CN116628263A - Video retrieval method and device based on multiple modes, electronic equipment and storage medium - Google Patents
Video retrieval method and device based on multiple modes, electronic equipment and storage medium Download PDFInfo
- Publication number
- CN116628263A CN116628263A CN202310672336.1A CN202310672336A CN116628263A CN 116628263 A CN116628263 A CN 116628263A CN 202310672336 A CN202310672336 A CN 202310672336A CN 116628263 A CN116628263 A CN 116628263A
- Authority
- CN
- China
- Prior art keywords
- feature
- modal
- video
- features
- text
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 55
- 239000013598 vector Substances 0.000 claims abstract description 125
- 230000004927 fusion Effects 0.000 claims abstract description 48
- 238000011176 pooling Methods 0.000 claims abstract description 20
- 230000008569 process Effects 0.000 claims abstract description 13
- 230000000007 visual effect Effects 0.000 claims description 25
- 238000000605 extraction Methods 0.000 claims description 20
- 210000002569 neuron Anatomy 0.000 claims description 15
- 238000004590 computer program Methods 0.000 claims description 11
- 238000012545 processing Methods 0.000 claims description 11
- 238000013528 artificial neural network Methods 0.000 claims description 8
- 238000004422 calculation algorithm Methods 0.000 claims description 6
- 238000004458 analytical method Methods 0.000 description 12
- 230000006870 function Effects 0.000 description 9
- 238000004891 communication Methods 0.000 description 8
- 238000007726 management method Methods 0.000 description 6
- 238000013473 artificial intelligence Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 4
- 238000012937 correction Methods 0.000 description 3
- 238000007405 data analysis Methods 0.000 description 3
- 238000007499 fusion processing Methods 0.000 description 3
- 238000002789 length control Methods 0.000 description 3
- 244000046052 Phaseolus vulgaris Species 0.000 description 2
- 235000010627 Phaseolus vulgaris Nutrition 0.000 description 2
- 201000010099 disease Diseases 0.000 description 2
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 230000036541 health Effects 0.000 description 2
- 239000004973 liquid crystal related substance Substances 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 230000002093 peripheral effect Effects 0.000 description 2
- 235000007688 Lycopersicon esculentum Nutrition 0.000 description 1
- 240000003768 Solanum lycopersicum Species 0.000 description 1
- 244000061456 Solanum tuberosum Species 0.000 description 1
- 235000002595 Solanum tuberosum Nutrition 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000000802 evaporation-induced self-assembly Methods 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 238000002560 therapeutic procedure Methods 0.000 description 1
- 238000012549 training Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/70—Information retrieval; Database structures therefor; File system structures therefor of video data
- G06F16/78—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/783—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/70—Information retrieval; Database structures therefor; File system structures therefor of video data
- G06F16/71—Indexing; Data structures therefor; Storage structures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/70—Information retrieval; Database structures therefor; File system structures therefor of video data
- G06F16/75—Clustering; Classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Databases & Information Systems (AREA)
- Multimedia (AREA)
- Library & Information Science (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Medical Treatment And Welfare Office Work (AREA)
Abstract
The invention relates to the field of intelligent decision making and digital medical treatment, and discloses a video retrieval method, a device, electronic equipment and a storage medium based on multiple modes, wherein the method comprises the following steps: inquiring the multi-modal attribute of the video to be searched, selecting a multi-modal network, and extracting multi-modal characteristics in the video to be searched; extracting a characteristic sequence, constructing a fixed cluster group of the characteristic sequence, carrying out mean pooling on the fixed cluster group to obtain a pooled cluster group, and carrying out fixed-length alignment on the characteristic sequence according to the pooled cluster group to obtain an alignment sequence; extracting feature dimensions, aligning the dimension lengths to obtain alignment dimensions, and determining alignment multi-modal features of the multi-modal features according to the alignment sequences and the alignment dimensions; performing feature splicing and feature fusion; and carrying out feature vector coding on the search text, calculating cosine similarity, and determining a video search result of the search text according to the cosine similarity. The invention can realize automatic indexing and searching of the video and improve the comprehensiveness of the whole video searching process.
Description
Technical Field
The invention relates to the field of intelligent decision making and digital medical treatment, in particular to a video retrieval method and device based on multiple modes, electronic equipment and a computer readable storage medium.
Background
The multi-mode-based video retrieval is to extract multi-mode features contained in the dynamic video, classify the video according to the multi-mode features, and match the video with an input search text, so that the process from the input text to the retrieval of the video is realized.
