature from the neighbor features through a max-pooling operation.

[0010] Optionally, the M different resolution scales comprise a first resolution, a second resolution, and a third resolution, and the feature fusion on the M features of different resolution scales to obtain the fusion feature comprises: calculating a similarity between the first resolution feature and the second resolution feature to obtain a first attention weight, performing weighted processing on the first resolution feature based on the first attention weight to obtain a first feature, wherein the first resolution is greater than the second resolution; calculating a similarity between the third resolution feature and the second resolution feature to obtain a second attention weight, performing weighted processing on the third resolution feature based on the second attention weight to obtain a second feature, wherein the second resolution is greater than the third resolution; and performing weighted summation on the first feature, the second feature, and the second resolution feature to obtain the fusion feature.

[0011] Optionally, the obtaining the target point cloud data set based on the missing point agent to complete the missing point cloud in the point cloud data set comprises: extracting a shape feature from the missing point agent, determining an activation state of the missing point agent; generating a missing point cloud feature based on the activation state and the shape feature, determining a shape code of the missing point agent through a point inflation strategy and the missing point cloud feature; concatenating the shape code and the missing point cloud feature to generate a rough point cloud, integrating the rough point cloud with the point cloud data set to obtain an updated point cloud data set; generating a point cloud seed through a far point sampling algorithm and the updated point cloud data set, and refining the point cloud seed through an upsampling operation to obtain the target point cloud data set.

[0012] Optionally, the refining the point cloud seed through the upsampling operation to obtain the target point cloud data set comprises: extracting a feature of the point cloud seed through a local spatial attention mechanism and a global channel attention mechanism to obtain a target feature; inputting the target feature into an upsampling block to complete the point cloud from the point cloud seed through a step-by-step upsampling operation to obtain an initial completed point cloud; and adjusting the initial completed point cloud through a contrastive learning module to obtain the target point cloud data set.

[0013] Optionally, the adjusting the initial completed point cloud through the contrastive learning module to obtain the target point cloud data set comprises: determining a first point cloud feature extracted from the N point cloud agents by the encoder, and determining a second point cloud feature extracted from the initial completed point cloud; comparing the first point cloud feature and the second point cloud feature to obtain a contrastive feature, and adjusting the initial completed point cloud based on the contrastive feature to obtain the target point cloud data set.

[0014] To achieve the above object, according to another aspect of the present application, an asset evaluation device is provided. The device comprises: a receiving unit configured to receive a point cloud data set of a target collateral, and extract N groups of point cloud clusters from the point cloud data set, wherein N is a positive integer; a first determining unit configured to determine, for each group of point cloud clusters, a point cloud center of the point cloud cluster, extract a local feature of the point cloud cluster, and determine a point cloud agent of the point cloud cluster based on the point cloud center and the local feature, to obtain N point cloud agents; an input unit configured to input the N point cloud agents into an encoder of a target model to obtain an encoding vector of each point cloud agent, and input the N encoding vectors into a decoder of the target model to obtain a missing point agent; and a completion unit configured to complete a missing point cloud in the point cloud data set based on the missing point agent to obtain a target point cloud data set, and perform asset evaluation on the target collateral through the target point cloud data set to obtain an asset evaluation result.

[0015] In the embodiment of the present application, the point cloud data set of the target collateral is received, N groups of point cloud clusters are extracted from the point cloud data set, wherein N is a positive integer; for each group of point cloud clusters, the point cloud center of the point cloud cluster is determined, the local feature of the point cloud cluster is extracted, the point cloud agent of the point cloud cluster is determined based on the point cloud center and the local feature, and N point cloud agents are obtained; the N point cloud agents are input into the encoder of the target model to obtain the encoding vector of each point cloud agent, and the N encoding vectors are input into the decoder of the target model to obtain the missing point agent; the missing point cloud in the point cloud data set is completed based on the missing point agent, the target point cloud data set is obtained, the asset evaluation of the target collateral is performed through the target point cloud data set, and the asset evaluation result is obtained. The method of extracting point cloud clusters from point cloud data sets and determining point cloud centers and local features can accurately reconstruct missing point clouds, greatly improving the integrity and quality of point cloud data. The point cloud agent determined based on the point cloud center and the local feature captures and restores the detailed features of the point cloud, making the completed point cloud more realistic and detailed. Using the complete and fine-grained target point cloud data set for asset evaluation can more accurately measure and analyze the key parameters such as size, shape and structure of the collateral, thereby obtaining more accurate asset evaluation results. The purpose of improving the quality and integrity of point cloud data is achieved, thereby realizing the technical effect of improving the accuracy of asset evaluation, and further solving the technical problem of inaccurate asset evaluation due to missing point cloud data. BRIEF DESCRIPTION OF DRAWINGS

[0016] The accompanying drawings, which form a part of this application, are intended to provide further understanding of the application and are incorporated herein in their entirety. The schematic embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0017] Figure 1 A hardware structure block diagram of a computer terminal (or mobile device) for implementing an asset evaluation method is shown;

[0018] Figure 2 is a structural schematic diagram of a point cloud completion system provided according to an embodiment of the present application;

[0019] Figure 3 is a flowchart of an asset evaluation method provided according to an embodiment of the present application;

[0020] Figure 4 is a flowchart of a point agent generation module generating a point cloud agent according to an embodiment of the present application;

[0021] Figure 5 is a flowchart of a self-encoder architecture generating an encoding vector according to an embodiment of the present application;

[0022] Figure 6is a flowchart of extracting a topological feature by a topological perception block according to an embodiment of the present application;

[0023] Figure 7 is a flowchart of feature fusion by a hierarchical feature fusion block according to an embodiment of the present application;

[0024] Figure 8 is a flowchart of generating a target point cloud data set by a point cloud cluster generation module according to an embodiment of the present application;

[0025] Figure 9 is a flowchart of upsampling by a space-channel upsampling block according to an embodiment of the present application;

[0026] Figure 10 is a flowchart of contrastive learning by a multi-layer point cloud cluster contrastive learning module according to an embodiment of the present application;

[0027] Figure 11 is a schematic diagram of an asset evaluation device according to an embodiment of the present application;

[0028] Figure 12 is a structural block diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0029] In order to enable persons skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by persons skilled in the art without creative labor should fall within the scope of protection of the present application.

[0030] It should be noted that the terms "first", "second", and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0031] It should be noted that the information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for display, analyzed data, etc.) involved in the present application are information and data authorized by the user or authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of related data comply with relevant laws, regulations and standards, necessary security measures are taken, do not violate public order and good customs, and provide corresponding operation portal for user to choose authorization or refusal. For example, the system and the interface between the related users or institutions are provided with the corresponding operation portal for the user to choose to agree or refuse the automatic decision result; if the user chooses to refuse, the expert decision process is entered.

[0032] Embodiment 1

[0033] According to the embodiments of the present application, a method for asset evaluation is also provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that here.

