Methods, apparatus, equipment and media for financial collateral survey based on computer vision

Abstract

This application discloses a computer vision-based method, apparatus, equipment, and medium for surveying financial collateral, relating to the field of target recognition technology. The method includes: acquiring a target video of a target property as collateral and determining target keyframes from the video; determining a set of keyframes from the target keyframes that correspond to the target property layout and room type, thus obtaining a target keyframe album, and determining the target financial business template of the property as collateral; performing target recognition on the target keyframes in the target keyframe album based on the target financial business template and a preset real-time target detection algorithm to obtain target recognition results; generating a survey and analysis result of the property as collateral based on the target recognition result, the target keyframe album, and a preset evaluation strategy; and generating a target survey report of the property as collateral based on preset artificial intelligence technology, preset prompts, and the survey and analysis result. This application enables automated surveying of property as collateral.

Description

Technical Field

[0001] This invention relates to the field of target recognition technology, and in particular to a method, apparatus, equipment and medium for surveying financial collateral based on computer vision. Background Technology

[0002] With the rapid development of computer vision technology in the field of artificial intelligence, traditional image processing technology has evolved into an advanced stage capable of deep understanding and analysis of video content. In the financial sector, collateral surveying is a crucial step in risk management and loan approval processes, with its core objective being the accurate valuation and risk control of pledged assets.

[0003] Currently, in commercial banks' home renovation loan business, account managers are required to personally visit the client's renovation site to conduct on-site surveys, including data collection, photography, and video recording. They then manually write survey reports to verify the authenticity of the loan funds' intended use. Given the typically large loan amounts, banks need to disburse funds in stages according to the renovation progress. Account managers must make multiple site visits to manually compare the renovation progress and ensure the authenticity of the funds' intended use and the alignment of loan disbursement with construction progress. This business demands significant manpower from commercial banks, is inefficient and time-consuming for account managers, and suffers from the potential for inaccurate and subjective human judgment.

[0004] In conclusion, how to automate the surveying of financial collateral and improve the accuracy and objectivity of valuation are pressing technical problems that need to be solved. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a computer vision-based method, apparatus, equipment, and medium for surveying financial collateral, which can achieve automated surveying of financial collateral and improve the accuracy and objectivity of valuation. The specific solution is as follows:

[0006] Firstly, this application provides a computer vision-based method for surveying financial collateral, including:

[0007] Acquire the target video of the target house collateral, and decode the target video to determine several target keyframes of the target video;

[0008] From the plurality of target keyframes, determine the set of keyframes corresponding to the target house layout and preset room type of the target house collateral, so as to obtain the target keyframe album and determine the target financial business template corresponding to the target house collateral.

[0009] Based on the target financial business template and the preset real-time target detection algorithm, target recognition is performed on the target keyframes in the target keyframe image set to obtain the corresponding target recognition results;

[0010] Based on the target identification results, the target keyframe album, and the preset evaluation strategy, the survey and analysis results of the target house collateral are generated, and a target survey report of the target house collateral is generated based on preset artificial intelligence technology, preset prompt words, and the survey and analysis results.

[0011] Optionally, before acquiring the target video of the target house collateral, the process further includes:

[0012] Based on the target selection instruction, select the house type of the target house collateral to obtain a reference house layout diagram corresponding to the house type.

[0013] By using the target drag command, and dragging the room elements in the reference house layout diagram based on the orientation structure of each room in the target house collateral, the target house layout diagram corresponding to the target house collateral is generated.

[0014] Based on the target house layout map, a target survey route corresponding to the target house collateral is generated, so as to collect the target video of the target house collateral based on the target survey route.

[0015] Optionally, determining a plurality of target keyframes of the target video includes:

[0016] The image data obtained after decoding the target video is stored in a preset distributed storage system;

[0017] Several target keyframes of the target video are determined from the image data.

[0018] Optionally, storing the image data obtained after decoding the target video into a preset distributed storage system includes:

[0019] Configure the target decoding parameters corresponding to the target video; the target decoding parameters include the target frame rate and the target resolution.

[0020] The target video is decoded based on the target decoding parameters to obtain the initial image data corresponding to the target video, and the initial image data is stored in memory.

[0021] The initial image data is preprocessed according to a preset image preprocessing method to obtain target image data, and the target image data in memory is stored in the preset distributed storage system based on a preset data redundancy strategy.

[0022] Optionally, before decoding the target video based on the target decoding parameters, the method further includes:

[0023] Determine the total number of frames in the target video and set the corresponding target counter in the target decoding program;

[0024] Accordingly, the process of decoding the target video based on the target decoding parameters further includes:

[0025] The number of decoded frames corresponding to the target video is determined in real time, and the value of the target counter is updated in real time based on the number of decoded frames; there is a positive correlation between the number of decoded frames and the value of the target counter;

[0026] The ratio of the target counter value to the total number of frames in the target video is determined in real time to monitor the decoding progress of the target video;

[0027] Monitor the utilization of target resources during the decoding process of the target video, and identify error information based on the decoding progress and the utilization of target resources;

[0028] If the error message is detected, the error message is recorded to the target log file and processed based on a preset error correction mechanism.

[0029] Optionally, the step of performing image preprocessing on the initial image data according to a preset image preprocessing method to obtain the target image data includes:

[0030] The initial image data is preprocessed according to a preset image noise reduction method and a preset contrast enhancement method, and the size of the preprocessed initial image data is adjusted to the target size based on a preset image size condition to obtain the target image data, so as to determine a number of target keyframes of the target video from the target image data.

[0031] Optionally, determining the target keyframes of the target video from the target image data includes:

[0032] Perform a two-dimensional Fourier transform on each of the target image data, and determine the target high-frequency region in each transformed target image data;

[0033] Determine the energy percentage corresponding to each target high-frequency region, and determine the energy percentage that is lower than the preset energy percentage threshold as the target energy percentage;

[0034] The target image data corresponding to the target energy percentage is determined as a blurred image, and the blurred image is filtered out from the target image data;

[0035] The filtered target image data is determined to be a clear image, and based on the clear image, several target keyframes of the target video are determined from the target image data.

[0036] Optionally, determining the set of keyframes from the plurality of target keyframes corresponding to the target house layout map and preset room type of the target house collateral includes:

[0037] Based on the target house layout map and the preset room type, determine the survey room in the target survey route corresponding to the target house collateral, and match each target keyframe with each survey room to determine the target survey room corresponding to each target keyframe;

[0038] The set of keyframes corresponding to each target keyframe is determined based on the target survey room corresponding to each target keyframe.

[0039] Optionally, determining the set of keyframes from the plurality of target keyframes corresponding to the target house layout map and preset room type of the target house collateral includes:

[0040] Based on the preset real-time positioning and map building technology and the target house layout map, construct the target three-dimensional spatial model corresponding to the target house collateral;

[0041] A corresponding target three-dimensional rectangular coordinate system is established in the target three-dimensional spatial model; the target three-dimensional rectangular coordinate system is a spatial coordinate system constructed with the front of the entrance of the target house as the X-axis direction, the vertical upward direction as the positive Z-axis direction, and the Y-axis direction determined according to the right-hand rule;

[0042] Based on the target three-dimensional rectangular coordinate system, determine the target three-dimensional coordinate data corresponding to each target key frame, and based on the target three-dimensional coordinate data, the target house layout diagram, and the preset room type, determine the target survey room corresponding to each target key frame;

[0043] The set of keyframes corresponding to each target keyframe is determined based on the target survey room corresponding to each target keyframe.

[0044] Optionally, the step of obtaining corresponding target recognition results by performing target recognition on the target keyframes in the target keyframe image set based on the target financial business template and a preset real-time target detection algorithm includes:

[0045] The preset real-time target detection algorithm is used to divide each target keyframe in each target keyframe image set according to the preset grid size condition, so as to obtain each grid corresponding to the target keyframe; the grid is used to detect the target item in the target keyframe.

[0046] The preset real-time target detection algorithm is used to generate a preset number of initial bounding boxes corresponding to the grid, and the initial confidence level and the class probability corresponding to the initial bounding box are determined.

[0047] The target keyframe is identified based on the initial bounding box, the initial confidence level corresponding to the initial bounding box, the class probability corresponding to the initial bounding box, the preset nonmaximum suppression algorithm, and the target financial business template to obtain the target identification result.

[0048] Optionally, the step of performing target recognition on the target keyframe based on the initial bounding box, the initial confidence level corresponding to the initial bounding box, the class probability corresponding to the initial bounding box, a preset non-maximum suppression algorithm, and the target financial business template to obtain the target recognition result includes:

[0049] The target confidence level of the initial bounding box is determined based on the product of the initial confidence level corresponding to the initial bounding box and the class probability corresponding to the initial bounding box.

[0050] Based on the target confidence level, the preset nonmaximum suppression algorithm, and the target financial business template, target recognition is performed on the target keyframe to obtain the target recognition result.

[0051] Optionally, the step of performing target recognition on the target keyframe based on the target confidence level, the preset non-maximum suppression algorithm, and the target financial business template to obtain the target recognition result includes:

[0052] Based on the target confidence level, the initial bounding boxes are sorted to obtain a first bounding box list, and the target result set is initialized.