At present, with the rise of video retrieval technology, people can support functions of disease auxiliary diagnosis, health management and the like by inputting a section of simple medical related text description or a plurality of keywords so as to search for a medical video retrieval mode of a medical video, a health maintenance video and the like which are wanted by the people, the traditional medical video retrieval mode marks the video manually, and marks a proper title label or content description and the like according to the content of the medical video, in the current self-media age, the information is usually added by a doctor, a hospital and other video creators while uploading the video, but the title label or the medical related description information of the medical video can not be fully summarized in the video sometimes, and because a small disease can involve a plurality of medical knowledge, the medical information in the title labels is even absent in some cases, for example, the video creators do not add description information themselves, or the medical information in the video comes from other sources such as a network, and the retrieval mode depending on a hard label can not accurately complete the retrieval task. Therefore, the medical video classification is realized by manually labeling, so that the video retrieval is not comprehensive enough.
Disclosure of Invention
The invention provides a video retrieval method, a device, electronic equipment and a computer readable storage medium based on multiple modes, which mainly aim to realize automatic indexing and retrieval of medical videos and improve the comprehensiveness of the whole medical video retrieval process.
In order to achieve the above object, the present invention provides a video retrieval method based on multiple modes, including:
acquiring a video to be searched, inquiring the multi-modal attribute of the video to be searched, selecting a multi-modal network corresponding to the multi-modal attribute, and extracting multi-modal characteristics in the video to be searched by utilizing the multi-modal network;
extracting a characteristic sequence of the multi-mode characteristic, constructing a fixed cluster group of the characteristic sequence, carrying out mean pooling on the fixed cluster group to obtain a pooled cluster group, and carrying out fixed-length alignment on the characteristic sequence according to the pooled cluster group to obtain an alignment sequence;
extracting feature dimensions of the multi-mode features, and aligning dimension lengths of the feature dimensions by utilizing a pre-constructed multi-layer perceptron to obtain pairs Ji Weidu;
determining an aligned multi-modal feature of the multi-modal features according to the aligned sequence and the aligned dimension, performing feature stitching on a plurality of features in the aligned multi-modal features to obtain a stitched multi-modal feature, and performing feature fusion on the stitched multi-modal feature by using a full-connection visual network to obtain a fused multi-modal feature;
And obtaining a search text, carrying out feature vector coding on the search text to obtain a coded text vector, calculating cosine similarity between the fused multi-mode features and the coded text vector, and determining a video search result of the search text according to the cosine similarity.
Optionally, the querying the multimodal attribute of the video to be retrieved includes:
performing playing processing on the video to be searched to obtain a playing video;
carrying out integrity check on the video elements of the playing video;
when the integrity check of the video elements of the played video is successful, determining the multi-modal attribute of the video to be retrieved;
when the integrity check of the video element of the played video fails, extracting a video failure part of which the integrity check of the video fails;
and determining the multi-modal attribute of the video to be retrieved according to the video failure part.
Optionally, the selecting the multi-mode network corresponding to the multi-mode attribute includes:
identifying a feature objective of the multimodal property;
acquiring a neural network structure, and inquiring a network purpose of the neural network structure;
matching the network purpose with the characteristic purpose;
And when the network purpose is successfully matched with the characteristic purpose, determining the multi-mode network corresponding to the multi-mode attribute.
Optionally, the aligning the dimension length of the feature dimension by using a pre-built multi-layer perceptron to obtain an aligned dimension includes:
configuring a target length of the feature dimension;
selecting a multi-layer neuron in the pre-constructed multi-layer perceptron according to the target length and the characteristic dimension;
and performing dimension length alignment operation of the characteristic dimension by using the multi-layer neuron to obtain the pair Ji Weidu.
Optionally, the feature fusion is performed on the spliced multi-modal feature by using a fully connected visual network to obtain a fused multi-modal feature, including:
and carrying out feature column fusion on the spliced multi-mode features by using the following formula to obtain column fusion features:
U *, ＝ *, + 2 σ(W 1 LayerNorm() *, )
wherein U is *, Representing the column fusion feature, X *, A column vector set representing a feature vector set in the stitched multimodal feature, i representing a column of the feature vector set in the stitched multimodal feature, W 1 And W is equal to 2 Respectively represent the feature sequence fusion timeWeights of front and rear layers NLP in the fully connected visual network, σ represents bias in the fully connected visual network, layerNorm represents an algorithm for normalizing the spliced multimodal features, layerNorm () *, Representing a process of normalizing a column vector set of a feature vector set in the spliced multi-modal feature, and X represents the feature vector set in the spliced multi-modal feature;
and carrying out feature row fusion on the column fusion features by using the following formula to obtain the fusion multi-modal features:
Y j,* ＝ j,* + 4 σ(W 3 LayerNorm() j,* )
wherein Y is j,* Representing the fused multi-modal feature, U j,* Representing a row feature vector, W, in the column fusion feature 3 And W is equal to 4 Respectively represent weights of front and rear layers NLP in the fully connected visual network when feature lines are fused, sigma represents bias in the fully connected visual network, layerNorm represents an algorithm for standardizing the spliced multi-modal features, layerNorm () j,* Representing the process of normalizing the row feature vectors in the column fusion feature, and U represents the column fusion feature.