[0034] The method provided by the embodiment of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Figure 1 A hardware structure block diagram of a computer terminal (or mobile device) for implementing the asset evaluation method is shown. As shown in Figure 1 , the computer terminal 10 (or mobile device) can include one or more (in the figure, 102a, 102b, …, 102n are shown) processors 102 (the processor 102 can include but not limited to a processing device such as MCU (Microcontroller Unit, microprocessor) or FPGA (Field-Programmable Gate Array, programmable logic device) etc.), a memory 104 for storing data, and a transmission device 106 for communication function. In addition, it can also include a display, an input / output interface (I / O interface), a USB (Universal Serial Bus, Universal Serial Bus) port (which can be included as one of the ports in the BUS (Business, Bus) port), a network interface, a power supply and / or a camera. Those skilled in the art can understand that Figure 1 The structure shown is only schematic, which does not limit the structure of the above-mentioned electronic device. For example, the computer terminal 10 can also include more or less components than those shown in Figure 1 , or have a different configuration from that shown in Figure 1 .

[0035] It should be noted that the one or more processors 102 and / or other data processing circuitry described above can be generally referred to herein as "data processing circuitry". The data processing circuitry can be embodied in whole or in part as software, hardware, firmware, or any combination thereof. Furthermore, the data processing circuitry can be a single standalone processing module, or incorporated in whole or in part within any of the other elements of the computer terminal 10 (or mobile device). As referred to in embodiments of the present application, the data processing circuitry functions as a processor to control, for example, the selection of the variable resistance terminal path connected to the interface.

[0036] The memory 104 can be used to store software programs of application software and modules, such as program instructions / data storage means corresponding to the asset evaluation method in embodiments of the present application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, i.e. implements the asset evaluation method described above. The memory 104 can include a high-speed random access memory, and can also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 can further include a memory remotely arranged with respect to the processor 102, which can be connected to the computer terminal 10 through a network. Examples of the above-mentioned network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0037] The transmission device 106 is used to receive or send data via a network. Specific examples of the above-mentioned network can include a wireless network provided by a communication provider of the computer terminal 10. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency (Radio Frequency, RF) module, which is used to communicate with the Internet in a wireless manner.

[0038] The display can be, for example, a touch screen type liquid crystal display, which can enable a user to interact with the user interface of the computer terminal 10 (or mobile device).

[0039] Under the above operating environment, the present application provides a point cloud completion system. Figure 2 is a structural schematic diagram of the point cloud completion system provided according to embodiments of the present application, as shown in Figure 2 The system includes a point agent generation module 201, a self-encoder architecture 202, a point cloud cluster generation module 203, a point cloud generation module 204, and a multi-layer point cloud cluster contrast learning module 205.

[0040] The point proxy generation module 201 generates a set of point cloud proxies, which can be regarded as point cloud clusters. The auto-encoder architecture 202 restores the missing point proxies. In addition, the multi-level point cloud cluster contrast learning module 205 maximizes the mutual information between two point patches extracted by the encoder and the decoder. Then, the point cloud cluster generation module 203 outputs the sketch of the point cloud. The point cloud generation module 204 reconstructs the fine-grained and locally structured point cloud.

[0041] The auto-encoder architecture 202 includes a multi-level feature learning block, a hierarchical feature fusion block, and a topology-aware block. The point cloud generation module 204 includes a spatial-channel upsampling block. Specifically, the multi-level feature learning block includes three different branch networks: high-resolution feature learning, medium-resolution feature learning, and low-resolution feature learning. Each branch contains different information from multiple levels and extracts features for different levels of regions. The hierarchical feature fusion block introduces high-resolution output features, medium-resolution output features, and low-resolution output features into the hierarchical feature fusion block according to the processing of the encoder on the hierarchical point cloud proxies. This enables the block to effectively integrate information from regions of different scales, facilitating learning and combination across different resolution features. The topology-aware block enhances the auto-encoder architecture by introducing topology awareness. This extension can improve the utilization of topological information bias present in the point cloud. The point cloud generation module 204 is decomposed into three layers. Each layer contains a spatial-channel upsampling block, which includes spatial window attention and channel group attention mechanisms.

[0042] The multi-level point cloud cluster contrast learning module 205 of the point cloud completion system maximizes the mutual information between the encoder and the decoder, fully utilizes the local geometric structure in the local input, and designs a spatial-channel upsampling block to guide the decoder to construct a complete and fine-grained point cloud. It is widely used in the financial industry, especially in asset evaluation fields such as agriculture and fisheries, and can provide high-precision data support and decision-making basis, significantly improving decision-making efficiency and accuracy, and having important practical significance and commercial value.

[0043] The present application provides an asset evaluation method using the above point cloud completion system. Figure 3 The flowchart of the asset evaluation method provided by the embodiments of the present application is shown in FIG. 1. Figure 3 The method comprises the following steps.

[0044] In step S301, a set of point cloud data of a target collateral is received, and N groups of point cloud clusters are extracted from the set of point cloud data, where N is a positive integer.

[0045] In step S301, the point cloud data set of the target collateral is processed by the point proxy generation module 201 to generate a point cloud proxy. The goal of the point proxy generation module 201 is to convert the point cloud into a sequence of vectors by generating a set of point cloud proxies. The point cloud data set can be obtained by 3D scanning devices such as lidar, depth camera or satellite remote sensing technology, etc. It contains a large amount of point coordinate data in three-dimensional space, which can accurately describe the shape, surface properties and environmental details of the object. A clustering algorithm is used to segment N point cloud clusters from the point cloud data set.

[0046] In step S302, for each point cloud cluster, the point cloud center of the point cloud cluster is determined, the local features of the point cloud cluster are extracted, and the point cloud proxy of the point cloud cluster is determined based on the point cloud center and the local features, and N point cloud proxies are obtained.

[0047] In step S302, Figure 4 The point proxy generation module provided by the embodiment of the present application is shown in the flow chart of generating a point cloud proxy as shown in Figure 4 The point proxy generation module locates a set of point cloud centers from the given incomplete point cloud (i.e. the point cloud data set). The local features and the position information of the point cloud center are extracted by a lightweight dynamic graph convolutional neural network and a multi-layer perceptron, respectively. The point proxy generation module combines the extracted local features and the position information of the point cloud center to obtain the point cloud proxy.

[0048] In step S303, the N point cloud proxies are input into the encoder of the target model to obtain the encoding vector of each point cloud proxy, and the N encoding vectors are input into the decoder of the target model to obtain the missing point proxy.

[0049] In step S303, the input of the encoder is N point cloud proxies, each containing the position information of the point cloud center and the local features related thereto. These proxies are generated from the original point cloud data set by clustering and feature extraction. The task of the encoder is to convert each point cloud proxy into an encoding vector. The encoding vector can effectively capture the key information of the point cloud proxy while reducing the redundancy of the data, facilitating network transmission and processing. The encoder can use a deep learning model, which can process sequence data and encode the features of the point cloud proxy in multiple layers. Through the self-attention mechanism, the model can learn the relationship between the point cloud proxies and the internal points of the point cloud proxy, thereby generating a more rich encoding vector.

[0050] The goal of the decoder is to convert the encoding vector into point cloud data, i.e. to generate a missing point proxy through the decoding process. The decoder is also based on deep learning and uses a matching architecture with the encoder to restore the information in the encoding vector and generate the missing point proxy. The decoder will gradually increase the dimension of the data until a predicted point cloud data similar to the original point cloud is generated.