[0053] The initial bounding box with the highest target confidence is determined from the first bounding box list, and the initial bounding box with the highest target confidence is determined as the target bounding box;

[0054] Add the target bounding box to the target result set, and remove the target bounding box from the first bounding box list;

[0055] Determine the initial intersection-union ratio (IUU) between the target bounding box and each of the remaining initial bounding boxes in the first bounding box list, and determine the target IUU that is greater than a preset IUU threshold from the initial IUUU ratio;

[0056] Remove the initial bounding box corresponding to the target intersection-union ratio from the first bounding box list, and jump to the step of determining the initial bounding box with the highest confidence of the target as the target bounding box, until the first bounding box list is empty, and generate the target recognition result corresponding to the target keyframe based on the target financial business template and the target result set.

[0057] Optionally, generating the survey and analysis results of the target house collateral based on the target identification results, the target keyframe image set, and the preset evaluation strategy includes:

[0058] Generate a target survey summary of the target property collateral based on the target identification results;

[0059] Based on the target keyframe image set, the target identification results are deduplicated, and the value assessment result corresponding to the target house collateral is generated based on the deduplicated target identification results.

[0060] Based on the deduplicated target identification result and the target keyframe album, an integrity assessment result corresponding to the target house collateral is generated, and a risk assessment is performed on the target house collateral based on the deduplicated target identification result and preset risk assessment indicators to obtain the corresponding risk assessment result;

[0061] The survey analysis results of the target property collateral are generated based on the target survey summary, the value assessment results, the integrity assessment results, and the risk assessment results.

[0062] Optionally, the step of conducting a risk assessment on the target property collateral based on the deduplicated target identification result and preset risk assessment indicators to obtain the corresponding risk assessment result includes:

[0063] Based on the valuation results, the reasonableness of the valuation of the target house collateral is assessed to obtain a reasonable valuation assessment result. Based on the deduplicated target identification results, an abnormal risk assessment is performed on the target house collateral to obtain an abnormal risk assessment result.

[0064] The risk assessment result for the target property collateral is generated based on the valuation reasonableness assessment result, the integrity assessment result, the abnormal risk assessment result, and the overall risk assessment result for the target property collateral.

[0065] Optionally, generating the target survey report for the target property collateral based on preset artificial intelligence technology, preset prompts, and the survey analysis results includes:

[0066] The initial survey data corresponding to the survey analysis results is processed by data structuring based on a preset data structure, and the target survey data corresponding to the initial survey data after data structuring is determined according to a preset text format.

[0067] The target survey report for the target house collateral is generated based on preset generative artificial intelligence technology, preset prompts, and target survey data.

[0068] Secondly, this application provides a computer vision-based financial collateral surveying device, comprising:

[0069] The target keyframe determination module is used to acquire the target video of the target house collateral and decode the target video to determine several target keyframes of the target video.

[0070] The target keyframe set determination module is used to determine the set of keyframes corresponding to the target house layout and preset room type of the target house collateral from the plurality of target keyframes, so as to obtain the target keyframe set and determine the target financial business template corresponding to the target house collateral.

[0071] The target recognition result determination module is used to perform target recognition on the target keyframes in the target keyframe photo set based on the target financial business template and the preset real-time target detection algorithm to obtain the corresponding target recognition result.

[0072] The target survey report generation module is used to generate survey and analysis results of the target house collateral based on the target identification results, the target keyframe album and the preset evaluation strategy, and to generate a target survey report of the target house collateral based on preset artificial intelligence technology, preset prompt words and the survey and analysis results.

[0073] Thirdly, this application provides an electronic device, comprising:

[0074] Memory, used to store computer programs;

[0075] A processor is used to execute the computer program to implement the aforementioned computer vision-based financial collateral survey method.

[0076] Fourthly, this application provides a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned computer vision-based financial collateral survey method.

[0077] In this application, the target video of the target property collateral is first acquired and decoded to determine several target keyframes. Then, a set of keyframes corresponding to the target property layout and preset room types of the target property collateral is determined from these keyframes to obtain a target keyframe album, and a target financial business template corresponding to the target property collateral is determined. Subsequently, based on the target financial business template and a preset real-time target detection algorithm, target recognition is performed on the target keyframes in the target keyframe album to obtain corresponding target recognition results. Finally, based on the target recognition results, the target keyframe album, and a preset evaluation strategy, a survey and analysis result of the target property collateral is generated, and a target survey report of the target property collateral is generated based on preset artificial intelligence technology, preset prompts, and the survey and analysis results. As can be seen, this application first acquires and decodes the target video of the target property collateral to determine several target keyframes. Then, based on the target property layout and preset room types of the target property collateral, a set of keyframes corresponding to the target keyframes is determined to obtain a target keyframe album. Then, based on the target financial business template corresponding to the target property collateral and a preset real-time target detection algorithm, target keyframes are identified to obtain the corresponding target identification results. Finally, based on the target identification results, the target keyframe album, and the preset evaluation strategy, survey and analysis results are generated, and then a target survey report is generated using preset artificial intelligence technology, preset prompts, and the survey and analysis results. In this way, by acquiring the target video of the target property collateral and directly performing target identification on the target video, this application can significantly reduce human resource costs and improve business processing efficiency; by using the target financial business template and the preset real-time target detection algorithm, target identification can be automatically performed on the target video of the target property collateral; and by using preset artificial intelligence technology and preset prompts, a target survey report of the target property collateral can be automatically output based on the target identification results. Thus, this application can achieve automated surveying of target property collateral, improve the accuracy and objectivity of value assessment, and reduce operating costs. Attached Figure Description

[0078] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0079] Figure 1 A flowchart of a computer vision-based financial collateral survey method provided for this application;

[0080] Figure 2A flowchart illustrating the video capture process for a specific target property mortgage provided in this application;

[0081] Figure 3 A schematic diagram of the layout of a specific target property as collateral for this application;

[0082] Figure 4 This application provides a specific video capture illustration of a target property as collateral.

[0083] Figure 5 A flowchart for identifying a specific type of target mortgage provided in this application;

[0084] Figure 6 A schematic diagram illustrating the identification results of a specific target property as collateral provided in this application;

[0085] Figure 7 A schematic diagram of a computer vision-based financial collateral surveying device provided for this application;

[0086] Figure 8 This application provides a structural diagram of an electronic device. Detailed Implementation

[0087] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0088] With the rapid development of computer vision technology in the field of artificial intelligence, traditional image processing technology has evolved into an advanced stage capable of deep understanding and analysis of video content. In the financial sector, collateral inspection is a crucial link in risk management and loan approval processes, with its core objective being the accurate valuation and risk control of mortgaged assets. Currently, in commercial banks' home renovation loan business, account managers must personally visit the client's renovation site to conduct on-site inspections, including data collection, photography, and video recording, and manually write inspection reports to verify the authenticity of the loan funds' intended use. Given that loan amounts are typically large, banks need to disburse funds in stages according to the renovation progress, requiring account managers to make multiple site visits to manually compare the renovation progress and ensure the authenticity of the funds' intended use and the matching of loan disbursement progress with construction progress. This business requires a high level of human resources from commercial banks, is inefficient and time-consuming for account managers, and suffers from the problems of inaccurate and subjective human judgment. Therefore, this application provides a computer vision-based financial collateral inspection solution that can automate the inspection of financial collateral and improve the accuracy and objectivity of valuation.

[0089] See Figure 1 As shown in the figure, this invention discloses a computer vision-based method for surveying financial collateral, which may include:

[0090] Step S11: Acquire the target video of the target house collateral and decode the target video to determine several target keyframes of the target video.

[0091] In this embodiment, the customer can use a personal mobile device to independently complete the video capture of the target property as collateral and upload the captured video to the system server. See also Figure 2As shown, the specific operation process is as follows: First, the customer accesses the H5 website (HTML, HyperText Markup Language) address provided in this embodiment via a mobile device to authorize login and real-name authentication, ensuring the security and legality of the operation. After completing login authentication, the customer needs to complete the association with a specific financial business in the system interface, such as a commercial bank's home renovation loan business. Before obtaining the target video of the target house collateral, the process may further include: first, selecting the house type structure of the target house collateral based on the target selection instruction to obtain a reference house layout map corresponding to the house type structure; then, dragging the room elements in the reference house layout map based on the orientation structure of each room in the target house collateral using the target drag instruction to generate the target house layout map corresponding to the target house collateral; finally, generating the target survey route corresponding to the target house collateral based on the target house layout map, so as to collect the target video of the target house collateral based on the target survey route. See also... Figure 3 As shown, in the system interface, the customer first selects the floor plan of the target property as collateral, such as a three-bedroom, two-living-room house. The system then generates a corresponding reference floor plan. Afterward, the customer can drag and drop images on the interface to position the rooms in the reference floor plan, configuring the floor plan and orientation of the target property as collateral, and generating a detailed floor plan of the target property. See also... Figure 4 As shown, based on the target property's layout map and the collateral, the system can automatically generate a target survey route to guide the client in recording on-site video. During video recording, the system can provide real-time route prompts to ensure the comprehensiveness and accuracy of video capture.

[0092] It is understood that determining several target keyframes of the target video as described above may include: first, storing the image data obtained after decoding the target video into a preset distributed storage system; then, determining several target keyframes of the target video from the image data. Specifically, storing the image data obtained after decoding the target video into the preset distributed storage system may include: first, configuring target decoding parameters corresponding to the target video; the target decoding parameters include the target frame rate and target resolution; then, decoding the target video based on the target decoding parameters to obtain the initial image data corresponding to the target video, and storing the initial image data in memory; finally, performing image preprocessing on the initial image data according to a preset image preprocessing method to obtain target image data, and storing the target image data in memory into the preset distributed storage system based on a preset data redundancy strategy. In this way, storing the target image data in memory into the preset distributed storage system through a data redundancy strategy can solve the problem of data loss.