Optionally, the feature vector encoding of the search text to obtain an encoded text vector includes:
determining a text category of the retrieved text;
constructing an initial code of the text category;
matching the search text with the text category;
and when the search text is successfully matched with the text category, carrying out feature vector coding on the search text according to the initial coding to obtain a coded text vector.
Optionally, the calculating cosine similarity between the fused multimodal feature and the encoded text vector includes:
calculating cosine similarity between the fused multi-modal feature and the encoded text vector using the formula
Wherein W is uv And (3) representing cosine similarity between the fused multi-modal feature and the coded text vector, wherein N (u) represents a feature corresponding to u in the fused multi-modal feature set N, and M (v) represents a text vector corresponding to v in the coded text vector set M.
In order to solve the above problems, the present invention further provides a video retrieval device based on multiple modes, the device comprising:
the feature extraction module is used for acquiring the video to be searched, inquiring the multi-modal attribute of the video to be searched, selecting a multi-modal network corresponding to the multi-modal attribute, and extracting multi-modal features in the video to be searched by utilizing the multi-modal network;
the fixed-length alignment module is used for extracting the characteristic sequences of the multi-mode characteristics, constructing fixed cluster groups of the characteristic sequences, carrying out mean pooling on the fixed cluster groups to obtain pooled cluster groups, and carrying out fixed-length alignment on the characteristic sequences according to the pooled cluster groups to obtain alignment sequences;
The dimension alignment module is used for extracting the characteristic dimension of the multi-mode characteristic, and aligning the dimension length of the characteristic dimension by utilizing a pre-constructed multi-layer perceptron to obtain a pair Ji Weidu;
the feature fusion module is used for determining the aligned multi-modal features in the multi-modal features according to the aligned sequences and the aligned dimensions, performing feature stitching on a plurality of features in the aligned multi-modal features to obtain stitched multi-modal features, and performing feature fusion on the stitched multi-modal features by using a full-connection visual network to obtain fused multi-modal features;
the retrieval determining module is used for obtaining a retrieval text, carrying out feature vector coding on the retrieval text to obtain a coded text vector, calculating cosine similarity between the fused multi-mode features and the coded text vector, and determining a video retrieval result of the retrieval text according to the cosine similarity.
In order to solve the above-mentioned problems, the present invention also provides an electronic apparatus including:
at least one processor; the method comprises the steps of,
a memory communicatively coupled to the at least one processor; wherein,,
the memory stores a computer program executable by the at least one processor to implement the multi-modality based video retrieval method described above.
In order to solve the above-mentioned problems, the present invention also provides a computer-readable storage medium having stored therein at least one computer program that is executed by a processor in an electronic device to implement the above-mentioned multi-modality based video retrieval method.
It can be seen that, in the embodiment of the present invention, by acquiring the video to be searched for performing label distribution on the medical video uploaded in the platform, further, by querying the multi-modal attribute of the video to be searched for implementing the extraction of the medical key feature of the video to be searched from multiple dimensions, the comprehensiveness of the medical information analysis of the video to be searched is improved, further, by selecting the multi-modal network corresponding to the multi-modal attribute for extracting the medical feature of the video by using the different types of feature extraction network, the single-mode of the label of the medical video is reduced, the comprehensiveness of the medical video analysis is improved, further, by extracting the multi-modal feature in the video to be searched by using the multi-modal network, in order to obtain medical key information attached to the video to be searched from a plurality of feature angles and ensure subsequent comprehensive analysis of medical information carried in the video to be searched, the embodiment of the invention performs unified length adjustment on the lengths of different medical feature sequences generated by different medical video time lengths and ensures the uniformity of medical feature data by extracting the feature sequences of the multi-modal features, further, the embodiment of the invention realizes the length control of the feature sequences by constructing fixed cluster groups of the feature sequences and dividing different types of medical feature vectors into different clusters, namely one cluster comprises one type of medical feature vector and ensures classification processing of the medical feature vectors, further, the embodiment of the invention performs average pooling on the fixed cluster groups to determine the average length of the feature sequences, furthermore, the embodiment of the invention aims at the characteristic sequences to adjust the sequences with different lengths into sequences with uniform lengths according to the pooling cluster group, ensures that the sequences are not influenced by disordered medical data when being analyzed, improves the efficiency of medical data analysis, and further aims at the non-standard dimension lengths by extracting the characteristic dimension of the multi-modal characteristics to adjust the non-standard dimension lengths, and aims at correcting the dimension lengths with different lengths by utilizing a pre-constructed multi-layer perceptron to improve the standard degree of the medical data, and determines the aligned multi-modal characteristics in the multi-modal characteristics to recombine the medical characteristics after the correction format according to the aligned sequences and the aligned dimensions, the order of medical data is ensured, further, the embodiment of the invention combines the medical feature vector with other medical feature vectors by performing feature stitching on a plurality of medical features in the aligned multi-modal features, and ensures subsequent fusion processing, further, the embodiment of the invention fuses the stitched multi-modal features by utilizing a fully connected vision network, so as to fuse the extracted multi-modal features to obtain more hidden medical features, promote the comprehensiveness of medical feature extraction, and ensure the comprehensiveness of medical video retrieval, and the embodiment of the invention obtains a retrieval text for matching the medical key text of a search video input by a user with the corresponding medical video, further, the embodiment of the invention encodes the feature vector by using the retrieval text, the method and the device are used for converting the search text into a medical feature vector form so as to be convenient for matching with the feature vector of the medical video, further, the embodiment of the invention realizes artificial intelligent automation by calculating cosine similarity between the fused multi-modal feature and the coded text vector so as to be used for determining the medical video to be searched which is close to the search text, and improves the comprehensiveness of the search by searching through the multi-dimensional feature. Therefore, the multi-mode-based video retrieval method, the multi-mode-based video retrieval device, the electronic equipment and the computer-readable storage medium can realize automatic indexing and retrieval of medical videos, and improve the comprehensiveness of the whole medical video retrieval process.