[0051] Step S304, based on the missing point proxy, the missing point cloud in the point cloud data set is completed to obtain the target point cloud data set, and the target collateral is asset evaluated through the target point cloud data set to obtain the asset evaluation result.

[0052] In step S304, the missing point proxy is the result of the decoder output in the encoder-decoder architecture, which represents the prediction proxy for the missing or incomplete part of the point cloud data set. Each missing point proxy contains the predicted point cloud center position and local features, which will be used to complete the blank or damaged part in the original point cloud data set. Point cloud completion is achieved by a contrastive learning mechanism, which fuses the missing point proxy with the original point cloud data set to generate a complete target point cloud data set.

[0053] After point cloud completion, the target collateral whose point cloud data is missing due to occlusion, transmission error or device limitation has a complete, continuous and detailed point cloud data set, which provides a high-quality three-dimensional data basis for subsequent asset evaluation. The complete target point cloud data set contains comprehensive three-dimensional information of the target collateral, including but not limited to its shape, size, texture and position, etc. Using the complete point cloud data set, financial institutions can apply three-dimensional analysis techniques such as volume calculation, surface analysis and structure recognition to accurately measure the real size and structure of the collateral. In addition, combined with industry knowledge and market data, a special valuation model can be built to calculate a reasonable asset evaluation result based on the point cloud data.

[0054] The asset evaluation method of the present embodiment has broad application prospects in the financial industry, especially when financial institutions use satellite remote sensing technology to evaluate agricultural and fishery assets, it can provide more accurate data support and decision basis. This basis not only improves the accuracy of the data, but also provides more accurate analysis and decision support for related industries in asset evaluation, environmental monitoring and other aspects, which has important practical significance and commercial value.

[0055] The asset evaluation method provided in the embodiments of the present application comprises the following steps: receiving a point cloud data set of a target mortgage, extracting N groups of point cloud clusters from the point cloud data set, wherein N is a positive integer; for each group of point cloud clusters, determining a point cloud center of the point cloud cluster, extracting a local feature of the point cloud cluster, determining a point cloud agent of the point cloud cluster based on the point cloud center and the local feature, and obtaining N point cloud agents; inputting the N point cloud agents into an encoder of a target model to obtain an encoding vector of each point cloud agent, inputting the N encoding vectors into a decoder of the target model to obtain a missing point agent, and based on the missing point agent, complementing missing point clouds in the point cloud data set to obtain a target point cloud data set, performing asset evaluation on the target mortgage through the target point cloud data set to obtain an asset evaluation result. By extracting point cloud clusters from the point cloud data set and determining point cloud centers and local features, the missing point clouds are accurately reconstructed, and the integrity and quality of the point cloud data are greatly improved. The point cloud agent determined based on the point cloud center and the local feature captures and restores the detailed features of the point cloud, so that the complemented point cloud is more realistic and has more details. Using the complete and fine-grained target point cloud data set for asset evaluation can more accurately measure and analyze key parameters such as the size, shape and structure of the mortgage, so as to obtain a more accurate asset evaluation result. The purpose of improving the quality and integrity of the point cloud data is achieved, thereby realizing the technical effect of improving the accuracy of asset evaluation, and further solving the technical problem of inaccurate asset evaluation due to missing point cloud data.

[0056] The encoder generates an encoding vector of each point cloud agent. Optionally, in the asset evaluation method provided in the embodiments of the present application, inputting the N point cloud agents into the encoder of the target model to obtain the encoding vector of each point cloud agent comprises: converting each point cloud agent into a feature vector, performing linear projection and max-pooling operation on each feature vector to obtain a global feature of each point cloud agent; performing linear transformation on the global feature of each point cloud agent to obtain the coordinates of the point cloud agent; determining a concatenated representation of the global feature and the coordinates of each point cloud agent, and inputting each concatenated representation into a multi-layer perception machine to obtain the encoding vector of the point cloud agent.

[0057] In some embodiments, each point cloud agent is first transformed into a feature vector. The encoder quantizes the geometry, position information, and local features of the point cloud agent, enabling it to be represented in the form of a numerical vector. Next, a linear projection is performed on the feature vector of each point cloud agent, mapping the data from a high-dimensional space to a lower-dimensional space, which helps with data dimensionality reduction and feature extraction. Subsequently, a max-pooling operation is performed, which effectively captures the dominant features of the point cloud agent, i.e., the global features, by selecting the maximum values in the feature map, while reducing computational complexity. After obtaining the global features, a linear transformation is performed on each global feature to generate the coordinates of the point cloud agent. The abstract global features are remapped to the physical space, allowing the center position of each point cloud agent to be explicitly represented.

[0058] The concatenated representation of the global features and coordinates of each point cloud agent is input into a multi-layer perceptron for non-linear data transformation and feature learning. In this process, the multi-layer perceptron further processes the concatenated representation of the point cloud agent through multiple layers of non-linear transformations, ultimately outputting an encoding vector for the point cloud agent. The encoding vector is a high-order feature representation of the point cloud agent.

[0059] For example, Figure 5 is a flowchart of generating an encoding vector based on the autoencoder architecture provided in the embodiments of the present application, as Figure 5 shown, the encoder based on the autoencoder architecture 202 receives input data and transforms it into an intermediate representation or feature vector. Operations are performed on each feature vector to generate global features using linear projection and max-pooling. Linear transformation is performed on each global feature to generate coordinates. By combining the global features and coordinates, an MLP (Multi-layer Perceptron) is applied on their concatenated representation to derive a query embedding dynamic query as the initial state of the predicted agent. The intermediate representation is converted back to a part of the original data space by the decoder of the autoencoder architecture 202, which is used in turn to predict the missing point agent.

[0060] The embodiments effectively extract and condense the features of the point cloud agent through the encoder encoding process, from geometric details to global characteristics, by linear projection, max-pooling, and linear transformation operations, transforming complex point cloud data into compact encoding vectors, which facilitates efficient processing in the subsequent decoding stage. The feature condensation and abstraction of the point cloud agent are achieved, and through dimensionality reduction and information preservation strategies, efficient processing of the point cloud agent and preservation of key information are ensured, and the final output encoding vector provides a solid data foundation for subsequent point cloud completion and asset evaluation.

[0061] In order to extract features in the point cloud agent, the point cloud agent needs to be converted into a feature vector. Optionally, in the asset evaluation method provided in the embodiments of the present application, converting each point cloud agent into a feature vector includes: performing feature extraction on the point cloud agent through a multi-level feature learning block of an encoder to obtain M feature maps of different resolution scales, wherein the multi-level feature learning block includes M branch networks, each branch network is used to extract a feature map of a resolution scale from the point cloud agent, and M is a positive integer; capturing topological features of the point cloud agent through a topological perception block of the encoder, wherein the topological features are used to represent features of local geometric structures of the point cloud agent; extracting features in the M feature maps of different resolution scales through a hierarchical feature fusion block of the encoder to obtain M features of different resolution scales, performing feature fusion on the M features of different resolution scales to obtain a fusion feature; and integrating the fusion feature and the topological features to obtain a feature vector of the point cloud agent, wherein the integration processing includes at least one of the following: feature splicing, weighted summation, and network layer mapping.