[0093] It should be noted that the above-mentioned image preprocessing of the initial image data according to the preset image preprocessing method to obtain the target image data may include: preprocessing the initial image data according to the preset image denoising method and the preset contrast enhancement method, and adjusting the size of the preprocessed initial image data to the target size based on the preset image size conditions to obtain the target image data, so as to determine several target keyframes of the target video from the target image data. Specifically, in order to reduce image noise, the mean value of pixels in the pixel neighborhood can be calculated to replace the current pixel value, thereby smoothing the noise; the median value in the pixel neighborhood can be used to replace the current pixel value, thereby removing impulse noise such as salt and pepper noise; a weighted average can be applied to the pixel neighborhood according to the Gaussian distribution, thereby effectively removing Gaussian noise; wavelet transform can be used to decompose the image into different frequency sub-bands, and thresholding of the high-frequency sub-bands can be used to remove noise while preserving important details of the image. To enhance image contrast, histogram transformation can be performed to make the gray-level distribution more uniform, thereby enhancing overall contrast. Histogram equalization can be performed in local areas to better adapt to the contrast requirements of different regions in the image, avoiding the problem of local over-enhancement or under-enhancement that may occur with global histogram equalization. In this embodiment, to reduce computational load and meet the image size constraints of the algorithm input, the target image size can be determined, and the size of the preprocessed initial image data can be adjusted to the target size to obtain the target image data.

[0094] Specifically, after video acquisition, the system automatically decodes the uploaded target video, converting the video file into an analyzable data format. Subsequently, the system efficiently stores the decoded image and audio data and performs a series of preprocessing operations, including but not limited to image denoising, contrast enhancement, and resizing, to optimize image data quality and lay the foundation for subsequent feature extraction and analysis. The logic for decoding and storing the target video is as follows:

[0095] A [Start] --> B [Client uses acquisition equipment to capture video]

[0096] B-->C [The acquisition device saves the video in a specific format]

[0097] C --> D [The customer uploads the video to the system server]

[0098] D-->E [The server receives the video and performs initial storage]

[0099] E-->F [The system decodes the video]

[0100] F --> G [The decoded image sequence and audio signal are stored in memory]

[0101] G --> H [Preprocess the image]

[0102] H-->I [Store the preprocessed image to a distributed storage system]

[0103] I-->J [Data processing complete; image data can be used for subsequent analysis]

[0104] J-->K [End]

[0105] Image preprocessing includes:

[0106] G-->G1 [Image Denoising]

[0107] G1->G2 [Image Contrast Enhancement]

[0108] G2 --> G3 [Image Resizing]

[0109] In this embodiment, to avoid errors caused by randomly determining keyframes, the image sharpness can first be determined to prevent blurry image data from being identified as keyframes. Specifically, the image sharpness can be determined using methods such as FFT (Fast Fourier Transform) frequency domain analysis and Laplacian (Laplace Operator) gradient methods.

[0110] In one specific implementation, in determining the image sharpness using FFT frequency domain analysis, a two-dimensional Fourier transform is first performed on each target image data to identify the target high-frequency regions in each transformed target image data. Then, the energy percentage corresponding to each target high-frequency region is determined, and the energy percentage below a preset energy percentage threshold is defined as the target energy percentage. Subsequently, the target image data corresponding to the target energy percentage is identified as a blurred image, and the blurred image is filtered out of the target image data. Finally, the filtered target image data is identified as a sharp image. It should be noted that the high-frequency regions of different image data need to be specifically set. For low-texture images, such as solid color areas, the radius of the high-frequency region can be reduced; for high-texture images, such as furniture details, the radius of the high-frequency region can be expanded. Since the high-frequency distribution of different image content, such as complex textures and large areas of solid color, varies significantly, to avoid the inability of a single preset threshold to adapt to the energy distribution differences in different scenes, image noise can first be estimated by spatial domain variance or frequency domain low-frequency energy, and the preset energy percentage threshold can be dynamically adjusted. Then, spatial domain gradient features and frequency domain energy features can be calculated in parallel, and the accuracy of the judgment can be improved through weighted fusion, such as SVM (Support Vector Machine) scoring. Furthermore, to avoid filtering images with locally sharp areas as a whole—such as frames in a video where the focus is on furniture but the background is blurred—images with slightly blurred overall appearance but whose high-frequency energy does not fall below a threshold are retained. This can be achieved by dividing the image into overlapping sub-blocks, such as 8×8 pixel windows, calculating the high-frequency energy independently for each sub-block, and filtering only images where the sub-block-level blur exceeds a certain proportion. This can be combined with edge detection algorithms, such as the Sobel operator, to calculate the average gradient magnitude of edge pixels, thus supplementing the insufficient high-frequency energy.

[0111] In another specific implementation, in the process of determining the image sharpness using the Laplacian gradient method, a Laplacian operator, such as a 3×3 convolution kernel, can be used to calculate the second derivative of the image, and then the absolute value or variance of the Laplacian response is calculated; the larger the absolute value or variance, the sharper the image.

[0112] It should be noted that after obtaining a clear image, several target keyframes of the target video can be determined from the target image data based on the clear image. In one specific implementation, target keyframes can be determined from the clear image based on time intervals. Specifically, the sampling time interval can be dynamically adjusted according to the rate of change of image content. When the image content changes rapidly, the sampling interval is shortened; when the image content changes slowly, the sampling interval is lengthened. The rate of change of image content can be determined by calculating the difference between adjacent frames, such as pixel difference, color histogram difference, etc. In another specific implementation, motion detection algorithms, such as background subtraction and frame difference methods, can be used to detect moving objects in the video. When a moving object is detected or the motion state changes significantly, the current image frame is taken as a keyframe. Specifically, the optical flow field between adjacent images can be calculated, and the degree of dynamic change in the image can be determined by the magnitude and direction of the optical flow vector. When the optical flow change exceeds a certain threshold, it indicates that an object is moving in the picture or the viewpoint has changed significantly, and the image frame can be retained as a keyframe. For example, when the photographer moves through different rooms or discovers new objects, the corresponding images may contain valuable information. Furthermore, it can detect shot transitions in the video. The first frame after a shot transition often marks the beginning of a new scene or perspective. Using this first frame as a keyframe can effectively segment different content segments. In a third specific implementation, stability can be assessed by calculating the degree of jitter in the image sequence. Motion estimation parameters of the image, such as translation, rotation, and scaling parameters, can be used to determine the degree of jitter. Images with severe jitter may be difficult to accurately identify due to unstable shooting and should be discarded. Images that are stable in shooting, clear, and without significant shaking should be retained as keyframes to ensure the accuracy of subsequent identification.

[0113] Step S12: Determine the set of keyframes corresponding to the target house layout and preset room type of the target house collateral from the plurality of target keyframes to obtain the target keyframe album, and determine the target financial business template corresponding to the target house collateral.

[0114] In this embodiment, features of the target keyframes are first extracted, converting the image data in the video into numerical feature vectors that can represent the content. These feature vectors can be used for subsequent target detection, recognition, and behavior analysis, forming the foundation for automated surveying. It is understood that this embodiment can pre-define multiple financial business templates, such as a construction progress survey tree. A financial business template refers to a series of predefined standard models based on the needs of different financial businesses. For example, in home renovation loan business, the financial business template may include renovation progress standards for different stages, the layout of key areas, etc.

[0115] See Figure 5As shown, to improve the accuracy of target recognition and reduce the computational burden of the model, target keyframes can be matched and reordered according to the target survey route and preset room types. Target keyframes that successfully match the route are assigned to the corresponding keyframe set, resulting in a target keyframe album. The survey route includes each surveyed room, such as the master bedroom, secondary bathroom #1, secondary bathroom #2, and entrance hall; each room type is represented as a keyframe album.

[0116] In one specific implementation, determining the set of keyframes corresponding to the target house layout map and preset room types of the target house collateral from the plurality of target keyframes may include: first, determining the survey rooms in the target survey route corresponding to the target house collateral based on the target house layout map and the preset room types, and matching each target keyframe with each survey room to determine the target survey room corresponding to each target keyframe; then, determining the set of keyframes corresponding to each target keyframe based on the target survey rooms corresponding to each target keyframe. Specifically, the target survey route can be stored in a structured data format, such as using a list or tree structure, clearly defining the order and hierarchy of each room, for example, [entrance hall, master bedroom, secondary bathroom #1, secondary bathroom #2]. A corresponding keyframe set is established for each room type, which can be stored using a dictionary or hash table, with the key being the room type name and the value being the container storing the keyframes. During the target video capture process, metadata can be added to each frame of the video using methods such as GPS (Global Positioning System) positioning, inertial sensors, or manual annotation. This metadata records the shooting location and room information corresponding to that frame. If GPS positioning is used, a mapping between geographic coordinates and rooms needs to be performed using a target building layout map. If manual annotation is used, the camera operator must simultaneously mark the room being filmed. Then, the metadata information of the keyframes is matched with the room information along the survey route. The shooting location or room annotation information of each keyframe is checked and compared with the room names along the survey route. If the names match, the keyframe is considered to have successfully matched the corresponding room. Subsequently, the successfully matched keyframes are rearranged according to the order of the rooms along the survey route. Keyframes corresponding to rooms that appear earlier are placed first, and keyframes corresponding to rooms that appear later are placed later. During the sorting process, sorting algorithms in data structures, such as bubble sort and quicksort, can be used to sort the keyframes based on the position of the room to which they belong along the survey route. Then, the reordered target keyframes are stored in the corresponding target keyframe albums according to the room type.

