Visual model engine platform based on AI deep learning

Abstract

The present invention discloses a visual model engine platform based on AI deep learning, relates to the field of deep learning technology, and comprises a control center, wherein the control center is connected with a data acquisition module, a front-end preprocessing module, a data processing module and a training and analysis module; a target image and a target video are collected by the data acquisition module; a query image is set in the front-end preprocessing module for image-text conversion and keyword extraction to obtain a query text library and a keyword library; a calibration size is set in the data processing module to perform standard proofreading and signal conversion on the input image to obtain a measured pixel sequence and a feature filtering end is set to perform feature filtering to obtain a measured filter sequence; a feature measured spectrum diagram is generated in the training and analysis module according to the measured filter sequence, and segment matching is performed on the feature measured spectrum diagram to obtain a target query text library and a video query text library; the efficiency of users in obtaining and understanding visual information is improved, and the intelligence and inclusiveness of the service are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a visual model engine platform based on AI deep learning. Background Art

[0002] Artificial intelligence is a branch of computer science that aims to enable computer systems to mimic various aspects of human intelligence. It covers a range of technologies and methods from simple rule-based systems to advanced systems capable of complex learning and problem solving. Deep learning uses artificial neural network models to simulate the workings of the human brain and learns data feature representation and pattern recognition through a multi-level neural network structure. The core idea of ​​deep learning is to represent the high-order features of data through a multi-level neural network, thereby achieving the ability to abstract and generalize data.

[0003] Visual models refer to mathematical models or computational models used to understand and interpret visual data. These models are usually used in computer vision tasks such as image recognition, object detection, image segmentation, face recognition, etc. Traditional visual model engines usually rely on hand-designed feature extractors and complex algorithms, which are difficult to cope with the challenges of complex visual tasks and large-scale data; therefore, it is of great theoretical and practical significance to intelligently process the visual information uploaded by users and obtain the corresponding textual interpretation and labeling of the visual information.

[0004] How to use deep learning technology to set the calibration size to standardize the input image and perform signal conversion, obtain the measured pixel sequence and set the feature filter end to perform feature filtering, obtain the measured filter sequence, generate the feature measured spectrum diagram according to the measured filter sequence, perform segment matching on the feature measured spectrum diagram, and obtain the target query text library and the video query text library are the problems we need to solve. To this end, we now provide a visual model engine platform based on AI deep learning. Summary of the invention

[0005] The purpose of the present invention can be achieved through the following technical solutions:

[0006] A visual model engine platform based on AI deep learning, including a control center, wherein the control center is connected to a data acquisition module, a front-end preprocessing module, a data processing module, and a training analysis module;

[0007] The process of the data acquisition module acquiring the target image and the target video includes:

[0008] Setting an access port and a sample collection end, wherein the access port is connected to a plurality of collection devices;

[0009] The sample collection end is used to set an inquiry image and obtain inquiry text data according to the inquiry image;

[0010] Capture the comprehensive information of the acquisition device through the access port to obtain comprehensive visual data;

[0011] The visual comprehensive data includes a target image and a target video.

[0012] The process of building a query library based on query text data and a keyword library based on text keywords includes:

[0013] An inquiry image is set according to the set sample collection terminal, and the obtained inquiry image is uploaded to the control center, and the control center sends an inquiry instruction to the sample user according to the obtained inquiry image;

[0014] Convert the query image into text according to the received query instruction to obtain query text data;

[0015] Building a query library based on the obtained query text data, uploading the obtained query text data to the query library, and associating the obtained query library with the corresponding query image;

[0016] Extract keywords from the query text data according to the query instruction to obtain text keywords;

[0017] A keyword library is constructed according to the obtained text keywords, the obtained text keywords are uploaded to the keyword library, and the obtained keyword library is associated with the corresponding query image.

[0018] The process of obtaining a reference image includes:

[0019] Set a frame node, perform frame extraction on the obtained target video according to the obtained frame node, and obtain a video frame image;

[0020] Marking the obtained query image, video frame image and target image as input images;

[0021] The calibration size is set to perform standard calibration on the input image to obtain a reference image.

[0022] The process of obtaining characteristic parameters includes:

[0023] Marking pixels of the obtained reference image to obtain measured pixel values;

[0024] Performing signal conversion on the obtained measured pixel value to obtain a measured pixel signal;

[0025] The obtained measured pixel signals are combined according to the obtained reference image to obtain a measured pixel sequence, and the obtained measured pixel sequence is characteristically marked to obtain characteristic parameters.