[0062] In some embodiments, the multi-level feature learning block is an important component in the encoder, which is used to extract multi-scale feature information from the point cloud agent to enhance the understanding ability of the model to the details and macro structure of the point cloud. The multi-level feature learning block contains M branch networks, each of which focuses on extracting a feature map of a specific resolution scale. From high-resolution local details to low-resolution macro structures, each part of the point cloud agent has the opportunity to be analyzed in detail by different levels of networks. After the point cloud agent is input into the multi-level feature learning block and processed by the M branch networks, M feature maps of different resolution scales are generated. This ensures that the feature information of the point cloud agent can be fully captured at multiple scales.

[0063] The topological perception block plays a role in enhancing the understanding of the local geometric structure of the point cloud in the encoder. Topological features refer to the connectivity and geometric layout between points in the point cloud agent, which can reveal the local shape and texture information of the point cloud. The topological perception block retrieves features related to the nearest neighbor points through KNN (K-Nearest Neighbors), calculates the distance between each point in the point cloud agent and all other points, and selects the K points with the smallest distance as the neighborhood points. After obtaining the local neighborhood points, the retrieved features are further mapped through a linear layer and a max-pooling operation, thereby extracting the local geometric structure features of the point cloud agent.

[0064] The hierarchical feature fusion block is used to integrate features of different resolution scales, ensuring that the model can learn using both local details and global information. The hierarchical feature fusion block receives feature maps of M different resolution scales from the multi-level feature learning block, calculates the similarity between high-resolution features and medium-resolution features using a fine-to-coarse attention mechanism, and generates attention weights. Then, the high-resolution features are weighted based on these weights, and the low-resolution features are combined to update the weighted features using a coarse-to-fine attention mechanism, ensuring that key global information is retained during information fusion. The processed features of different scales are weighted and summed to obtain a fused feature map, which is then passed through an activation function to generate new hierarchical information. Finally, all weighted features of different scales are concatenated into a final output feature, i.e., the fused feature.

[0065] The final output of the encoder is the feature vector of the point cloud agent, which is integrated from the fused features and the topological features. The integration method includes feature splicing, weighted summation, and mapping through network layers. Feature splicing concatenates different types of features to form a vector containing more information; weighted summation weights the features according to their importance and then sums them to integrate the effects of different features; network layer mapping converts the features through additional network layers to make them more suitable as components of the encoder output feature vector.

[0066] The combination of multi-level feature learning and topology perception technology enables the encoder to understand the point cloud agent from multiple perspectives, capturing not only local details but also global topological structures, greatly enhancing the model's expressive power. The hierarchical feature fusion block and integration processing ensure that the feature vector output by the encoder is an efficient integration of multi-scale and multi-type features, making the feature information more rich and providing strong support for subsequent decoder processing. Through the multi-level feature learning, topology perception, and hierarchical feature fusion techniques of the encoder, multi-level feature information can be extracted and integrated from the point cloud agent, significantly enhancing the robustness and accuracy of the point cloud completion method.

[0067] After obtaining the hierarchical features of different scales of the point cloud agent, the topology features of the point cloud agent need to be extracted. Optionally, in the asset evaluation method provided in the present application, the topology features of the point cloud agent are captured by the topology perception block of the encoder, including: determining the query coordinates and key value coordinates of the point cloud agent, wherein the query coordinates are used to focus on the local area of the point cloud agent, and the key value coordinates are used to represent the context information of the local area; retrieving K points near the query coordinates through a pre-set machine learning model and the key value coordinates, and determining the K points as neighborhood points, wherein K is a positive integer; extracting features of the K neighborhood points to obtain neighborhood features, and filtering out topology features from the neighborhood features through a max-pooling operation.

[0068] In some embodiments, in the point cloud proxy processing, the query coordinates play a role in guiding the model to focus on a specific local area of the point cloud. The key-value coordinates provide contextual information of the local area, including the location and features of neighboring points, which is crucial for capturing the local topological properties of the point cloud proxy. Through the KNN model, the encoder retrieves the K nearest neighbor points closest to the query coordinates using the key-value coordinates. The number K of these neighbor points carries detailed information of the point cloud proxy around the query point, and the choice of K affects the depth and accuracy of the capture of local information. From the retrieved K neighbor points, the topology-aware block extracts the feature information of each point, including location, shape, texture, and other geometric properties, forming the neighborhood features. In order to select topological features from the neighborhood features, the topology-aware block uses the max-pooling operation. Max-pooling is a dimensionality reduction and feature selection method that selects the maximum or most significant features from the neighborhood features, thus retaining the most important local geometric information. This operation helps to remove redundant information and highlight key topological properties.

[0069] For example, Figure 6 is a flowchart of the topology-aware block extracting topological features according to an embodiment of the present application, as Figure 6 shown, the topology-aware block uses query coordinates to help the model focus on a specific area, and key-value coordinates to provide contextual information of the area. The KNN model is used to retrieve features related to the nearest key-value, and the distances between the query coordinates and all other points are calculated, and the K points with the smallest distance are selected as the neighborhood points. These features are passed through a linear layer and max-pooling is applied to capture the local geometric structure present in the point cloud. By merging the features of the K neighborhood points into a new space, further feature mapping is performed on the features retrieved by KNN. The max-pooling operation is used to extract the most significant features from the K neighborhood point features, and the most important information is retained by selecting the maximum value in the local area. Gradually integrate local geometric features with global semantic features, integrate information at multiple levels, and finally output features with the same original dimension, but with richer semantic and geometric information inside.

[0070] This embodiment can effectively capture the local geometric structure features of the point cloud proxy, especially the topological properties, through the query coordinate and key-value coordinate mechanism, combined with KNN retrieval and max-pooling operation. This not only improves the accuracy and efficiency of point cloud data processing, but also provides a more powerful algorithm tool for asset evaluation and credit risk management based on point clouds, thereby enhancing the decision-making ability of banks and financial institutions in processing complex remote sensing data.

[0071] After obtaining the features of M different resolution scales, feature fusion is performed through a hierarchical feature fusion block. Optionally, in the asset evaluation method provided in the embodiments of the present application, the M different resolution scales include a first resolution, a second resolution, and a third resolution. The features of the M different resolution scales are fused to obtain a fused feature, including: calculating the similarity between the first resolution feature and the second resolution feature to obtain a first attention weight, performing weighted processing on the first resolution feature based on the first attention weight to obtain a first feature, wherein the first resolution is greater than the second resolution; calculating the similarity between the third resolution feature and the second resolution feature to obtain a second attention weight, performing weighted processing on the third resolution feature based on the second attention weight to obtain a second feature, wherein the second resolution is greater than the third resolution; and performing weighted summation on the first feature, the second feature, and the second resolution feature to obtain the fused feature.

[0072] In some embodiments, the similarity between the first resolution feature and the second resolution feature is first calculated to obtain a first attention weight. The attention weight here reflects the correlation between the two features and can help the model identify which features should be given more attention. Then, the similarity between the third resolution feature and the second resolution feature is calculated to obtain a second attention weight. Based on the first attention weight, the first resolution feature is weighted processed to obtain a first feature that focuses more on key information. Based on the second attention weight, the third resolution feature is weighted processed to obtain a second feature that emphasizes local details. The weighted first feature, the second feature, and the unweighted second resolution feature are weighted summed to finally obtain the fused feature. This fusion method considers the importance and complementarity of features of different resolutions, ensuring that the fused feature can cover both macro structures and local details.