[0117] In another specific implementation, determining the set of keyframes corresponding to the target house layout map and preset room type of the target house collateral from the aforementioned target keyframes may include: first, constructing a target three-dimensional spatial model corresponding to the target house collateral based on preset real-time positioning and mapping technology and the target house layout map; then, establishing a corresponding target three-dimensional rectangular coordinate system in the target three-dimensional spatial model; the target three-dimensional rectangular coordinate system is a spatial coordinate system constructed with the front of the entrance of the target house collateral as the X-axis direction, the vertical upward direction as the positive Z-axis direction, and the Y-axis direction determined according to the right-hand rule; subsequently, determining the target three-dimensional coordinate data corresponding to each target keyframe based on the target three-dimensional rectangular coordinate system, and determining the target survey room corresponding to each target keyframe based on the target three-dimensional coordinate data, the target house layout map, and the preset room type; finally, determining the set of keyframes corresponding to each target keyframe based on the target survey room corresponding to each target keyframe. Specifically, the customer can choose a suitable shooting device, such as a camera with a depth sensor or a device that simultaneously has high-definition video recording and an IMU (Inertial Measurement Unit). Depth sensors can directly acquire depth information of the scene, while IMUs can record the device's motion and posture, such as rotation and acceleration, providing data support for SLAM (Simultaneous Localization and Mapping) algorithms. When the camera starts working, the SLAM system is activated and initialized. Using the image frames acquired by the device initially and the sensor data, the initial camera position and posture are determined as reference points for subsequent calculations. The SLAM algorithm extracts features from each frame. Through feature matching algorithms, feature points between adjacent frames are compared to find identical feature point pairs, thereby determining the relative motion relationship of the camera at different times. Then, combining the motion information provided by the IMU and the feature point matching results, the camera's trajectory in space, including translation and rotation, is calculated. Through continuous iterative optimization, the camera's pose at each moment is accurately estimated. Furthermore, based on the camera's pose and image information, a target 3D spatial model corresponding to the target house collateral is gradually constructed. Finally, a 3D Cartesian coordinate system is established with the house entrance as the origin. Typically, the positive X-axis is defined as the direction directly in front of the building entrance, the positive Z-axis is defined as vertically upward, and the Y-axis direction is determined using the right-hand rule. Then, the camera pose information corresponding to the target keyframe is determined, i.e., the camera's position and orientation in 3D space when the target keyframe was captured. Based on the camera pose information of the target keyframe, the XYZ offset of the target keyframe relative to the building entrance, i.e., the origin, is calculated. Using spatial geometric transformations and coordinate transformation formulas, the camera's position information is converted into 3D coordinate values ​​in the defined coordinate system, obtaining the target's 3D coordinate data corresponding to the target keyframe.By combining the target 3D spatial model of the target house collateral, the region where the target keyframe is located is determined, and then the room type corresponding to the target keyframe is determined. The target keyframes are then stored in the corresponding target keyframe photo set according to the room type.

[0118] It should be noted that each target keyframe album can be stored using a file system, with folders named according to room type, and the target keyframes stored as images in the corresponding folders; alternatively, a database can be used to store the target keyframe data and metadata, such as shooting time and room type, in a table structure.

[0119] Step S13: Based on the target financial business template and the preset real-time target detection algorithm, target recognition is performed on the target keyframes in the target keyframe image set to obtain the corresponding target recognition results.

[0120] In this embodiment, the above-mentioned target recognition result obtained by performing target recognition on the target keyframes in the target keyframe image set based on the target financial business template and the preset real-time target detection algorithm may include: firstly, using the preset real-time target detection algorithm and dividing each target keyframe in the target keyframe image set into grids based on a preset grid size condition, to obtain grids corresponding to the target keyframes; the grids are used to detect target items in the target keyframes; then, using the preset real-time target detection algorithm to generate a preset number of initial bounding boxes corresponding to the grids, and determining the initial confidence and class probability corresponding to the initial bounding boxes; finally, performing target recognition on the target keyframes based on the initial bounding boxes, the initial confidence, the class probability, the preset non-maximum suppression algorithm, and the target financial business template, to obtain the target recognition result. In one specific implementation, the system divides the target keyframes into grids one by one based on the YOLO (You Only Look Once) algorithm, dividing the target keyframes into 7×7 grids. Each grid is responsible for detecting target items whose center point falls within the grid. Each grid cell can predict B initial bounding boxes, each containing location information (x, y, w, h) and a confidence score. The system predicts initial bounding boxes, including their location (x, y, w, h) and confidence score. (x, y) represents the offset of the initial bounding box's center point relative to the top-left corner of the grid cell; (w, h) represents the width and height of the initial bounding box. The confidence score is measured by IOU (Intersection over Union), representing the probability that the target is contained within the initial bounding box and the accuracy of the bounding box. Simultaneously, the system predicts the probability distribution of the category to which each initial bounding box belongs, obtaining the category probability corresponding to the initial bounding box. In the current scenario of house mortgage surveying, categories include "walls," "ground," and "indoor facilities." The confidence score calculation formula is as follows:

[0121] ;

[0122] in, This indicates the probability that the target item is contained within the initial bounding box; This represents the intersection-union ratio (IoU) between the predicted initial bounding box and the actual bounding box.

[0123] Subsequently, to perform target recognition on the target keyframe by combining the confidence level and class probability of the initial bounding box, the specific item identified and its location are determined. The above-mentioned target recognition of the target keyframe based on the initial bounding box, the initial confidence level corresponding to the initial bounding box, the class probability corresponding to the initial bounding box, a preset non-maximum suppression algorithm, and the target financial business template to obtain the target recognition result may include: firstly, determining the target confidence level corresponding to the initial bounding box based on the product of the initial confidence level and the class probability corresponding to the initial bounding box; finally, performing target recognition on the target keyframe based on the target confidence level, the preset non-maximum suppression algorithm, and the target financial business template to obtain the target recognition result. That is, in this embodiment, the product of the initial confidence level and the class probability corresponding to the initial bounding box can be determined as the target confidence level corresponding to the initial bounding box. Subsequently, to ensure that each target item is detected only once, this embodiment can use non-maximum suppression to remove redundant bounding boxes in the initial bounding box, retaining the recognition result with the highest confidence level and low overlap with other bounding boxes.

[0124] It should be noted that the above-mentioned target recognition of the target keyframe based on the target confidence level, the preset non-maximum suppression algorithm, and the target financial business template to obtain the target recognition result may include: firstly, sorting each of the initial bounding boxes based on the target confidence level to obtain a first bounding box list, and initializing the target result set; then, determining the initial bounding box with the highest target confidence level from the first bounding box list, and determining the initial bounding box with the highest target confidence level as the target bounding box; then, adding the target bounding box to the target result set, and removing the target bounding box from the first bounding box list; subsequently, determining the initial intersection-union ratio (IU / R) between the target bounding box and each of the remaining initial bounding boxes in the first bounding box list, and determining the target IU / R greater than a preset IU / R threshold from the initial IU / R; finally, removing the initial bounding box corresponding to the target IU / R from the first bounding box list, and jumping to the step of determining the initial bounding box with the highest target confidence level as the target bounding box, until the first bounding box list is empty, and generating the target recognition result corresponding to the target keyframe based on the target financial business template and the target result set. Specifically, this embodiment can combine the location information of the initial bounding box, the house type corresponding to the target keyframe album, and the target confidence of the initial bounding box to finally determine the detected target item and its location. The specific process is as follows:

[0125] (1) Sort by target confidence of initial bounding boxes: Sort all initial bounding boxes from high to low according to the target confidence of the initial bounding boxes, and generate the first bounding box list based on the sorted initial bounding boxes, while initializing the target result set;

[0126] (2) Select the initial bounding box with the highest target confidence: take the initial bounding box with the highest confidence in the first bounding box list as the target bounding box, add the target bounding box to the target result set, and remove the target bounding box from the first bounding box list;

[0127] (3) Calculate the target IOU: Calculate the target IOU between the target bounding box and each of the remaining initial bounding boxes in the first bounding box list; where the formula for calculating IOU is: IOU = intersection area / union area;

[0128] (4) Remove overlapping initial bounding boxes: If the target IOU between the initial bounding box and the target bounding box is greater than the preset intersection-union ratio threshold, such as 0.5, then remove it from the first bounding box list;

[0129] (5) Repeat steps (2)-(4): until all initial bounding boxes in the first bounding box list have been processed and the target result set is obtained.

[0130] See Figure 6 As shown, in this embodiment, through the YOLO algorithm and the Mechanism of Predefined Survey Routes, the system can identify the target items and their states in the target keyframes.

[0131] Step S14: Generate the survey and analysis results of the target house collateral based on the target identification results, the target keyframe album and the preset evaluation strategy, and generate the target survey report of the target house collateral based on the preset artificial intelligence technology, preset prompt words and the survey and analysis results.

[0132] In this embodiment, the system can automatically assess the value of the target property collateral based on the target identification results and generate the survey and analysis results of the target property collateral. Then, the system can automatically generate a target survey report of the target property collateral based on the survey and analysis results of the target property collateral, according to the preset prompt words, i.e., the preset prompt, and using generative artificial intelligence technology.