[0026] The process of obtaining the measured filter sequence includes:

[0027] Setting the extraction base according to the obtained characteristic parameters;

[0028] Setting a characteristic filter end, performing a filter design on the obtained extraction base through the characteristic filter end, and obtaining a filter characteristic coefficient;

[0029] The obtained measured pixel sequence is uploaded to the feature filtering end, and the measured pixel sequence is feature filtered by the feature filtering end to obtain a measured filter sequence.

[0030] The process of obtaining the characteristic measured curve includes:

[0031] generating a measured signal spectrum diagram according to the obtained measured filter sequence;

[0032] Marking the spectrum lines of the obtained measured signal spectrum diagram to obtain a measured filtering curve;

[0033] The measured filter curve is discretely sampled according to the obtained measured filter sequence to obtain measured sampling points, a filter acquisition window is set according to the obtained measured sampling points, and the obtained measured filter curve is discretely windowed according to the filter acquisition window to obtain a characteristic measured curve.

[0034] The process of obtaining the target inquiry text library and the video inquiry text library includes:

[0035] The measured signal spectrum is updated and marked according to the obtained characteristic measured curve to obtain a characteristic query spectrum, a characteristic video spectrum and a characteristic target spectrum;

[0036] A query sample library is constructed according to the obtained characteristic query spectrum diagram, and segment matching is performed on the characteristic target spectrum diagram according to the obtained query sample library to obtain a segment matching coefficient;

[0037] Perform image-text matching on the feature target graph according to the segment matching coefficient and the query text library to obtain the target query text library;

[0038] Perform segment matching on the characteristic video spectrum graph according to the query sample library to obtain the video matching coefficient;

[0039] The characteristic video spectrum graph is matched with keywords according to the video matching coefficient and the keyword library to obtain the video query text library.

[0040] Compared with the prior art, the invention has the following beneficial effects: setting the inquiry image for image-text conversion to obtain inquiry text data, constructing an inquiry library based on the inquiry text data, extracting keywords from the inquiry text data, and obtaining a keyword library; by setting the image samples and text samples in advance, it is beneficial to improve the image comparison efficiency and text matching rate of visual information;

[0041] Setting the calibration size to perform standard calibration and signal conversion on the video data, image data and query image, obtaining the measured pixel sequence and setting the feature filter end to perform feature filtering to obtain the measured filter sequence; processing the visual information of different acquisition times and different formats into the signal sequence form of the same standard is conducive to enhancing the feature extraction of the image, facilitating the matching of the visual image information with the image sample, and increasing the matching success rate and matching accuracy;

[0042] A characteristic measured spectrum graph is generated according to the measured filtering sequence, and segment matching is performed on the characteristic measured spectrum graph to obtain the target query text library and the video query text library; providing intelligent visual information conversion text data service is conducive to improving user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0044] Figure 1 It is a schematic diagram of the present invention. DETAILED DESCRIPTION

[0045] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] like Figure 1 As shown, the visual model engine platform based on AI deep learning includes a control center, which is connected to a data acquisition module, a front-end preprocessing module, a data processing module, and a training analysis module;

[0047] The data acquisition module is used to acquire target images and target videos, and the specific process includes:

[0048] Setting an access port and a sample collection end, wherein the access port is connected to a plurality of collection devices;

[0049] The sample collection end is used to set an inquiry image and obtain inquiry text data according to the inquiry image;

[0050] Furthermore, the acquisition equipment includes cameras, drones, satellites, etc. for acquiring images and video data of target objects;

[0051] The comprehensive information of the target object is collected by the collection device, and the collected comprehensive information is uploaded to the access port, and the uploaded comprehensive information is captured by the access port to obtain visual comprehensive data;

[0052] The visual comprehensive data includes a target image and a target video.

[0053] The front-end preprocessing module is used to set the query image to perform image-text conversion and keyword extraction to obtain the query library and keyword library. The specific process includes:

[0054] Setting the query image according to the set sample collection end, and uploading the obtained query image to the control center;

[0055] The control center sends an inquiry instruction to the sample user according to the obtained inquiry image;

[0056] Convert the query image into text according to the received query instruction to obtain query text data;

[0057] It should be further explained that, in the specific implementation process, the query image represents image data that can cover various types and different fields; the sample users represent the pre-selected group of people for describing the query image; the query instruction represents the required instruction for describing the query image, for example, the query instruction is "describe the color information of the query image", "ask whether the query image contains a person, if it contains a person, please describe the person's appearance and clothing";