[0073] For example, Figure 7 is a flowchart of feature fusion performed by the hierarchical feature fusion block provided in the embodiments of the present application, as shown in Figure 7As shown, the hierarchical feature fusion block extracts local region features from high-resolution and medium-resolution feature maps. The high-resolution feature map contains more detailed information, while the medium-resolution feature map provides coarser semantic information. A fine-to-coarse attention mechanism is used to generate attention weights by calculating the similarity between high-resolution features and medium-resolution features. Based on the calculated attention weights, the high-resolution features are weighted to ensure that the high-resolution features can be adjusted appropriately according to the semantic information of the medium-resolution features. The low-resolution is processed by the convolution layer to obtain the low-resolution feature representation. Using a coarse-to-fine attention mechanism, the similarity between the low-resolution features and the medium-resolution features is first calculated. According to the calculated correlation weight, the representation of the low-resolution features is adjusted. The low-resolution features will be updated by weighting according to the semantic correlation of the medium-resolution features, ensuring that the key global information will not be lost during information fusion. The high-resolution weighted features, low-resolution weighted features and medium-resolution features are weighted and summed to obtain a fused feature map. The fused feature map is passed into an activation function to generate new hierarchical information. All weighted features of different scales are concatenated into a final output feature.

[0074] This embodiment effectively integrates feature information of different resolution scales by calculating attention weights and performing multi-stage weighted summation, improving the comprehensiveness and fineness of the point cloud agent processing process. The accuracy of data processing is improved.

[0075] After obtaining the missing point agent, the target point cloud data combination is generated. Optionally, in the asset evaluation method provided in the present application, the missing point cloud in the point cloud data set is completed based on the missing point agent to obtain the target point cloud data set, including: extracting shape features from the missing point agent, determining the activation state of the missing point agent; generate missing point cloud features based on the activation state and shape features, determine the shape code of the missing point agent through the point inflation strategy and the missing point cloud features; concatenate the shape code and the missing point cloud features to generate a rough point cloud, integrate the rough point cloud with the point cloud data set to obtain an updated point cloud data set; generate a point cloud seed through a far point sampling algorithm and the updated point cloud data set, refine the point cloud seed through an upsampling operation, and obtain the target point cloud data set.

[0076] In some embodiments, first, shape features are extracted from the missing point proxies, which contain geometric properties of the missing regions, such as edges, curvatures, etc., helping to understand and recover the local structure of the missing parts. Then, the activation states of the missing point proxies are determined, which reflect the activity or importance of the point cloud data in the completion process, helping to guide the subsequent point cloud generation process. Based on the extracted shape features and activation states, missing point cloud features are generated. Through the point inflation strategy, the influence range of the missing point cloud features is expanded, which helps to better fill the blank areas in the completion process and ensures that the completed point cloud data can be naturally integrated into the original point cloud. The shape encoding of the missing point proxy is generated by combining the missing point cloud features with the point inflation strategy. Shape encoding is a higher-level feature representation that contains shape information and contextual clues of the missing point cloud. The shape encoding and the missing point cloud features are concatenated as the input of the shared MLP, and processed by the ReLU (Rectified Linear Unit, an activation function) activation function to generate the initial version of the rough point cloud.

[0077] The generated rough point cloud is integrated with the existing point cloud dataset to obtain an updated point cloud data set. A far point sampling algorithm is used to generate point cloud seeds from the updated point cloud data set. Far point sampling aims to ensure that the generated seed points are uniformly distributed on the point cloud. Through upsampling operations, the point cloud seeds are gradually refined. Additional points are added within the neighborhood of the point cloud seeds to increase the details and resolution of the point cloud. The upsampling operation can gradually build a more detailed point cloud structure until the desired level of detail is reached. The upsampling operation is repeatedly executed until a sufficiently detailed and accurate target point cloud data set is obtained. This set contains the completed and refined point cloud data, which can more accurately reflect the details and structure of the original point cloud.

[0078] For example, Figure 8 is a flowchart of generating a target point cloud data set by a point cloud cluster generation module according to an embodiment of the present application, as Figure 8 shown, the point cloud cluster generation module extracts features from the missing point proxies. Shape features of each missing point proxy are captured, and sub-point features are generated by combining the activation states. This process allows the method to obtain existing and missing shape encodings through point inflation. A shared MLP consisting of three linear layers is used, and each point feature is concatenated with the shape encoding using the ReLU activation to generate a rough point cloud. The rough point cloud is integrated with each input point cloud, and further point cloud seeds are generated through FPS (Farthest Point Sampling, farthest point sampling). The output is passed through three consecutive layers. Each layer takes the point cloud from the previous layer to perform progressive upsampling operations and gradually refine the point cloud data, thereby generating a dense output.

[0079] This embodiment forms a systematic point cloud completion process from shape feature extraction to point inflation strategy, to rough point cloud generation and upsampling operation. It effectively supplements the missing point cloud data and can gradually refine the point cloud to ultimately obtain a complete and detailed target point cloud data set.

[0080] After obtaining the point cloud seed, the point cloud seed is refined by upsampling. In the asset evaluation method provided in the embodiments of the present application, the target point cloud data set obtained by refining the point cloud seed through upsampling operation includes: extracting the features of the point cloud seed through a local spatial attention mechanism and a global channel attention mechanism to obtain target features; inputting the target features into an upsampling block to complete the point cloud from the point cloud seed through step-by-step upsampling operation to obtain an initial completed point cloud; adjusting the initial completed point cloud through a contrast learning module to obtain the target point cloud data set.

[0081] In some embodiments, first, local spatial features are extracted from the point cloud seed, focusing on the position of each point in the point cloud seed in the local space and the nearby point cloud structure, capturing the local geometric information of the point cloud. Then, a global channel attention mechanism is used to analyze the global features of the point cloud seed, which focuses on the long-distance dependency relationship between point cloud channel labels and can extract the overall structure and semantic information of the point cloud. Combining these two mechanisms, the algorithm generates target features, which is a high-level feature representation containing local and global information of the point cloud seed, providing a rich information base for subsequent upsampling operations. The target features are input into the upsampling block, which increases the number of points in the point cloud seed through step-by-step upsampling operation to supplement the missing point cloud details.

[0082] The output of the upsampling block is the initial completed point cloud, which is more dense than the original point cloud seed and contains more point cloud details and structure information, but may still need further optimization in some aspects. The initial completed point cloud is then sent to the contrast learning module, which adjusts the geometric shape and details of the point cloud through the contrast learning mechanism to ensure the accuracy of the point cloud completion. The goal of the contrast learning module is to minimize the difference between point cloud clusters to make the initial completed point cloud closer to the structure of the real point cloud. After optimization by the contrast learning module, the details and structure of the point cloud are further improved, and the final generated target point cloud data set has higher quality and more accurate geometric shape.