[0133] As can be seen from the above, in this embodiment, the target video of the target house collateral is first acquired and decoded to determine several target keyframes. Then, a set of keyframes corresponding to the target house layout and preset room type of the target house collateral is determined from these target keyframes to obtain a target keyframe album, and a target financial business template corresponding to the target house collateral is determined. Subsequently, based on the target financial business template and a preset real-time target detection algorithm, target recognition is performed on the target keyframes in the target keyframe album to obtain corresponding target recognition results. Finally, based on the target recognition results, the target keyframe album, and a preset evaluation strategy, a survey and analysis result of the target house collateral is generated, and a target survey report of the target house collateral is generated based on preset artificial intelligence technology, preset prompts, and the survey and analysis result. As can be seen from the above, in this embodiment, the target video of the target house collateral is first acquired and decoded to determine several target keyframes. Then, based on the target house layout and preset room type of the target house collateral, a set of keyframes corresponding to the target keyframes is determined to obtain a target keyframe album. Then, based on the target financial business template corresponding to the target property collateral and a preset real-time target detection algorithm, target keyframes are identified to obtain the corresponding target identification results. Finally, based on the target identification results, the target keyframe album, and the preset evaluation strategy, a survey and analysis result is generated. Then, a target survey report is generated using preset artificial intelligence technology, preset prompts, and the survey and analysis result. In this way, by acquiring the target video of the target property collateral and directly performing target identification on the video, this embodiment can significantly reduce human resource costs and improve business processing efficiency; the target financial business template and preset real-time target detection algorithm can automatically identify the target video of the target property collateral; and the preset artificial intelligence technology and preset prompts can automatically output a target survey report of the target property collateral based on the target identification results. Thus, this embodiment can achieve automated surveying of target property collateral, improve the accuracy and objectivity of value assessment, and reduce operating costs.

[0134] As described in the previous embodiment, this application can acquire target video of the target house collateral, decode the target video, determine target keyframes, and then perform target recognition on the target keyframes to obtain the target recognition result. Next, this embodiment will elaborate on how to handle error information during the video decoding process.

[0135] In this embodiment, to promptly handle error information during the target video decoding process, before decoding the target video based on the target decoding parameters, the process may further include: determining the total number of frames in the target video and setting a corresponding target counter in the target decoding program. That is, during the decoding process, to monitor the decoding progress of the target video in real time, the total number of frames in the target video can be obtained when the target video decoding task starts, and this total number of frames can be used as a total indicator of the decoding progress.

[0136] Accordingly, the above-mentioned decoding process of the target video based on the target decoding parameters may further include: firstly, determining the number of decoding frames corresponding to the target video in real time, and updating the value of the target counter in real time based on the number of decoding frames; the number of decoding frames and the value of the target counter are positively correlated; then, determining the ratio of the value of the target counter to the total number of frames of the target video in real time to monitor the decoding progress of the target video; monitoring the utilization of target resources during the decoding process of the target video, and identifying error information based on the decoding progress and the utilization of target resources; if the error information is identified, recording the error information to the target log file, and processing the error information based on a preset error correction mechanism. Specifically, in this embodiment, a counter can be set in the decoding program, and the counter value is incremented by one each time a frame of the target video is successfully decoded. By calculating the ratio of the number of decoded frames to the total number of frames, the decoding progress can be displayed intuitively, for example, as a percentage. At the same time, a visual interface can be used to provide real-time feedback on the decoding progress to the system administrator or relevant operators so as to understand the progress of the decoding task at any time. If the decoding progress stalls for an extended period, such as exceeding a preset time threshold (e.g., no change after 10 minutes), the system automatically triggers an early warning mechanism, sending alerts to relevant personnel and indicating a potential decoding blockage. Furthermore, the system can monitor target resource utilization during the video decoding process. In one implementation, CPU resources can be monitored using metrics such as CPU utilization and the number of cores used. If CPU utilization remains consistently high (e.g., exceeding 80%), it may lead to decreased decoding efficiency or even slow system response. In this case, the system can implement dynamic adjustment strategies, such as lowering the priority of the decoding task or pausing other non-critical tasks to free up CPU resources and ensure smooth decoding. In another implementation, parameters such as the memory size and memory usage growth rate of the decoding process can be monitored in real time. When memory usage reaches a preset threshold, such as 80% of physical memory, the system can initiate a memory optimization mechanism to clean up invalid cached data and free up memory space. If memory remains scarce, some other non-essential processes can be terminated to ensure sufficient memory resources are available for the decoding task. In a third implementation, metrics such as the frequency and volume of disk read / write operations during the decoding process can be monitored in real time.

[0137] It should be noted that when the system identifies an error message, it can record key data such as the time of the error, the specific error message, and the current decoding progress, and store the error message in the corresponding log file. Simultaneously, the system can handle the error message based on a preset error correction mechanism. In one specific implementation, if the current error message indicates a decoding anomaly caused by incomplete data transmission due to network fluctuations, the system can attempt automatic re-decoding. Before re-decoding, an integrity check can be performed on the target video to ensure it is not further damaged. During re-decoding, decoding parameters can be adjusted, such as reducing the decoding frame rate or decreasing the resolution, to improve the decoding success rate. In another specific implementation, when encountering an undecoding video frame, frame interpolation technology can be used. Based on the information from preceding and following frames, an approximate image frame is generated using an algorithm, skipping the erroneous frame and continuing decoding to ensure the continuity of video decoding. Simultaneously, the skipped video frames are marked for subsequent manual inspection or special processing.

[0138] As can be seen from the above, this embodiment determines the decoding progress of the target video in real time and monitors the utilization of target resources during the decoding process. Based on the decoding progress and resource utilization, it identifies error information and processes it using a preset error correction mechanism. In this way, this embodiment can ensure the data quality of keyframes, improve the accuracy of keyframe recognition, and guarantee the stable operation of the entire process from video decoding to keyframe recognition, avoiding the accumulation of underlying problems that could affect the final recognition result.

[0139] As can be seen from the above embodiments, this application can obtain target videos of target property collateral and directly identify targets in the target keyframes corresponding to the target videos using a target financial business template and a preset real-time target detection algorithm to obtain target identification results. Next, this embodiment will elaborate on how to determine the target survey report of the target property collateral.

[0140] In this embodiment, the system can utilize pre-collected external data sources such as e-commerce platform data, industry association data, professional valuation agency data, and home furnishing market research reports, combined with a comprehensive evaluation model based on the market comparison method, to determine the completeness score and valuation of the target property collateral. Subsequently, based on the target identification results, target keyframe album, and preset evaluation strategies, a survey and analysis result of the target property collateral can be generated. Specifically, this can include: first, generating a target survey summary of the target property collateral based on the target identification results; then, deduplicating the target identification results based on the target keyframe album, and generating a corresponding value assessment result for the target property collateral based on the deduplicated target identification results; subsequently, generating a completeness assessment result for the target property collateral based on the deduplicated target identification results and the target keyframe album, and performing a risk assessment on the target property collateral based on the deduplicated target identification results and preset risk assessment indicators to obtain a corresponding risk assessment result; finally, generating the survey and analysis result of the target property collateral based on the target survey summary, the value assessment result, the completeness assessment result, and the risk assessment result. Specifically, the system first generates a basic information summary of the survey and analysis results, including (1) report number; (2) collateral type, such as: home renovation, commercial equipment, etc.; (3) customer name, such as: customer name, company name, etc.; (4) business type, such as: home renovation loan, equipment mortgage loan, etc.; (5) survey date; (6) survey personnel, such as: customer self-service, customer manager, or others. In this embodiment, multiple keyframes may be selected for the same room, which may lead to duplicate recognition. Therefore, the identified items in the same target keyframe image set can be deduplicated to obtain the deduplicated target recognition results. The system summarizes and sorts the surveyed and identified items according to the target keyframe image set to generate value assessment results. The system generates a master table and multiple sub-tables, as shown in Table 1. A sub-table refers to a set of keyframe images; the master table refers to a set of multiple keyframe images.

[0141] Table 1 Valuation Table

[0142]

[0143] In addition, as shown in Table 2, the system will also conduct an integrity assessment of the overall survey of the target property collateral and generate an integrity assessment result corresponding to the target property collateral, with a completeness score of 100.

[0144] Table 2 Integrity Assessment Table

[0145]

[0146] Subsequently, the system can perform a risk assessment on the target property collateral based on the deduplicated target identification results and preset risk assessment indicators, obtaining the corresponding risk assessment results. The specific process may include: first, assessing the reasonableness of the valuation of the target property collateral based on the value assessment results to obtain a reasonableness valuation result; and then performing an anomaly risk assessment on the target property collateral based on the deduplicated target identification results to obtain an anomaly risk assessment result. Next, based on the reasonableness valuation result, the completeness assessment result, the anomaly risk assessment result, and the overall risk assessment result corresponding to the target property collateral, a risk assessment result corresponding to the target property collateral is generated. Specifically, the preset risk assessment indicators include: reasonableness valuation, completeness, anomaly risk, and overall risk assessment. That is, the system can determine the reasonableness valuation result of the target property collateral based on the combined value assessment results and perform anomaly risk assessment on the target property collateral. Finally, as shown in Table 3, the system can generate the risk assessment result corresponding to the target property collateral based on the reasonableness valuation result, completeness assessment result, anomaly risk assessment result, and overall risk assessment result.

[0147] Table 3 Risk Assessment Table

[0148]

[0149] Finally, as shown in Table 4, the system will automatically generate the survey and analysis results of the target property collateral by combining the above-mentioned valuation results, integrity evaluation results, risk comprehensive assessment results, and target survey summary of the target property collateral.