[0058] The image-text conversion means answering the inquiry instruction according to the inquiry image, recording the answer information, and marking the obtained answer information as inquiry text data; the inquiry text data is the text data obtained by the sample user answering the inquiry image according to the inquiry instruction;

[0059] Building a query library based on the obtained query text data, uploading the obtained query text data to the query library, and associating the obtained query library with the corresponding query image;

[0060] Extract keywords from the query text data according to the query instruction to obtain text keywords;

[0061] Further, the keyword extraction means screening the query text data according to the question words in the query instruction. For example, according to the query instructions of "describe the color information of the query image" and "whether the query image contains people, if it contains people, please describe the appearance and clothing of the people", the question words are "color", "people" and "clothing appearance", and the corresponding answers of the sample users also contain the color information, person information and clothing appearance information in the picture, then the obtained color information, person information and clothing appearance information are the text keywords of the query picture;

[0062] Build a keyword library based on the obtained text keywords, and upload the obtained text keywords to the keyword library;

[0063] The obtained keyword library is associated with the corresponding query image.

[0064] The data processing module is used to set the calibration size to perform standard calibration and signal conversion on the input image, obtain the measured pixel sequence and set the feature filter end to perform feature filtering to obtain the measured filter sequence. The specific process includes:

[0065] Set a frame node, perform frame extraction on the obtained target video according to the obtained frame node, and obtain a video frame image;

[0066] It should be further explained that, in the specific implementation process, the frame node is determined according to the sampling rate and frame rate of the target video, and represents the time node of extracting the image in the target video; the frame extraction means reading the target video frame by frame through the video processing software, that is, pausing at each frame node to obtain the data of the current frame node, that is, the video frame image;

[0067] Marking the obtained query image, video frame image and target image as input images;

[0068] Setting a calibration size, the calibration size including an image size and a channel criterion;

[0069] Furthermore, the image size includes image width and image length, which are used to control the resolution of the input image. The image width represents the number of pixels in the vertical direction, and the image length represents the number of pixels in the horizontal direction. The channel criterion represents the number of color channels in the image. For example, if the number of channels is 3, it means that three colors are allowed to pass through.

[0070] Performing standard calibration on the input image according to the obtained calibration size to obtain a reference image;

[0071] The reference images include a reference query image, a reference video image, and a reference target image;

[0072] Furthermore, the standard calibration means that each reference image is uniformed in resolution and color according to the calibration size;

[0073] Marking pixels of the obtained reference image to obtain measured pixel values;

[0074] It should be further explained that, in the specific implementation process, the pixel marking means marking the pixels in the reference image, that is, marking the pixel position and pixel value, wherein the pixel value is the measured pixel value, and a reference image includes a plurality of measured pixel values, and each pixel corresponds to a measured pixel value;

[0075] Performing signal conversion on the obtained measured pixel value to obtain a measured pixel signal;

[0076] Combining the obtained measured pixel signals according to the obtained reference image to obtain a measured pixel sequence;

[0077] The measured pixel sequence includes a reference query sequence, a reference video sequence and a reference target sequence;

[0078] It should be further explained that, in a specific implementation process, since the reference image is composed of a plurality of measured pixel values, a measured pixel sequence composed of a combination of measured pixel signals converted from measured pixel values ​​in the reference image can represent the reference image; the measured pixel sequence is a signal sequence composed of a plurality of measured pixel signals;

[0079] Furthermore, if the target video is composed of a plurality of video frame images, the reference video sequence is obtained by first combining the measured pixel signals corresponding to the measured pixel values ​​in the video frame images to obtain a reference frame sequence corresponding to a single video frame image, and then combining the reference frame sequence with the plurality of video frame images included in the target video to obtain a reference video sequence;

[0080] Performing characteristic marking on the obtained measured pixel sequence to obtain characteristic parameters;

[0081] Furthermore, the characteristic parameters include frequency range, stability, time domain characteristics, and frequency domain characteristics;

[0082] Setting an extraction basis according to the obtained characteristic parameters, wherein the extraction basis includes a cutoff frequency, an order, a bandwidth, and a ripple;

[0083] It should be further explained that, in the specific implementation process, the extraction cardinality is determined according to the characteristic parameters of the measured pixel sequence, and the set extraction cardinality is the premise of setting the filter. Therefore, different extraction cardins can be obtained according to the characteristic parameters of different signals, and then the corresponding filters are set to process the signals.