Drawings
Fig. 1 is a flow chart of a multi-mode-based video retrieval method according to an embodiment of the present invention;
FIG. 2 is a schematic block diagram of a multi-mode video retrieval device according to an embodiment of the present invention;
fig. 3 is a schematic diagram of an internal structure of an electronic device implementing a multi-mode video retrieval method according to an embodiment of the present invention;
the achievement of the objects, functional features and advantages of the present invention will be further described with reference to the accompanying drawings, in conjunction with the embodiments.
Detailed Description
It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
The embodiment of the invention provides a video retrieval method based on multiple modes. The execution subject of the multi-mode video retrieval method includes, but is not limited to, at least one of a server, a terminal, and the like, which can be configured to execute the method provided by the embodiment of the invention. In other words, the multi-modality based video retrieval method may be performed by software or hardware installed in a terminal device or a server device, and the software may be a blockchain platform. The service end includes but is not limited to: a single server, a server cluster, a cloud server or a cloud server cluster, and the like. The server may be an independent server, or may be a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (Content Delivery Network, CDN), and basic cloud computing services such as big data and artificial intelligence platforms.
Referring to fig. 1, a flowchart of a multi-mode-based video retrieval method according to an embodiment of the invention is shown. In the embodiment of the invention, the multi-mode-based video retrieval method comprises the following steps S1-S5:
s1, acquiring a video to be searched, inquiring the multi-modal attribute of the video to be searched, selecting a multi-modal network corresponding to the multi-modal attribute, and extracting multi-modal characteristics in the video to be searched by utilizing the multi-modal network.
The embodiment of the invention is used for distributing the labels to the medical videos uploaded in the platform by acquiring the videos to be searched. The video to be searched refers to a video uploaded to a platform by a user, is generated based on different business scenes, and can be a medical video, a science popularization knowledge video, a propaganda video of a therapy, a medical payment video and the like in a digital medical industry scene; in the financial business scenario, the video to be searched can be a fund science popularization video, a stock rise and fall analysis video, a bank card handling video and the like.
Further, the embodiment of the invention is used for extracting the medical key characteristics of the video to be searched from multiple dimensions by inquiring the multi-modal attribute of the video to be searched, so that the comprehensiveness of medical information analysis of the video to be searched is improved. The multi-modal attribute refers to an analysis dimension of the video to be retrieved, and includes attributes such as object category information in the video, video scene information, dynamic time sequence information in the video, audio in the video, and subtitles in the video.
In an embodiment of the present invention, the querying the multimodal attribute of the video to be retrieved includes: performing playing processing on the video to be searched to obtain a playing video; carrying out integrity check on the video elements of the playing video; when the integrity check of the video elements of the played video is successful, determining the multi-modal attribute of the video to be retrieved; when the integrity check of the video element of the played video fails, extracting a video failure part of which the integrity check of the video fails; and determining the multi-modal attribute of the video to be retrieved according to the video failure part.
The method includes the steps that the video to be searched is played, when the video to be searched is played, whether video elements of the video to be searched are complete or not can be inquired, for example, whether sound of the video to be searched is missing or not, if the sound is not missing, the sound of the video to be searched is complete enough, and then audio information is used as the multi-mode attribute; if the sound is missing, the sound representing the video is not complete enough, and the missing sound is not used as the multi-modal attribute of the video.
Further, the embodiment of the invention selects the multi-modal network corresponding to the multi-modal attribute to extract the medical features of the video by utilizing the different types of feature extraction networks, thereby reducing the singleization of the labels of the medical video and improving the comprehensiveness of the medical video analysis. The multi-modal network refers to a feature extraction network for extracting video features, which is also called as an expert, and comprises that for 1, static image information, a ResNet network trained on an ImageNet can be adopted as an expert to extract object category information, and 2, a SENET network pre-trained on a Place365 is adopted to extract scene information; for dynamic time sequence information, 3, video time sequence characteristics can be extracted by adopting models such as I3D, 4, for audio, audio can be sampled and characteristics can be extracted by adopting VGGish network, 5, and for subtitles, text characteristics can be encoded by adopting pre-training models such as BERT, GPT and the like.