[0150] Table 4 Survey and Analysis Table

[0151]

[0152] It should be noted that generating a target survey report for a target property mortgage based on preset artificial intelligence technology, preset prompts, and survey analysis results can include: first, performing data structuring on the initial survey data corresponding to the survey analysis results based on a preset data structure, and determining the target survey data corresponding to the initial survey data after data structuring according to a preset text format; then, generating the target survey report for the target property mortgage based on preset generative artificial intelligence technology, the preset prompts, and the target survey data. Specifically, the system first performs data structuring on the initial survey data corresponding to the survey analysis results, converting the initial survey data into key-value pairs or a table format. Then, based on preset generative artificial intelligence technology, a generative model is determined, and the initial survey data after data structuring is converted into a text format that the generative model can understand to obtain the target survey data, for example, converting JSON data into a natural language description. Using the generative model, combined with the target survey data and the preset Prompt, i.e., the preset prompts, a target survey report for the target property mortgage is generated. The preset Prompt includes:

[0153] ① Basic information;

[0154] ② Collateral valuation and assessment;

[0155] ③ Integrity assessment;

[0156] ④ Abnormal situations;

[0157] ⑤ Comprehensive risk assessment;

[0158] ⑥ Conclusions and Recommendations;

[0159] ⑦ Review information.

[0160] The system can call the generative model via API (Application Programming Interface), inputting a preset prompt as a request parameter. Finally, the generated initial survey report needs further formatting to ensure it meets the requirements of financial institutions. The system extracts key information from the generated initial survey report text and converts it into tables or paragraphs. For example, regular expressions or natural language processing tools can be used to convert the extracted key information from the generated initial survey report text into tabular data. Then, the extracted content is formatted according to a preset format, such as PDF or Word. Specifically, Python's ReportLab can be used to generate PDF or the python-docx library can be used to generate Word to complete the formatting, resulting in the target survey report for the target property mortgage. Partial code examples are shown below:

[0161] # Create a Word document

[0162] doc = Document()

[0163] doc.add_heading('Financial Collateral Survey and Assessment Report', level=1)

[0164] # Add basic information

[0165] doc.add_heading('I. Basic Information', level=2)

[0166] doc.add_paragraph('Report No.: 20250120-001')

[0167] doc.add_paragraph('Collection Type: Home Renovation')

[0168] doc.add_paragraph('Customer Name: Zhang San')

[0169] doc.add_paragraph('Business Type: Home Renovation Loan')

[0170] doc.add_paragraph('Survey Date: 2025-01-20')

[0171] doc.add_paragraph('Surveyors: Automatically generated by the system')

[0172] # Add collateral valuation assessment

[0173] doc.add_heading('II. Collateral Valuation and Assessment', level=2)

[0174] table = doc.add_table(rows=1, cols=5)

[0175] table.cell(0, 0).text = 'Project'

[0176] table.cell(0, 1).text = 'Description'

[0177] table.cell(0, 2).text = 'Quantity'

[0178] table.cell(0, 3).text = 'Status'

[0179] table.cell(0, 4).text = 'Estimated Value (Unit: Yuan)'

[0180] # Add data row

[0181] data = [

[0182] ["Wall decoration", "Completed", "100 square meters", "Completed", "5000"],

[0183] ["Floor renovation", "Completed", "80 square meters", "Completed", "8000"],

[0184] ["Indoor Facilities", "5 Lights, 3 Sets of Sanitary Ware", "Completed", "Completed", "3000"],

[0185] ["Total Valuation", "", "", "", "16000"] ]

[0187] for row_data in data:

[0188] row = table.add_row().cells

[0189] for i, cell_text in enumerate(row_data):

[0190] row[i].text = cell_text

[0191] # Save document

[0192] doc.save('Survey and Assessment Report.docx')

[0193] As can be seen from the above, this embodiment can generate survey and analysis results of the target property collateral based on the target recognition results of the target keyframes in the target video of the target property and a preset evaluation strategy. Then, based on preset artificial intelligence technology, preset prompts, and the survey and analysis results, a target survey report of the target property collateral is generated. In this way, through generative artificial intelligence technology, this embodiment can output a target survey report of the target property collateral, automating the survey of financial collateral, significantly reducing the human resource investment of financial institutions, improving business processing efficiency, and reducing financial risks.

[0194] Accordingly, see Figure 7 As shown in the figure, this application embodiment also provides a computer vision-based financial collateral surveying device, which may include:

[0195] The target keyframe determination module 11 is used to acquire the target video of the target house collateral and decode the target video to determine several target keyframes of the target video.

[0196] The target keyframe set determination module 12 is used to determine the set of keyframes corresponding to the target house layout map and preset room type of the target house collateral from the plurality of target keyframes, so as to obtain the target keyframe set and determine the target financial business template corresponding to the target house collateral.

[0197] The target recognition result determination module 13 is used to perform target recognition on the target key frames in the target key frame photo set based on the target financial business template and the preset real-time target detection algorithm to obtain the corresponding target recognition result;

[0198] The target survey report generation module 14 is used to generate the survey and analysis results of the target house collateral based on the target identification results, the target keyframe album and the preset evaluation strategy, and to generate the target survey report of the target house collateral based on the preset artificial intelligence technology, preset prompt words and the survey and analysis results.

[0199] As can be seen from the above, this application first acquires the target video of the target house collateral and decodes the target video to determine several target keyframes. Then, it determines a set of keyframes from these target keyframes that correspond to the target house layout and preset room types of the target house collateral, thus obtaining a target keyframe album, and determines the target financial business template corresponding to the target house collateral. Subsequently, based on the target financial business template and a preset real-time target detection algorithm, it performs target recognition on the target keyframes in the target keyframe album to obtain corresponding target recognition results. Finally, it generates a survey and analysis result of the target house collateral based on the target recognition result, the target keyframe album, and a preset evaluation strategy, and generates a target survey report of the target house collateral based on preset artificial intelligence technology, preset prompts, and the survey and analysis results. As can be seen from the above, this application first acquires the target video of the target house collateral and decodes the video to determine several target keyframes. Then, based on the target house layout and preset room types of the target house collateral, it determines the set of keyframes corresponding to the target keyframes, thus obtaining a target keyframe album. Then, based on the target financial business template corresponding to the target property collateral and a preset real-time target detection algorithm, target keyframes are identified to obtain the corresponding target identification results. Finally, based on the target identification results, the target keyframe album, and the preset evaluation strategy, survey and analysis results are generated, and then a target survey report is generated using preset artificial intelligence technology, preset prompts, and the survey and analysis results. In this way, by acquiring the target video of the target property collateral and directly performing target identification on the target video, this application can significantly reduce human resource costs and improve business processing efficiency; by using the target financial business template and the preset real-time target detection algorithm, target identification can be automatically performed on the target video of the target property collateral; and by using preset artificial intelligence technology and preset prompts, a target survey report of the target property collateral can be automatically output based on the target identification results. Thus, this application can achieve automated surveying of target property collateral, improve the accuracy and objectivity of value assessment, and reduce operating costs.

[0200] In some specific embodiments, the computer vision-based financial collateral surveying device may further include:

[0201] The reference house layout diagram determination module is used to select the house type structure of the target house collateral based on the target selection instruction, so as to obtain a reference house layout diagram corresponding to the house type structure.

[0202] The target house layout generation module is used to drag room elements in the reference house layout diagram based on the orientation structure of each room in the target house collateral by using a target drag command, so as to generate the target house layout diagram corresponding to the target house collateral.

[0203] The target survey route generation module is used to generate a target survey route corresponding to the target house collateral based on the target house layout map, so as to collect the target video of the target house collateral based on the target survey route.

[0204] In some specific embodiments, the target keyframe determination module 11 may include:

[0205] The image data storage submodule is used to store the image data obtained after decoding the target video into a preset distributed storage system;

[0206] The target keyframe determination submodule is used to determine a number of target keyframes of the target video from the image data.

[0207] In some specific embodiments, the image data storage submodule may include:

[0208] A decoding parameter configuration unit is used to configure the target decoding parameters corresponding to the target video; the target decoding parameters include the target frame rate and the target resolution.

[0209] The target video decoding unit is used to decode the target video based on the target decoding parameters to obtain the initial image data corresponding to the target video, and store the initial image data in memory;

[0210] An image data storage unit is used to perform image preprocessing on the initial image data according to a preset image preprocessing method to obtain target image data, and to store the target image data in memory to the preset distributed storage system based on a preset data redundancy strategy.

[0211] In some specific embodiments, the computer vision-based financial collateral surveying device may further include:

[0212] The total frame count determination module is used to determine the total number of frames of the target video and set the corresponding target counter in the target decoding program;

[0213] Accordingly, the target video decoding unit is specifically used to determine the number of decoding frames corresponding to the target video in real time, and update the value of the target counter in real time based on the number of decoding frames; there is a positive correlation between the number of decoding frames and the value of the target counter; the ratio of the value of the target counter to the total number of frames of the target video is determined in real time to monitor the decoding progress of the target video; the target resource utilization is monitored during the decoding process of the target video, and error information is identified based on the decoding progress and the target resource utilization; if the error information is identified, the error information is recorded to the target log file, and the error information is processed based on a preset error correction mechanism.