[0084] Setting a characteristic filter end, performing a filter design on the obtained extraction base through the characteristic filter end, and obtaining a filter characteristic coefficient;

[0085] Furthermore, the characteristic filter end is a filter for processing the measured pixel sequence. In this embodiment, the characteristic filter end set for processing the measured pixel sequence is a bandpass filter;

[0086] The obtained extraction base is uploaded to the filter design software, for example, the filter design software is MATLAB; the extraction base is input through the design function of the filter design software to generate the filter characteristic coefficient;

[0087] The obtained measured pixel sequence is uploaded to the feature filtering end, and the measured pixel sequence is subjected to feature filtering by the feature filtering end to obtain a measured filter sequence;

[0088] The measured filtering sequence includes a filtering query sequence, a filtering video sequence and a filtering target sequence;

[0089] Furthermore, the feature filtering means filtering the filter feature coefficients in the feature filtering end and the measured pixel sequence to obtain a unilateral filtering sequence, performing logarithmic value extraction on the obtained unilateral filtering sequence to obtain a logarithmic filtering sequence, and performing inverse discrete cosine transform on the obtained logarithmic filtering sequence to obtain a measured filtering sequence;

[0090] In particular, the measured filter sequence is obtained by filtering the measured pixel sequence. According to the reference query sequence, reference video sequence and reference target sequence included in the measured pixel sequence, the measured filter sequence correspondingly includes the filter query sequence, the filter video sequence and the filter target sequence.

[0091] The training analysis module is used to generate a characteristic measured spectrum diagram according to the measured filter sequence, perform segment matching on the characteristic measured spectrum diagram, and obtain a target query text library and a video query text library. The specific process includes:

[0092] Generate a measured signal spectrum diagram according to the obtained measured filter sequence, wherein the measured signal spectrum diagram includes a measured query spectrum diagram, a measured video spectrum diagram, and a measured target spectrum diagram;

[0093] Further, the measured query spectrogram is generated based on the filtered query sequence, the measured video spectrogram is generated based on the filtered video sequence, and the measured target spectrogram is generated based on the filtered target sequence;

[0094] Marking the spectrum lines of the obtained measured signal spectrum diagram to obtain a measured filtering curve;

[0095] The measured filtering curves include a filtering inquiry curve, a filtering video curve and a filtering target curve;

[0096] Further, the measured filter curve is represented in the measured query spectrum diagram, the filter query curve is generated according to the filter query sequence, the filter video curve is represented in the measured video spectrum diagram according to the filter video sequence generated filter video curve, the filter target curve is represented in the measured target spectrum diagram according to the filter target sequence generated filter target curve;

[0097] Discretely sampling the obtained measured filtering curve according to the obtained measured filtering sequence to obtain measured sampling points;

[0098] The measured sampling points obtained are marked as k=0, 1, 2, 3, ..., v1, where v1 is a positive integer;

[0099] Furthermore, the measured sampling points represent discretization of the measured filtering curve according to the sampling frequency of the measured filtering sequence, and the corresponding discrete points are the measured sampling points;

[0100] A filtering acquisition window is set according to the obtained measured sampling points, and the length of the obtained filtering acquisition window is marked as C, and C is greater than 1;

[0101] The obtained filtering acquisition window is marked as G(k), where G(k)=a+0.46* , a represents the front-end factor, 0<a<1, and k=0, 1, 2, 3, ..., C-1;

[0102] Furthermore, the length of the filtering acquisition window is determined according to the signal length of the measured filtering sequence;

[0103] According to the obtained filtering acquisition window, discrete windowing is performed on the obtained measured filtering curve to obtain a characteristic measured curve;

[0104] The characteristic measured curves include a characteristic inquiry curve, a characteristic video curve and a characteristic target curve;

[0105] The obtained characteristic measured curve is marked as T(k), where T(k)=S(k)*G(k), and S(k) represents the measured filtering sequence;

[0106] According to the obtained characteristic measured curve, the measured signal spectrum diagram is updated and marked to obtain the characteristic measured spectrum diagram;

[0107] The characteristic measured spectrum graph includes a characteristic query spectrum graph, a characteristic video spectrum graph and a characteristic target spectrum graph; in particular, there is only one characteristic measured curve in a characteristic measured spectrum graph;

[0108] Building a query sample library according to the obtained characteristic query spectrum graph, and uploading the obtained characteristic query spectrum graph to the query sample library;

[0109] Performing segment matching on the characteristic target spectrum graph according to the obtained query sample library to obtain a segment matching coefficient;