In an embodiment of the present invention, the selecting the multi-modal network corresponding to the multi-modal attribute includes: identifying a feature objective of the multimodal property; acquiring a neural network structure, and inquiring a network purpose of the neural network structure; matching the network purpose with the characteristic purpose; and when the network purpose is successfully matched with the characteristic purpose, determining the multi-mode network corresponding to the multi-mode attribute.
The feature object refers to an object of extracting a feature corresponding to the multi-modal attribute, for example, when the multi-modal attribute is an object type, the feature object is to extract a foreground region feature of the video, and the network object refers to a feature extraction object of an existing neural network structure, for example, an object of a res net network is to extract a foreground region feature in the video.
Further, the embodiment of the invention extracts the multi-modal characteristics in the video to be searched by utilizing the multi-modal network, so as to be used for acquiring the medical key information attached to the video to be searched from a plurality of characteristic angles, thereby guaranteeing the subsequent comprehensive analysis of the medical information carried in the video to be searched. The multi-modal feature is a feature vector corresponding to the multi-modal attribute, and includes features such as object category information in video, video scene information, dynamic time sequence information in video, audio in video, and subtitles in video.
In an embodiment of the present invention, the extracting, by using the multi-modal network, multi-modal features in the video to be retrieved includes: inquiring the network structure of the multi-mode network; identifying a feature extraction link of the network structure; and executing the extraction of the multi-modal features in the video to be retrieved by utilizing the feature extraction link.
For example, if the multi-mode network is a res net network, and the network structure of the res net network is queried to obtain a final output by first a convolution layer, then a pooling layer, then a series of residual structures, and finally an average pooling downsampling and a full connection layer, if the network structure is the convolution layer, determining a feature extraction link of the convolution layer to convolve the size of a picture, the feature extraction can be implemented by using the feature extraction link, for example, an input image is 224×224×3 in size, and then 7×7,64 is convolved in the convolution layer link, stride=2, stride=3, and then the output size is (224-7+6)/2+1=112.5, and then 112×112×64 is output.
S2, extracting a characteristic sequence of the multi-mode characteristic, constructing a fixed cluster group of the characteristic sequence, carrying out mean pooling on the fixed cluster group to obtain a pooled cluster group, and carrying out fixed-length alignment on the characteristic sequence according to the pooled cluster group to obtain an alignment sequence.
The embodiment of the invention is used for carrying out uniform length adjustment on the lengths of different medical characteristic sequences generated by different medical video time lengths by extracting the characteristic sequences of the multi-mode characteristics, so as to ensure the uniformity of medical characteristic data. The feature sequence refers to the length of the feature vector sequence of the multi-modal feature, and may be understood as the length of a feature vector corresponding to a certain dimension among a plurality of dimensions of the extracted multi-modal feature.
In an embodiment of the present invention, the extracting the feature sequence of the multi-modal feature includes: acquiring a feature vector of the multi-mode feature, and extracting vector data in the feature vector; and identifying a vector sequence of the vector data, and taking the vector sequence as a characteristic sequence of the multi-mode characteristic.
The method includes the steps that when the feature vector of the multi-mode feature is obtained as a one-dimensional array, data in the array are extracted, the extracted data with a front-back relation are constructed as a sequence, and the vector sequence is obtained.
Further, the embodiment of the invention is used for dividing the medical feature vectors of different categories into different clusters by constructing the fixed cluster group of the feature sequences, namely, one cluster comprises the medical feature vector of one category, so that the classification processing of the medical feature vector is ensured. The fixed cluster group refers to a cluster containing vector data in a certain category of feature vectors.
In an embodiment of the present invention, the constructing the fixed cluster group of the feature sequence includes: inquiring the feature category of the feature sequence; determining a fixed category of the feature sequence according to the feature category; according to the fixed category; and determining a fixed cluster group of the characteristic sequences.
The feature class of the feature sequence can be queried through the number of feature vectors output by querying, if the number of the feature vectors is 5, the feature class is determined to be 5, meanwhile, the feature sequence with 5 classes is determined, the fixed class of the feature sequence is 5, and finally, 5 corresponding fixed clusters are obtained.
Further, the embodiment of the invention realizes the length control of the characteristic sequence by carrying out the mean value pooling on the fixed cluster group so as to be used for determining the average length of the characteristic sequence.
In an embodiment of the present invention, the fixed cluster group is subjected to mean pooling by using the following formula to obtain a pooled cluster group:
wherein g p A pooling cluster representing p-fixed clusters among the pooling clusters, E representing a total number of feature sequences in the p-fixed clusters among the fixed clusters, q representing q-feature sequences in the p-fixed clusters, g pq Representing the length of the q characteristic sequences in the p fixed cluster group in the fixed cluster group.