[0214] In some specific embodiments, the image data storage unit is specifically used to perform image preprocessing on the initial image data according to a preset image noise reduction method and a preset contrast enhancement method, and adjust the size of the preprocessed initial image data to a target size based on a preset image size condition to obtain the target image data, so as to determine a number of target keyframes of the target video from the target image data.

[0215] In some specific embodiments, the image data storage unit is specifically used to perform a two-dimensional Fourier transform on each of the target image data and determine the target high-frequency region in each transformed target image data; determine the energy percentage corresponding to each target high-frequency region, and determine the energy percentage below a preset energy percentage threshold as the target energy percentage; determine the target image data corresponding to the target energy percentage as a blurred image, and filter the blurred image from the target image data; determine the filtered target image data as a clear image, and determine several target keyframes of the target video from the target image data based on the clear image.

[0216] In some specific embodiments, the target keyframe set determination module 12 may include:

[0217] The first target survey room determination unit is used to determine the survey room in the target survey route corresponding to the target house collateral based on the target house layout map and the preset room type, and to match each target key frame with each survey room to determine the target survey room corresponding to each target key frame;

[0218] The first target keyframe set determination unit is used to determine the set of keyframes corresponding to each target keyframe based on the target survey room corresponding to each target keyframe.

[0219] In some specific embodiments, the target keyframe set determination module 12 may include:

[0220] The target 3D spatial model construction unit is used to construct the target 3D spatial model corresponding to the target house collateral based on the preset real-time positioning and map construction technology and the target house layout map;

[0221] The target three-dimensional rectangular coordinate system establishment unit is used to establish a corresponding target three-dimensional rectangular coordinate system in the target three-dimensional spatial model; the target three-dimensional rectangular coordinate system is a spatial coordinate system constructed with the front of the entrance of the target house as the X-axis direction, the vertical upward direction as the positive Z-axis direction, and the Y-axis direction determined according to the right-hand rule;

[0222] The second target survey room determination unit is used to determine the target three-dimensional coordinate data corresponding to each target key frame based on the target three-dimensional rectangular coordinate system, and to determine the target survey room corresponding to each target key frame based on the target three-dimensional coordinate data, the target house layout diagram and the preset room type;

[0223] The second target keyframe set determination unit is used to determine the set of keyframes corresponding to each target keyframe based on the target survey room corresponding to each target keyframe.

[0224] In some specific embodiments, the target recognition result determination module 13 may include:

[0225] The target keyframe segmentation submodule is used to divide each target keyframe in each target keyframe image set using the preset real-time target detection algorithm and based on the preset grid size condition, so as to obtain each grid corresponding to the target keyframe; the grid is used to detect target items in the target keyframe;

[0226] The initial bounding box generation submodule is used to generate a preset number of initial bounding boxes corresponding to the grid using the preset real-time target detection algorithm, and to determine the initial confidence level and the class probability corresponding to the initial bounding box.

[0227] The target recognition result determination submodule is used to perform target recognition on the target keyframe based on the initial bounding box, the initial confidence level corresponding to the initial bounding box, the class probability corresponding to the initial bounding box, the preset nonmaximum suppression algorithm, and the target financial business template, so as to obtain the target recognition result.

[0228] In some specific implementations, the target recognition result determination submodule may include:

[0229] The target confidence determination unit is used to determine the target confidence level corresponding to the initial bounding box based on the product between the initial confidence level corresponding to the initial bounding box and the class probability corresponding to the initial bounding box;

[0230] The target recognition result determination unit is used to perform target recognition on the target keyframe based on the target confidence level, the preset nonmaximum suppression algorithm and the target financial business template, so as to obtain the target recognition result.

[0231] In some specific embodiments, the target recognition result determination unit is specifically used to sort each of the initial bounding boxes based on the target confidence level to obtain a first bounding box list and initialize a target result set; determine the initial bounding box with the highest target confidence level from the first bounding box list and determine the initial bounding box with the highest target confidence level as the target bounding box; add the target bounding box to the target result set and remove the target bounding box from the first bounding box list; determine the initial intersection-union ratio (IU / R) between the target bounding box and each of the remaining initial bounding boxes in the first bounding box list, and determine the target IU / R greater than a preset IU / R threshold from the initial IU / R; remove the initial bounding box corresponding to the target IU / R from the first bounding box list, and jump to the step of determining the initial bounding box with the highest target confidence level as the target bounding box, until the first bounding box list is empty, and generate the target recognition result corresponding to the target keyframe based on the target financial business template and the target result set.

[0232] In some specific embodiments, the target survey report generation module 14 may include:

[0233] The target survey summary generation submodule is used to generate a target survey summary of the target house collateral based on the target identification result.

[0234] The value assessment result generation submodule is used to perform deduplication processing on the target identification result based on the target keyframe image set, and generate the value assessment result corresponding to the target house collateral based on the deduplicated target identification result;

[0235] The risk assessment result determination submodule is used to generate an integrity assessment result corresponding to the target house collateral based on the deduplicated target identification result and the target keyframe album, and to conduct a risk assessment on the target house collateral based on the deduplicated target identification result and preset risk assessment indicators to obtain the corresponding risk assessment result;

[0236] The survey and analysis results generation submodule is used to generate the survey and analysis results of the target property collateral based on the target survey summary, the value assessment results, the integrity assessment results, and the risk assessment results.

[0237] In some specific implementations, the risk assessment result determination submodule may include:

[0238] An abnormal risk assessment result determination unit is used to assess the valuation reasonableness of the target house collateral based on the valuation result to obtain a valuation reasonableness assessment result, and to conduct an abnormal risk assessment of the target house collateral based on the deduplicated target identification result to obtain an abnormal risk assessment result.

[0239] The risk assessment result determination unit is used to generate the risk assessment result corresponding to the target house collateral based on the valuation reasonableness assessment result, the integrity assessment result, the abnormal risk assessment result, and the overall risk assessment result corresponding to the target house collateral.

[0240] In some specific embodiments, the target survey report generation module 14 may include:

[0241] The target survey data determination unit is used to perform data structuring processing on the initial survey data corresponding to the survey analysis results based on a preset data structure, and to determine the target survey data corresponding to the initial survey data after data structuring processing according to a preset text format.

[0242] The target survey report generation unit is used to generate the target survey report of the target house collateral based on preset generative artificial intelligence technology, preset prompt words and target survey data.

[0243] Furthermore, embodiments of this application also disclose an electronic device, Figure 8 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the computer vision-based financial collateral survey method disclosed in any of the foregoing embodiments. Furthermore, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0244] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0245] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0246] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the computer vision-based financial collateral survey method executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0247] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned computer vision-based financial collateral survey method. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0248] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0249] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0250] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0251] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0252] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A computer vision-based method for surveying financial collateral, characterized in that, include: Acquire the target video of the target house collateral, and decode the target video to determine several target keyframes of the target video; From the plurality of target keyframes, determine the set of keyframes corresponding to the target house layout and preset room type of the target house collateral, so as to obtain the target keyframe album and determine the target financial business template corresponding to the target house collateral. Based on the target financial business template and the preset real-time target detection algorithm, target recognition is performed on the target keyframes in the target keyframe image set to obtain the corresponding target recognition results; Based on the target identification results, the target keyframe album, and the preset evaluation strategy, the survey and analysis results of the target house collateral are generated, and a target survey report of the target house collateral is generated based on preset artificial intelligence technology, preset prompt words, and the survey and analysis results. The step of generating the survey and analysis results of the target house collateral based on the target identification results, the target keyframe image set, and the preset evaluation strategy includes: Generate a target survey summary of the target property collateral based on the target identification results; Based on the target keyframe image set, the target identification results are deduplicated, and the value assessment result corresponding to the target house collateral is generated based on the deduplicated target identification results. Based on the deduplicated target identification result and the target keyframe album, an integrity assessment result corresponding to the target house collateral is generated, and a risk assessment is performed on the target house collateral based on the deduplicated target identification result and preset risk assessment indicators to obtain the corresponding risk assessment result; The survey analysis results of the target property collateral are generated based on the target survey summary, the value assessment results, the integrity assessment results, and the risk assessment results. The step of conducting a risk assessment of the target property collateral based on the deduplicated target identification result and preset risk assessment indicators to obtain the corresponding risk assessment result includes: Based on the valuation results, the reasonableness of the valuation of the target house collateral is assessed to obtain a reasonable valuation assessment result. Based on the deduplicated target identification results, an abnormal risk assessment is performed on the target house collateral to obtain an abnormal risk assessment result. The risk assessment result for the target property collateral is generated based on the valuation reasonableness assessment result, the integrity assessment result, the abnormal risk assessment result, and the overall risk assessment result for the target property collateral. The generation of the target survey report for the target property collateral based on preset artificial intelligence technology, preset prompts, and the survey analysis results includes: The initial survey data corresponding to the survey analysis results is processed by data structuring based on a preset data structure, and the target survey data corresponding to the initial survey data after data structuring is determined according to a preset text format. The target survey report for the target house collateral is generated based on preset generative artificial intelligence technology, preset prompts, and target survey data.

2. The computer vision-based financial collateral surveying method according to claim 1, characterized in that, Before acquiring the target video of the target house collateral, the process also includes: Based on the target selection instruction, select the house type of the target house collateral to obtain a reference house layout diagram corresponding to the house type. By using the target drag command, and dragging the room elements in the reference house layout diagram based on the orientation structure of each room in the target house collateral, the target house layout diagram corresponding to the target house collateral is generated. Based on the target house layout map, a target survey route corresponding to the target house collateral is generated, so as to collect the target video of the target house collateral based on the target survey route.