[0110] The obtained segment matching coefficient is marked as ,in, , is the equalization factor, j represents the number of the characteristic query spectrum in the query sample library, j=1, 2, 3, ..., v2, v2 is a positive integer, Represents the characteristic query curve in the characteristic query spectrum diagram, represents the characteristic target curve in the characteristic target spectrum diagram, and dk represents the differential of k;

[0111] Performing image-text matching on the obtained characteristic target graph according to the obtained segment matching coefficient and the query text library to obtain a target query text library;

[0112] It should be further explained that, in the specific implementation process, the image-text matching process includes:

[0113] The obtained segment matching coefficients are sorted in descending order to obtain a matching coefficient sorting;

[0114] Mark the segment matching coefficients ranked in the top p in the matching coefficient sorting as qualified matching query graphs;

[0115] Obtain the query library associated with the query images corresponding to the first p qualified matching query images, recorded as the qualified query library;

[0116] The obtained qualified query library is sorted and combined according to the matching coefficient to form a target query text library; in particular, if the query text data in the target query text library has repeated sorting, the sorting is updated according to the number of repetitions;

[0117] Associating the obtained target query text library with the target image corresponding to the qualified matching query graph;

[0118] The query text data corresponding to the target image is obtained by acquiring the target query text library associated with the target image, that is, feature processing is performed on the target image, the obtained features are converted into text data, and the target image is marked and interpreted;

[0119] Perform segment matching on the characteristic video spectrum graph according to the obtained query sample library to obtain a video matching coefficient;

[0120] The obtained video matching coefficient is marked as ,in, , is the impact factor, Representing a characteristic video curve in a characteristic video spectrogram;

[0121] Perform keyword matching on the obtained characteristic video spectrum graph according to the obtained video matching coefficient and the keyword library to obtain a video query text library;

[0122] It should be further explained that, in the specific implementation process, the keyword matching process includes:

[0123] The obtained video matching coefficients are sorted in descending order to obtain a video matching coefficient sorting;

[0124] Mark the video matching coefficients ranked in the top m in the video matching coefficient ranking as qualified video query graphs;

[0125] Obtain a keyword library associated with the query image corresponding to the qualified video query image, recorded as a video query text library;

[0126] Associating the obtained video query text library with the target video corresponding to the qualified video query graph;

[0127] Obtaining text keywords corresponding to the target video by acquiring a video query text library associated with the target video, that is, performing feature processing on the target video and converting the obtained features into text keywords;

[0128] Furthermore, based on the several video frame images included in the target video, text marking and interpretation of the target video will include query images that match the several video frame images. If the target video is marked using query text data, there will be too much text content, which is not conducive to users quickly understanding the video content. If the target video is marked using text keywords, it will be conducive to quickly understanding the video content and increase the speed of video searching.