Furthermore, according to the embodiment of the invention, the characteristic sequences are aligned in a fixed length mode according to the pooling cluster group, so that sequences with different lengths are adjusted to be sequences with uniform lengths, the situation that the sequence is not affected by disordered medical data during subsequent analysis is ensured, and the medical data analysis efficiency is improved.
In an embodiment of the present invention, the performing, according to the pooled cluster group, fixed-length alignment on the feature sequence to obtain an aligned sequence includes: extracting cluster length in the pooled cluster; and taking the cluster length as the length of the characteristic sequence to obtain the alignment sequence.
The cluster length refers to the average length of the feature sequence calculated in each pooled cluster.
And S3, extracting the characteristic dimension of the multi-mode characteristic, and aligning the dimension length of the characteristic dimension by utilizing a pre-constructed multi-layer perceptron to obtain an aligned dimension.
The embodiment of the invention is used for adjusting the non-standard dimension length by extracting the feature dimension of the multi-modal feature. The feature dimension refers to features of different categories, namely features of different dimensions, in a certain mode in the multi-mode feature.
In an embodiment of the present invention, the extracting feature dimensions of the multi-modal feature includes: inquiring the feature class of the feature sequence in the multi-mode feature; and identifying the feature dimension of the feature sequence according to the feature category.
Illustratively, if the feature classes of the feature sequence x (0, 1) and the feature sequence y (1, 0) are a and B, the feature sequences are combined according to the feature classes, so that (x, y) is the feature dimension.
Furthermore, the embodiment of the invention aligns the dimension lengths of the characteristic dimensions by utilizing the pre-constructed multi-layer perceptron so as to correct the dimension lengths with different lengths and improve the standardization of medical data. The multi-layer perceptron is an artificial neural network, layers of the multi-layer perceptron are fully connected, wherein the bottommost layer is an input layer, the middle layer is a hidden layer, and the output layer is finally arranged.
In an embodiment of the present invention, the aligning the dimension length of the feature dimension by using a pre-built multi-layer perceptron to obtain an aligned dimension includes: configuring a target length of the feature dimension; selecting a multi-layer neuron in the pre-constructed multi-layer perceptron according to the target length and the characteristic dimension; and performing dimension length alignment operation of the characteristic dimension by using the multi-layer neuron to obtain the pair Ji Weidu.
Illustratively, the multi-layer perceptron is composed of the input layer, the hidden layer and the output layer, each layer is composed of a plurality of neurons, and the neurons of each layer are connected with the neurons of the next layer, which can change the dimension length of the characteristic dimension, for example, the dimension length of the characteristic dimension is 3 dimensions, the dimension length of the target dimension is 4 dimensions, then the neurons of the first layer should select 3 neurons, the neurons of the second layer select 4 neurons, and the neurons of the first layer are connected with the neurons of the second layer through full connection.
S4, determining the aligned multi-modal features in the multi-modal features according to the aligned sequences and the aligned dimensions, performing feature stitching on a plurality of features in the aligned multi-modal features to obtain stitched multi-modal features, and performing feature fusion on the stitched multi-modal features by using a full-connection visual network to obtain fused multi-modal features.
According to the embodiment of the invention, the alignment multi-modal characteristics in the multi-modal characteristics are determined according to the alignment sequence and the alignment dimension, so that the medical characteristics after correction of the format are recombined, and the order of medical data is ensured. The alignment multi-modal feature refers to a feature vector formed by the alignment sequence and the alignment dimension after processing the alignment sequence and the alignment dimension.
In an embodiment of the present invention, the determining the aligned multi-modal feature of the multi-modal features according to the alignment sequence and the alignment dimension is implemented by stitching the alignment sequence and the pair Ji Weidu into the aligned multi-modal feature.
Further, the embodiment of the invention ensures the subsequent fusion processing by performing feature stitching on a plurality of medical features in the aligned multi-modal features, so as to combine the medical feature vector with other medical feature vectors.
In an embodiment of the present invention, the feature stitching is performed on a plurality of features in the aligned multi-modal feature, so as to obtain a stitched multi-modal feature, and the feature stitching is implemented by combining the plurality of features in the aligned multi-modal feature. Illustratively, stitching the a feature vector (1, 0) with the B feature vector (0, 1, 0) may be accomplished by (1,0,0,0,0,1,0,0).
Furthermore, the embodiment of the invention performs feature fusion on the spliced multi-modal features by utilizing the fully-connected visual network so as to fuse the extracted multi-modal features to obtain more hidden medical features, thereby improving the comprehensiveness of medical feature extraction and ensuring the comprehensiveness of medical video retrieval.