[0083] For example, Figure 9 is a flowchart of upsampling performed by the spatial-channel upsampling block provided in the embodiments of the present application, as Figure 9As shown, the spatial-channel upsampling block divides the feature space into non-overlapping local regions. Within each local region, feature points are indexed by spatial tokens and self-attention computation is performed separately. Tokens are divided into multiple groups along the sequence direction according to the channel dimension. The complete point cloud channel is treated as a token, and each divided group is independently subjected to self-attention operation. By combining the two mechanisms, the model can perform multi-dimensional processing on the point cloud at different levels.

[0084] The embodiment designs a spatial-channel upsampling block, which uses the enhanced point cloud clusters obtained by contrast learning to guide the decoder to construct complete and fine-grained point clouds. By combining local spatial attention mechanisms and global channel attention mechanisms, both local details and overall structures of the point cloud can be focused on, ensuring that the completed point cloud data has good performance in terms of detail richness and structural integrity. The combination of upsampling operations and contrast learning modules realizes gradual refinement and optimization during the point cloud completion process, effectively handling missing parts of the point cloud data and generating a high-quality point cloud data set close to the real point cloud. Through the contrast learning mechanism, the accuracy of the completed point cloud is improved.

[0085] To make the generated point cloud more accurate, the initial completed point cloud also needs to be adjusted through the contrast learning module. Optionally, in the asset evaluation method provided in the embodiment of the present application, adjusting the initial completed point cloud through the contrast learning module to obtain the target point cloud data set includes: determining the first point cloud features extracted by the encoder from the N point cloud agents, and determining the second point cloud features extracted from the initial completed point cloud; comparing the first point cloud features and the second point cloud features to obtain contrast features, and adjusting the initial completed point cloud based on the contrast features to obtain the target point cloud data set.

[0086] In some embodiments, the contrast learning module, i.e., the multi-layer point cloud cluster contrast learning module 205, first extracts the first point cloud features from the N point cloud agents by the encoder. Then, the second point cloud features are extracted from the initial completed point cloud. The second point cloud features reflect the current state and potential defects of the initial completed point cloud. By comparing the first point cloud features and the second point cloud features, contrast features are generated. These contrast features reveal the differences between the information extracted by the encoder and the actual state of the initial completed point cloud, providing a basis for subsequent adjustments. Based on the contrast features, the algorithm can identify and correct inaccuracies or missing details in the initial completed point cloud, such as geometric shape adjustment, point cloud density optimization, and local structure refinement, ensuring that the generated point cloud data set is significantly improved in terms of geometric shape, detail richness, and structural integrity.

[0087] Finally, after the adjustment and optimization of the contrast learning module, the initial completed point cloud is transformed into a more accurate and complete point cloud data set, i.e., the target point cloud data set. This set not only meets the target requirements in terms of point number, but also is closer to the real point cloud in terms of details and geometric accuracy, and can provide high-precision and reliable data support for banks and financial institutions, especially in the asset assessment of agricultural planting areas or fishery waters, the target point cloud data set can provide more detailed crop growth conditions or water environment information, providing a solid basis for credit decision-making and risk management.

[0088] For example, Figure 10 is a flowchart of the multi-layer point cloud cluster contrast learning module provided by the embodiment of the present application for contrast learning, as Figure 10 shown, the multi-layer point cloud cluster contrast learning module 205 strengthens the relationship between the features extracted by the encoder and the decoder, and adjusts the geometric shape of the generated point cloud: the multi-layer point cloud cluster contrast learning module 205 selects representative layers in the encoder. Stack these layers to generate a multi-layer feature stack. The output point cloud of each decoding stage is encoded. The features are divided into multiple patches (small regions or subsets extracted from the original data). Use PatchNCE (Patchwise Noise-Contrastive Estimation, a contrast learning strategy) loss to calculate the geometric distance or similarity between each pair of input and output blocks, and align the geometric information between the corresponding input and output blocks at a given position.

[0089] The embodiment maximizes the mutual information between the two point cloud clusters of the encoder and the decoder through the multi-layer point cloud cluster contrast learning module, and through contrast learning, it can identify specific defects in the point cloud data, such as geometric inaccuracy or detail loss, and then make targeted adjustments to improve the accuracy and detail richness of the point cloud completion. The contrast between the features extracted by the encoder and the features of the initial completed point cloud helps to enhance the detail expression and structural integrity of the point cloud data, making the generated point cloud data more realistic and reliable.

[0090] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown.

[0091] Embodiment 2

[0092] The embodiment of the present application also provides an asset evaluation device. It should be noted that the asset evaluation device of the embodiment of the present application can be used to execute the asset evaluation method provided by the embodiment of the present application. The asset evaluation device provided by the embodiment of the present application is introduced as follows.

[0093] According to the embodiments of the present application, a device for implementing the asset evaluation method is also provided, Figure 11 is a schematic diagram of an asset evaluation device provided by the embodiments of the present application, as shown in Figure 11 The device comprises:

[0094] The receiving unit 1101 is configured to receive a point cloud data set of a target collateral, and extract N groups of point cloud clusters from the point cloud data set, wherein N is a positive integer.

[0095] The first determining unit 1102 is configured to determine, for each group of point cloud clusters, a point cloud center of the point cloud cluster, extract a local feature of the point cloud cluster, determine a point cloud agent of the point cloud cluster based on the point cloud center and the local feature, and obtain N point cloud agents.

[0096] The input unit 1103 is configured to input the N point cloud agents into an encoder of a target model to obtain an encoding vector of each point cloud agent, input the N encoding vectors into a decoder of the target model to obtain a missing point agent.

[0097] The completion unit 1104 is configured to complete a missing point cloud in the point cloud data set based on the missing point agent to obtain a target point cloud data set, and perform asset evaluation on the target collateral through the target point cloud data set to obtain an asset evaluation result.

[0098] The asset evaluation device provided in the embodiments of the present application receives the point cloud data set of the target mortgage through the receiving unit 1101, extracts N groups of point cloud clusters from the point cloud data set, wherein N is a positive integer; the first determining unit 1102 determines the point cloud center of each group of point cloud clusters, extracts the local features of the point cloud cluster, determines the point cloud agent of the point cloud cluster based on the point cloud center and the local features, and obtains N point cloud agents; the input unit 1103 inputs the N point cloud agents into the encoder of the target model to obtain the encoding vector of each point cloud agent, inputs the N encoding vectors into the decoder of the target model to obtain the missing point agent, and the completion unit 1104 completes the missing point cloud in the point cloud data set based on the missing point agent to obtain the target point cloud data set, performs asset evaluation on the target mortgage through the target point cloud data set, and obtains the asset evaluation result. By extracting the point cloud cluster from the point cloud data set and determining the point cloud center and the local features, the missing point cloud is accurately reconstructed, and the integrity and quality of the point cloud data are greatly improved. The point cloud agent determined based on the point cloud center and the local features captures and restores the detailed features of the point cloud, so that the completed point cloud is more realistic and has more details. The asset evaluation is performed by using the complete and fine-grained target point cloud data set, which can more accurately measure and analyze the key parameters such as the size, shape and structure of the mortgage, so as to obtain a more accurate asset evaluation result. The purpose of improving the quality and integrity of the point cloud data is achieved, thereby realizing the technical effect of improving the accuracy of asset evaluation, and further solving the technical problem that the asset evaluation is inaccurate due to the missing of the point cloud data.