3. The computer vision-based financial collateral surveying method according to claim 1, characterized in that, The determination of several target keyframes of the target video includes: The image data obtained after decoding the target video is stored in a preset distributed storage system; Several target keyframes of the target video are determined from the image data.

4. The computer vision-based financial collateral surveying method according to claim 3, characterized in that, The step of storing the image data obtained after decoding the target video into a preset distributed storage system includes: Configure the target decoding parameters corresponding to the target video; the target decoding parameters include the target frame rate and the target resolution. The target video is decoded based on the target decoding parameters to obtain the initial image data corresponding to the target video, and the initial image data is stored in memory. The initial image data is preprocessed according to a preset image preprocessing method to obtain target image data, and the target image data in memory is stored in the preset distributed storage system based on a preset data redundancy strategy.

5. The computer vision-based financial collateral surveying method according to claim 4, characterized in that, Before decoding the target video based on the target decoding parameters, the method further includes: Determine the total number of frames in the target video and set the corresponding target counter in the target decoding program; Accordingly, the process of decoding the target video based on the target decoding parameters further includes: The number of decoded frames corresponding to the target video is determined in real time, and the value of the target counter is updated in real time based on the number of decoded frames; there is a positive correlation between the number of decoded frames and the value of the target counter; The ratio of the target counter value to the total number of frames in the target video is determined in real time to monitor the decoding progress of the target video; Monitor the utilization of target resources during the decoding process of the target video, and identify error information based on the decoding progress and the utilization of target resources; If the error message is detected, the error message is recorded to the target log file and processed based on a preset error correction mechanism.

6. The computer vision-based financial collateral surveying method according to claim 4, characterized in that, The step of preprocessing the initial image data according to a preset image preprocessing method to obtain the target image data includes: The initial image data is preprocessed according to a preset image noise reduction method and a preset contrast enhancement method, and the size of the preprocessed initial image data is adjusted to the target size based on a preset image size condition to obtain the target image data, so as to determine a number of target keyframes of the target video from the target image data.

7. The computer vision-based financial collateral surveying method according to claim 6, characterized in that, The step of determining a plurality of target keyframes of the target video from the target image data includes: Perform a two-dimensional Fourier transform on each of the target image data, and determine the target high-frequency region in each transformed target image data; Determine the energy percentage corresponding to each target high-frequency region, and determine the energy percentage that is lower than the preset energy percentage threshold as the target energy percentage; The target image data corresponding to the target energy percentage is determined as a blurred image, and the blurred image is filtered out from the target image data; The filtered target image data is determined to be a clear image, and based on the clear image, several target keyframes of the target video are determined from the target image data.

8. The computer vision-based financial collateral surveying method according to claim 1, characterized in that, The step of determining the set of keyframes from the plurality of target keyframes that correspond to the target house layout and preset room type of the target house collateral includes: Based on the target house layout map and the preset room type, determine the survey room in the target survey route corresponding to the target house collateral, and match each target keyframe with each survey room to determine the target survey room corresponding to each target keyframe; The set of keyframes corresponding to each target keyframe is determined based on the target survey room corresponding to each target keyframe.

9. The computer vision-based financial collateral surveying method according to claim 1, characterized in that, The step of determining the set of keyframes from the plurality of target keyframes that correspond to the target house layout and preset room type of the target house collateral includes: Based on the preset real-time positioning and map building technology and the target house layout map, construct the target three-dimensional spatial model corresponding to the target house collateral; A corresponding target three-dimensional rectangular coordinate system is established in the target three-dimensional spatial model; the target three-dimensional rectangular coordinate system is a spatial coordinate system constructed with the front of the entrance of the target house as the X-axis direction, the vertical upward direction as the positive Z-axis direction, and the Y-axis direction determined according to the right-hand rule; Based on the target three-dimensional rectangular coordinate system, determine the target three-dimensional coordinate data corresponding to each target key frame, and based on the target three-dimensional coordinate data, the target house layout diagram, and the preset room type, determine the target survey room corresponding to each target key frame; The set of keyframes corresponding to each target keyframe is determined based on the target survey room corresponding to each target keyframe.

10. The computer vision-based financial collateral surveying method according to claim 1, characterized in that, The step of performing target recognition on the target keyframes in the target keyframe set based on the target financial business template and a preset real-time target detection algorithm to obtain the corresponding target recognition results includes: The preset real-time target detection algorithm is used to divide each target keyframe in each target keyframe image set according to the preset grid size condition, so as to obtain each grid corresponding to the target keyframe; the grid is used to detect the target item in the target keyframe. The preset real-time target detection algorithm is used to generate a preset number of initial bounding boxes corresponding to the grid, and the initial confidence level and the class probability corresponding to the initial bounding box are determined. The target keyframe is identified based on the initial bounding box, the initial confidence level corresponding to the initial bounding box, the class probability corresponding to the initial bounding box, the preset nonmaximum suppression algorithm, and the target financial business template to obtain the target identification result.

11. The computer vision-based financial collateral surveying method according to claim 10, characterized in that, The step of performing target recognition on the target keyframe based on the initial bounding box, the initial confidence level corresponding to the initial bounding box, the class probability corresponding to the initial bounding box, a preset non-maximum suppression algorithm, and the target financial business template to obtain the target recognition result includes: The target confidence level of the initial bounding box is determined based on the product of the initial confidence level corresponding to the initial bounding box and the class probability corresponding to the initial bounding box. Based on the target confidence level, the preset nonmaximum suppression algorithm, and the target financial business template, target recognition is performed on the target keyframe to obtain the target recognition result.

12. The computer vision-based financial collateral surveying method according to claim 11, characterized in that, The step of performing target recognition on the target keyframe based on the target confidence level, the preset non-maximum suppression algorithm, and the target financial business template to obtain the target recognition result includes: Based on the target confidence level, the initial bounding boxes are sorted to obtain a first bounding box list, and the target result set is initialized. The initial bounding box with the highest target confidence is determined from the first bounding box list, and the initial bounding box with the highest target confidence is determined as the target bounding box; Add the target bounding box to the target result set, and remove the target bounding box from the first bounding box list; Determine the initial intersection-union ratio (IUU) between the target bounding box and each of the remaining initial bounding boxes in the first bounding box list, and determine the target IUU that is greater than a preset IUU threshold from the initial IUUU ratio; Remove the initial bounding box corresponding to the target intersection-union ratio from the first bounding box list, and jump to the step of determining the initial bounding box with the highest confidence of the target as the target bounding box, until the first bounding box list is empty, and generate the target recognition result corresponding to the target keyframe based on the target financial business template and the target result set.

13. A computer vision-based financial collateral surveying device, characterized in that, include: The target keyframe determination module is used to acquire the target video of the target house collateral and decode the target video to determine several target keyframes of the target video. The target keyframe set determination module is used to determine the set of keyframes corresponding to the target house layout and preset room type of the target house collateral from the plurality of target keyframes, so as to obtain the target keyframe set and determine the target financial business template corresponding to the target house collateral. The target recognition result determination module is used to perform target recognition on the target keyframes in the target keyframe photo set based on the target financial business template and the preset real-time target detection algorithm to obtain the corresponding target recognition result. The target survey report generation module is used to generate the survey and analysis results of the target house collateral based on the target identification results, the target keyframe album and the preset evaluation strategy, and to generate a target survey report of the target house collateral based on preset artificial intelligence technology, preset prompt words and the survey and analysis results; The target survey report generation module includes: The target survey summary generation submodule is used to generate a target survey summary of the target house collateral based on the target identification result. The value assessment result generation submodule is used to perform deduplication processing on the target identification result based on the target keyframe image set, and generate the value assessment result corresponding to the target house collateral based on the deduplicated target identification result; The risk assessment result determination submodule is used to generate an integrity assessment result corresponding to the target house collateral based on the deduplicated target identification result and the target keyframe album, and to conduct a risk assessment on the target house collateral based on the deduplicated target identification result and preset risk assessment indicators to obtain the corresponding risk assessment result; The survey and analysis results generation submodule is used to generate the survey and analysis results of the target property collateral based on the target survey summary, the value assessment results, the integrity assessment results, and the risk assessment results. The risk assessment result determination submodule includes: An abnormal risk assessment result determination unit is used to assess the valuation reasonableness of the target house collateral based on the valuation result to obtain a valuation reasonableness assessment result, and to conduct an abnormal risk assessment of the target house collateral based on the deduplicated target identification result to obtain an abnormal risk assessment result. The risk assessment result determination unit is used to generate the risk assessment result corresponding to the target house collateral based on the valuation reasonableness assessment result, the integrity assessment result, the abnormal risk assessment result, and the overall risk assessment result corresponding to the target house collateral. The target survey report generation module includes: The target survey data determination unit is used to perform data structuring processing on the initial survey data corresponding to the survey analysis results based on a preset data structure, and to determine the target survey data corresponding to the initial survey data after data structuring processing according to a preset text format. The target survey report generation unit is used to generate the target survey report of the target house collateral based on preset generative artificial intelligence technology, preset prompt words and target survey data.

14. An electronic device, characterized in that, The electronic device includes a processor and a memory; wherein the memory is used to store a computer program, which is loaded and executed by the processor to implement the computer vision-based financial collateral survey method as described in any one of claims 1 to 12.

15. A computer-readable storage medium, characterized in that, Used to store a computer program, which, when executed by a processor, implements the computer vision-based financial collateral survey method as described in any one of claims 1 to 12.

Images