In an embodiment of the present invention, the feature fusion of the spliced multi-modal features by using a fully connected visual network to obtain a fused multi-modal feature includes: and carrying out feature column fusion on the spliced multi-mode features by using the following formula to obtain column fusion features:
U *, ＝ *, + 2 σ(W 1 LayerNorm() *, )
wherein U is *, Representing the column fusion feature, X *, A column vector set representing a feature vector set in the stitched multimodal feature, i representing a column of the feature vector set in the stitched multimodal feature, W 1 And W is equal to 2 Respectively represent weights of front and rear layers NLP in the fully connected visual network when feature column fusion is performed, sigma represents bias in the fully connected visual network, layerNorm represents an algorithm for normalizing the spliced multi-modal features, layerNorm () *, Representing a process of normalizing a column vector set of a feature vector set in the spliced multi-modal feature, and X represents the feature vector set in the spliced multi-modal feature;
and carrying out feature row fusion on the column fusion features by using the following formula to obtain the fusion multi-modal features:
Y j,* ＝ j,* + 4 σ(W 3 LayerNorm() j,* )
wherein Y is j,* Representing the fused multi-modal feature, U j,* Representing a row feature vector, W, in the column fusion feature 3 And W is equal to 4 Respectively represent weights of front and rear layers NLP in the fully connected visual network when feature lines are fused, sigma represents bias in the fully connected visual network, layerNorm represents an algorithm for standardizing the spliced multi-modal features, layerNorm () j,* Representation ofAnd (3) a process of normalizing the row feature vectors in the column fusion features, wherein U represents the column fusion features.
S5, obtaining a search text, carrying out feature vector coding on the search text to obtain a coded text vector, calculating cosine similarity between the fused multi-mode features and the coded text vector, and determining a video search result of the search text according to the cosine similarity.
According to the embodiment of the invention, the search text is acquired to be used for matching the medical key text input by the user into the search video with the corresponding medical video. Wherein, the search text refers to text input in a search engine for searching video.
Furthermore, the embodiment of the invention is convenient for matching with the feature vector of the medical video by carrying out feature vector coding on the search text to convert the search text into the medical feature vector form.
In an embodiment of the present invention, the feature vector encoding of the search text to obtain an encoded text vector includes: determining a text category of the retrieved text; constructing an initial code of the text category; matching the search text with the text category; and when the search text is successfully matched with the text category, carrying out feature vector coding on the search text according to the initial coding to obtain a coded text vector.
For example, when the search text is "potato and tomato", it is determined that the search text is of a Chinese character type and no symbol is present, it is determined that the text category of the search text is all Chinese characters, the initial encoding of the text category is constructed by sorting all Chinese characters and encoding them to 0, the constructed initial encoding is used as a top row, the search text is divided into a plurality of single words and used as columns, and when the search text "bean" of the second row is matched with the Chinese character "bean" of the 3 rd column, the encoding of the second row is "001000 … …", and the encoding vector of the search text can be obtained in sequence.
Further, the embodiment of the invention calculates the cosine similarity between the fusion multimodal feature and the coded text vector to be used for determining the medical video to be searched, which is similar to the search text, and searches by utilizing the multidimensional feature, so that the comprehensiveness of searching is improved. The cosine similarity refers to the similarity of the proportions of two variables in all directions (attributes), and the cosine value of the included angle of two vectors in the vector space is used as a measure for measuring the difference between two individuals, and the closer the value is 1, the closer the included angle is to 0 degrees, namely the more similar the two vectors are, so that the vector similarity measurement is realized.
In an embodiment of the present invention, the method for calculating cosine similarity between the fused multi-modal feature and the encoded text vector using the following formula includes:
wherein W is uv And (3) representing cosine similarity between the fused multi-modal feature and the coded text vector, wherein N (u) represents a feature corresponding to u in the fused multi-modal feature set N, and M (v) represents a text vector corresponding to v in the coded text vector set M.
Further, according to the embodiment of the invention, the video retrieval result of the retrieval text is determined according to the cosine similarity, so that the medical video retrieval result is determined according to the calculation result, artificial intelligent automation is realized, and the medical video retrieval efficiency is improved.
In an embodiment of the present invention, the determining the video search result of the search text according to the cosine similarity includes: configuring a similarity threshold of the cosine similarity; judging whether the cosine similarity accords with the similarity threshold value or not; when the cosine similarity accords with the similarity threshold, taking the video to be searched corresponding to the cosine similarity as a video search result of the search text; and when the cosine similarity does not accord with the similarity threshold, the video retrieval result of the retrieval text is retrieval failure.
The cosine similarity can be set to 0.8 or set according to actual conditions.