[0099] Optionally, in the asset evaluation device provided in the embodiments of the present application, the input unit 1103 includes: a transformation module, configured to transform each point cloud agent into a feature vector, perform linear projection and maximum pooling operation on each feature vector to obtain the global feature of each point cloud agent; a transformation module, configured to perform linear transformation on the global feature of each point cloud agent to obtain the coordinates of the point cloud agent; a first determination module, configured to determine the concatenated representation of the global feature and the coordinates of each point cloud agent, and input each concatenated representation into a multi-layer perception machine to obtain the encoding vector of the point cloud agent.

[0100] Optionally, in the asset evaluation apparatus provided by the embodiment of the present application, the conversion module comprises: a first extraction submodule, configured to perform feature extraction on the point cloud agent through a multi-level feature learning block of the encoder to obtain feature maps of M resolution scales, wherein the multi-level feature learning block comprises M branch networks, each branch network is configured to extract a feature map of a resolution scale from the point cloud agent, and M is a positive integer; a capture submodule, configured to capture topological features of the point cloud agent through a topological perception block of the encoder, wherein the topological features are used to represent features of local geometric structures of the point cloud agent; a second extraction submodule, configured to extract features in the M resolution scale feature maps through a hierarchical feature fusion block of the encoder to obtain features of M different resolution scales, perform feature fusion on the features of M different resolution scales to obtain a fusion feature; and an integration submodule, configured to perform integration processing on the fusion feature and the topological features to obtain a feature vector of the point cloud agent, wherein the integration processing comprises at least one of feature splicing, weighted summation, and network layer mapping.

[0101] Optionally, in the asset evaluation apparatus provided by the embodiment of the present application, the capture submodule comprises: a coordinate query module, configured to determine a query coordinate and a key-value coordinate of the point cloud agent, wherein the query coordinate is used to focus on a local area of the point cloud agent, and the key-value coordinate is used to represent context information of the local area; a query module, configured to retrieve K points near the query coordinate through a preset machine learning model and the key-value coordinate, and determine the K points as neighborhood points, wherein K is a positive integer; and a screening module, configured to extract features of the K neighborhood points to obtain neighborhood features, and screen the topological features from the neighborhood features through a max-pooling operation.

[0102] Optionally, in the asset evaluation apparatus provided by the embodiment of the present application, the second extraction submodule comprises: a first calculation module, configured to calculate a similarity between a first resolution feature and a second resolution feature to obtain a first attention weight, perform weighted processing on the first resolution feature based on the first attention weight to obtain a first feature, wherein the first resolution is greater than the second resolution; a second calculation module, configured to calculate a similarity between a third resolution feature and the second resolution feature to obtain a second attention weight, perform weighted processing on the third resolution feature based on the second attention weight to obtain a second feature, wherein the second resolution is greater than the third resolution; and a summation module, configured to perform weighted summation on the first feature, the second feature, and the second resolution feature to obtain the fusion feature.

[0103] Optionally, in the asset evaluation device provided in the embodiment of the present application, the completion unit 1104 comprises: an extraction module, configured to extract shape features from the missing point agent, and determine an activation state of the missing point agent; a first generation module, configured to generate missing point cloud features based on the activation state and the shape features, and determine shape coding of the missing point agent through a point inflation strategy and the missing point cloud features; a concatenation module, configured to concatenate the shape coding and the missing point cloud features, generate a rough point cloud, integrate the rough point cloud with the point cloud data set to obtain an updated point cloud data set; and a second generation module, configured to generate a point cloud seed through a far point sampling algorithm and the updated point cloud data set, and refine the point cloud seed through an upsampling operation to obtain a target point cloud data set.

[0104] Optionally, in the asset evaluation device provided in the embodiment of the present application, the second generation module comprises: a third extraction submodule, configured to extract features of the point cloud seed through a local spatial attention mechanism and a global channel attention mechanism to obtain target features; an input submodule, configured to input the target features into an upsampling block, and complete the point cloud from the point cloud seed through a step-by-step upsampling operation to obtain an initial completed point cloud; and an adjustment submodule, configured to adjust the initial completed point cloud through a contrast learning module to obtain the target point cloud data set.

[0105] Optionally, in the asset evaluation device provided in the embodiment of the present application, the adjustment submodule comprises: a second determination module, configured to determine first point cloud features extracted by the encoder from the N point cloud agents, and determine second point cloud features extracted from the initial completed point cloud; and a contrast module, configured to compare the first point cloud features and the second point cloud features to obtain contrast features, and adjust the initial completed point cloud based on the contrast features to obtain the target point cloud data set.

[0106] It should be noted that the above receiving unit 1101, first determination unit 1102, input unit 1103 and completion unit 1104 correspond to steps S201 to S204 in Embodiment 1, and the four units have the same instances and application scenarios as the corresponding steps, but are not limited to the content disclosed in Embodiment 1. It should be noted that the above modules or units can be hardware components or software components stored in a memory (for example, memory 104) and processed by one or more processors (for example, processors 102a, 102b, …, 102n), or can be a part of the device and can run in the computer terminal 10 provided in Embodiment 1.

[0107] Embodiment 3

[0108] Embodiments of the present application can provide an electronic device, Figure 12 is a structural block diagram of an electronic device according to an embodiment of the present application. As Figure 12 shown, the electronic device can include one or more (Figure 12 The processor 1202, the memory 1204, the storage controller, and the peripheral interface are shown in the figure. The peripheral interface is connected with a radio frequency module, an audio module, and a display.

[0109] The memory can be used to store software programs and modules, such as program instructions / modules corresponding to the method and device in the embodiments of the present application. The processor executes various functions and data processing by running the software programs and modules stored in the memory, that is, implements the above method. The memory can include a high-speed random access memory, and can further include a non-volatile memory, such as one or more magnetic storage devices, flash memories, or other non-volatile solid-state memories. In some examples, the memory can further include a memory remotely arranged with respect to the processor, which can be connected to the terminal through a network. Examples of the network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0110] The processor can call information and applications stored in the memory through the transmission device to execute the asset evaluation method.

[0111] Those skilled in the art can understand that, Figure 12 The structure shown is only schematic, and the electronic device can also be a terminal device such as a smart phone, a tablet computer, a palm computer, a Mobile Internet Device (MID), a PAD, and the like. Figure 12 This does not limit the structure of the above-mentioned electronic device. For example, the electronic device can further include more or fewer components (such as a network interface, a display device, and the like) than those shown in the figure, or have a different configuration from that shown in the figure. Figure 12 For example, the electronic device can further include more or fewer components (such as a network interface, a display device, and the like) than those shown in the figure, or have a different configuration from that shown in the figure. Figure 12 For example, the electronic device can further include more or fewer components (such as a network interface, a display device, and the like) than those shown in the figure, or have a different configuration from that shown in the figure.

[0112] Those skilled in the art can understand that all or part of the steps of the various methods of the above embodiments can be completed by programs instructing the relevant hardware of the terminal device, and the programs can be stored in a computer readable storage medium, which can include a flash disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and the like.

[0113] Embodiment 4

[0114] The embodiments of the present application also provide a storage medium. Optionally, in the present embodiment, the above-mentioned storage medium can be used to save the program code executed by the asset evaluation method provided in Embodiment 1.