[0129] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to only specific implementation methods. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A visual model engine platform based on AI deep learning, including a control center, characterized by: The control center is connected to a data acquisition module, a front-end preprocessing module, a data processing module and a training analysis module; The data acquisition module is used to acquire target images and target videos; The front-end preprocessing module is used to set an inquiry image, send an inquiry instruction to a sample user and perform image-to-text conversion on the inquiry image to obtain inquiry text data, build an inquiry library based on the inquiry text data, extract keywords from the inquiry text data through the inquiry instruction, obtain text keywords, and build a keyword library based on the text keywords; The data processing module is used to set frame nodes to perform frame extraction on the target video to obtain a video frame image, set a calibration size to perform standard proofreading on the query image, the video frame image and the target image to obtain a reference image, the reference image includes a reference query image, a reference video image and a reference target image, perform pixel marking and signal conversion on the reference image to obtain a measured pixel sequence, the measured pixel sequence includes a reference query sequence, a reference video sequence and a reference target sequence, perform characteristic marking on the measured pixel sequence to obtain characteristic parameters, the characteristic parameters include frequency range, stationarity, time domain characteristics and frequency domain characteristics, set a characteristic filtering end according to the characteristic parameters and perform characteristic filtering on the measured pixel sequence to obtain a measured filter sequence, the measured filter sequence includes a filter query sequence, a filter video sequence and a filter target sequence; The training analysis module is used to generate a measured signal spectrum diagram according to the obtained measured filter sequence, and the measured signal spectrum diagram includes a measured query spectrum diagram, a measured video spectrum diagram and a measured target spectrum diagram; Marking the spectrum lines of the obtained measured signal spectrum diagram to obtain a measured filtering curve, wherein the measured filtering curve includes a filtering query curve, a filtering video curve, and a filtering target curve; Discretize and sample the obtained measured filter curve according to the sampling frequency of the obtained measured filter sequence to obtain corresponding discrete points, namely measured sampling points; A filtering acquisition window is set according to the obtained measured sampling points, and the length of the obtained filtering acquisition window is marked as C, where C is greater than 1. According to the obtained filtering acquisition window, a discrete window is performed on the obtained measured filtering curve to obtain a characteristic measured curve, wherein the characteristic measured curve includes a characteristic query curve, a characteristic video curve, and a characteristic target curve; Update and mark the measured signal spectrum diagram according to the obtained characteristic measured curve to obtain a characteristic measured spectrum diagram, wherein the characteristic measured spectrum diagram includes a characteristic query spectrum diagram, a characteristic video spectrum diagram, and a characteristic target spectrum diagram; Construct a query sample library according to the obtained characteristic query spectrum graph, upload the obtained characteristic query spectrum graph to the query sample library, perform segment matching on the characteristic target spectrum graph according to the obtained query sample library, obtain segment matching coefficients, sort the obtained segment matching coefficients in order from large to small, obtain matching coefficient ranking, mark the segment matching coefficients ranked in the top p in the matching coefficient ranking as qualified matching query graphs, obtain the query text library associated with the query images corresponding to the top p qualified matching query graphs, record them as qualified query libraries, combine the obtained qualified query libraries according to the matching coefficient ranking, and form a target query text library; The obtained target query text library is associated with the target image corresponding to the qualified matching query graph, and the query text data corresponding to the target image is obtained by acquiring the target query text library associated with the target image, that is, feature processing is performed on the target image, the obtained features are converted into text data, and the target image is marked and interpreted. According to the obtained query sample library, segment matching is performed on the feature video spectrum graph to obtain the video matching coefficient; The obtained characteristic video spectrum graph is subjected to keyword matching according to the obtained video matching coefficient and the keyword library to obtain a video query text library.

2. The AI ​​deep learning-based visual model engine platform according to claim 1, characterized in that: The process of the data acquisition module acquiring the target image and the target video includes: Setting an access port and a sample collection end, wherein the access port is connected to a plurality of collection devices; The sample collection end is used to set an inquiry image and obtain inquiry text data according to the inquiry image; Capture the comprehensive information of the acquisition device through the access port to obtain comprehensive visual data; The visual comprehensive data includes a target image and a target video.

3. The AI ​​deep learning-based visual model engine platform according to claim 2, characterized in that: The process of building a query library based on query text data and a keyword library based on text keywords includes: An inquiry image is set according to the set sample collection terminal, and the obtained inquiry image is uploaded to the control center, and the control center sends an inquiry instruction to the sample user according to the obtained inquiry image; Convert the query image into text according to the received query instruction to obtain query text data; Building a query library based on the obtained query text data, uploading the obtained query text data to the query library, and associating the obtained query library with the corresponding query image; Extract keywords from the query text data according to the query instruction to obtain text keywords; A keyword library is constructed according to the obtained text keywords, the obtained text keywords are uploaded to the keyword library, and the obtained keyword library is associated with the corresponding query image.

4. The AI ​​deep learning-based visual model engine platform according to claim 3, characterized in that: The process of obtaining a reference image includes: Set a frame node, perform frame extraction on the obtained target video according to the obtained frame node, and obtain a video frame image; Marking the obtained query image, video frame image and target image as input images; The calibration size is set to perform standard calibration on the input image to obtain a reference image.

5. The AI ​​deep learning-based visual model engine platform according to claim 4, characterized in that: The process of obtaining characteristic parameters includes: Marking pixels of the obtained reference image to obtain measured pixel values; Performing signal conversion on the obtained measured pixel value to obtain a measured pixel signal; The obtained measured pixel signals are combined according to the obtained reference image to obtain a measured pixel sequence, and the obtained measured pixel sequence is characteristically marked to obtain characteristic parameters.

6. The AI ​​deep learning-based visual model engine platform according to claim 5, characterized in that: The process of obtaining the measured filter sequence includes: Setting the extraction base according to the obtained characteristic parameters; Setting a characteristic filter end, performing a filter design on the obtained extraction base through the characteristic filter end, and obtaining a filter characteristic coefficient; The obtained measured pixel sequence is uploaded to the feature filtering end, and the measured pixel sequence is feature filtered by the feature filtering end to obtain a measured filter sequence.

Images