[0115] Optionally, in the embodiment, the storage medium can be located in any one of computer terminals in a computer terminal group in a computer network, or in any one of mobile terminals in a mobile terminal group.

[0116] The application also provides a computer program product adapted to perform the steps of the asset valuation method when executed on a data processing device.

[0117] The above-mentioned embodiment numbers of the application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0118] In the above-mentioned embodiments of the application, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0119] In several embodiments provided by the present application, it should be understood that the disclosed technology can be implemented in other ways. Of course, the embodiment described above is only a schematic, for example, the division of units is only a logical function division, and there can be another division manner in actual implementation, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, unit or module, and can be electrical or other forms.

[0120] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. According to actual needs, part or all of the units can be selected to achieve the purpose of the embodiment.

[0121] In addition, each functional unit in each embodiment of the present application can be integrated into a processing unit, or each unit can exist physically, or two or more units can be integrated into one unit. The above integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0122] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or say the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.

[0123] The above is only the preferred embodiment of the present application, and it should be pointed out that for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which should be considered as the protection scope of the present application.

Claims

1. An asset evaluation method, characterized in that: include: Receiving a point cloud data set of a target collateral, and extracting N groups of point cloud clusters from the point cloud data set, where N is a positive integer; For each group of point cloud clusters, determining the point cloud center of the point cloud cluster, extracting local features of the point cloud cluster, and determining a point cloud proxy of the point cloud cluster based on the point cloud center and the local features to obtain N point cloud proxies; Input the N point cloud proxies into the encoder of the target model to obtain an encoding vector of each point cloud proxy, and input the N encoding vectors into the decoder of the target model to obtain a missing point proxy; The missing point clouds in the point cloud data set are completed based on the missing point agent to obtain a target point cloud data set, and an asset evaluation is performed on the target collateral using the target point cloud data set to obtain an asset evaluation result.

2. The method according to claim 1, characterized in that Input the N point cloud proxies into the encoder of the target model, and obtain the encoding vector of each point cloud proxy including: Each point cloud proxy is converted into a feature vector, and each feature vector is linearly projected and max pooled to obtain the global features of each point cloud proxy; Performing a linear transformation on the global features of each point cloud proxy to obtain the coordinates of the point cloud proxy; A cascade representation of the global features and coordinates of each point cloud agent is determined, and each cascade representation is input into a multi-layer perceptron to obtain an encoding vector of the point cloud agent.

3. The method according to claim 2, characterized in that Converting each point cloud proxy into a feature vector involves: Performing feature extraction on the point cloud proxy through a multi-level feature learning block of the encoder to obtain feature maps of M resolution scales, wherein the multi-level feature learning block includes M branch networks, each branch network is used to extract a feature map of a resolution scale from the point cloud proxy, and M is a positive integer; Capturing topological features of the point cloud proxy through a topology perception block of the encoder, wherein the topological features are used to characterize features of a local geometric structure of the point cloud proxy; Extracting features from the feature maps of the M resolution scales through a hierarchical feature fusion block of the encoder to obtain features of M different resolution scales, and fusing the features of the M different resolution scales to obtain a fused feature; The fusion feature and the topological feature are integrated to obtain a feature vector of the point cloud proxy, wherein the integration process includes at least one of the following: feature splicing, weighted summation, and network layer mapping.

4. The method according to claim 3, characterized in that Capturing the topological features of the point cloud proxy through the topology perception block of the encoder includes: Determining query coordinates and key value coordinates of the point cloud proxy, wherein the query coordinates are used to focus on a local area of ​​the point cloud proxy, and the key value coordinates are used to represent context information of the local area; Retrieving K points near the query coordinates using a preset machine learning model and the key-value coordinates, and determining the K points as neighborhood points, where K is a positive integer; The features of the K neighborhood points are extracted to obtain neighborhood features, and the topological features are filtered out from the neighborhood features through a maximum pooling operation.

5. The method according to claim 3, characterized in that The M different resolution scales include a first resolution, a second resolution, and a third resolution. The features of the M different resolution scales are subjected to feature fusion to obtain fused features including: Calculating a similarity between a first resolution feature and a second resolution feature to obtain a first attention weight, performing weighted processing on the first resolution feature based on the first attention weight to obtain a first feature, wherein the first resolution is greater than the second resolution; Calculating a similarity between a third resolution feature and the second resolution feature to obtain a second attention weight, and performing weighted processing on the third resolution feature based on the second attention weight to obtain a second feature, wherein the second resolution is greater than the third resolution; A weighted sum is performed on the first feature, the second feature, and the second resolution feature to obtain the fused feature.

6. The method according to claim 1, characterized in that Completing the missing point clouds in the point cloud data set based on the missing point proxy to obtain the target point cloud data set includes: extracting shape features from the missing point agent and determining an activation state of the missing point agent; generating missing point cloud features based on the activation state and the shape features, and determining the shape encoding of the missing point proxy using a point expansion strategy and the missing point cloud features; cascading the shape code and the missing point cloud features to generate a rough point cloud, and integrating the rough point cloud with the point cloud data set to obtain an updated point cloud data set; A point cloud seed is generated by using a far-point sampling algorithm and the updated point cloud data set, and the point cloud seed is refined by an upsampling operation to obtain a target point cloud data set.

7. The method according to claim 6, characterized in that The point cloud seed is refined by upsampling operation to obtain a target point cloud data set including: Extracting the features of the point cloud seeds through the local spatial attention mechanism and the global channel attention mechanism to obtain the target features; Inputting the target features into an upsampling block, completing the point cloud starting from the point cloud seed through a step-by-step upsampling operation, and obtaining an initial completed point cloud; The initial completed point cloud is adjusted through a contrast learning module to obtain the target point cloud data set.

8. The method according to claim 7, characterized in that The target point cloud data set is obtained by adjusting the initial completed point cloud through the contrast learning module, including: Determining a first point cloud feature extracted by the encoder from the N point cloud proxies, and determining a second point cloud feature extracted from the initial completed point cloud; The first point cloud feature and the second point cloud feature are compared to obtain a comparison feature, and the initial completed point cloud is adjusted based on the comparison feature to obtain the target point cloud data set.

9. An asset evaluation device, characterized in that: include: A receiving unit, configured to receive a point cloud data set of a target collateral, and extract N groups of point cloud clusters from the point cloud data set, where N is a positive integer; a first determining unit, configured to, for each group of point cloud clusters, determine a point cloud center of the point cloud cluster, extract local features of the point cloud cluster, and determine a point cloud proxy of the point cloud cluster based on the point cloud center and the local features, to obtain N point cloud proxies; An input unit, configured to input the N point cloud proxies into an encoder of a target model to obtain an encoding vector of each point cloud proxy, and input the N encoding vectors into a decoder of the target model to obtain a missing point proxy; A completion unit is used to complete the missing point clouds in the point cloud data set based on the missing point agent to obtain a target point cloud data set, and perform asset evaluation on the target collateral through the target point cloud data set to obtain an asset evaluation result.

10. A computer program product comprising computer instructions, characterized in that When the computer instructions are executed by a processor, the steps of the asset evaluation method according to any one of claims 1 to 8 are implemented.

Images