It can be seen that, in the embodiment of the present invention, by acquiring the video to be searched for performing label distribution on the medical video uploaded in the platform, further, by querying the multi-modal attribute of the video to be searched for implementing the extraction of the medical key feature of the video to be searched from multiple dimensions, the comprehensiveness of the medical information analysis of the video to be searched is improved, further, by selecting the multi-modal network corresponding to the multi-modal attribute for extracting the medical feature of the video by using the different types of feature extraction network, the single-mode of the label of the medical video is reduced, the comprehensiveness of the medical video analysis is improved, further, by extracting the multi-modal feature in the video to be searched by using the multi-modal network, in order to obtain medical key information attached to the video to be searched from a plurality of feature angles and ensure subsequent comprehensive analysis of medical information carried in the video to be searched, the embodiment of the invention performs unified length adjustment on the lengths of different medical feature sequences generated by different medical video time lengths and ensures the uniformity of medical feature data by extracting the feature sequences of the multi-modal features, further, the embodiment of the invention realizes the length control of the feature sequences by constructing fixed cluster groups of the feature sequences and dividing different types of medical feature vectors into different clusters, namely one cluster comprises one type of medical feature vector and ensures classification processing of the medical feature vectors, further, the embodiment of the invention performs average pooling on the fixed cluster groups to determine the average length of the feature sequences, furthermore, the embodiment of the invention aims at the characteristic sequences to adjust the sequences with different lengths into sequences with uniform lengths according to the pooling cluster group, ensures that the sequences are not influenced by disordered medical data when being analyzed, improves the efficiency of medical data analysis, and further aims at the non-standard dimension lengths by extracting the characteristic dimension of the multi-modal characteristics to adjust the non-standard dimension lengths, and aims at correcting the dimension lengths with different lengths by utilizing a pre-constructed multi-layer perceptron to improve the standard degree of the medical data, and determines the aligned multi-modal characteristics in the multi-modal characteristics to recombine the medical characteristics after the correction format according to the aligned sequences and the aligned dimensions, the order of medical data is ensured, further, the embodiment of the invention combines the medical feature vector with other medical feature vectors by performing feature stitching on a plurality of medical features in the aligned multi-modal features, and ensures subsequent fusion processing, further, the embodiment of the invention fuses the stitched multi-modal features by utilizing a fully connected vision network, so as to fuse the extracted multi-modal features to obtain more hidden medical features, promote the comprehensiveness of medical feature extraction, and ensure the comprehensiveness of medical video retrieval, and the embodiment of the invention obtains a retrieval text for matching the medical key text of a search video input by a user with the corresponding medical video, further, the embodiment of the invention encodes the feature vector by using the retrieval text, the method and the device are used for converting the search text into a medical feature vector form so as to be convenient for matching with the feature vector of the medical video, further, the embodiment of the invention realizes artificial intelligent automation by calculating cosine similarity between the fused multi-modal feature and the coded text vector so as to be used for determining the medical video to be searched which is close to the search text, and improves the comprehensiveness of the search by searching through the multi-dimensional feature. Therefore, the multi-mode-based video retrieval method provided by the embodiment of the invention can realize automatic indexing and retrieval of medical videos and improve the comprehensiveness of the whole medical video retrieval process.
As shown in fig. 2, a functional block diagram of the multi-mode video search device according to the present invention is shown.
The multi-modality-based video retrieval apparatus 100 of the present invention may be installed in an electronic device. Depending on the implemented functionality, the multi-modality based video retrieval device may include a feature extraction module 101, a fixed length alignment module 102, a dimension alignment module 103, a feature fusion module 104, and a retrieval determination module 105. The module according to the invention, which may also be referred to as a unit, refers to a series of computer program segments, which are stored in the memory of the electronic device, capable of being executed by the processor of the electronic device and of performing a fixed function.
In the present embodiment, the functions concerning the respective modules/units are as follows:
the feature extraction module 101 is configured to obtain a video to be retrieved, query a multi-modal attribute of the video to be retrieved, select a multi-modal network corresponding to the multi-modal attribute, and extract multi-modal features in the video to be retrieved by using the multi-modal network;
the fixed-length alignment module 102 is configured to extract a feature sequence of the multi-mode feature, construct a fixed cluster group of the feature sequence, perform mean pooling on the fixed cluster group to obtain a pooled cluster group, and perform fixed-length alignment on the feature sequence according to the pooled cluster group to obtain an alignment sequence;
The dimension alignment module 103 is configured to extract feature dimensions of the multi-modal feature, and align dimension lengths of the feature dimensions by using a pre-built multi-layer perceptron to obtain a pair Ji Weidu;
the feature fusion module 104 is configured to determine an aligned multi-modal feature of the multi-modal features according to the aligned sequence and the aligned dimension, perform feature stitching on a plurality of features in the aligned multi-modal feature to obtain a stitched multi-modal feature, and perform feature fusion on the stitched multi-modal feature by using a fully connected visual network to obtain a fused multi-modal feature;
the search determining module 105 is configured to obtain a search text, perform feature vector encoding on the search text to obtain an encoded text vector, calculate cosine similarity between the fused multi-modal feature and the encoded text vector, and determine a video search result of the search text according to the cosine similarity.
In detail, the modules in the multi-mode-based video retrieval device 100 in the embodiment of the present invention use the same technical means as the multi-mode-based video retrieval method described in fig. 1, and can produce the same technical effects, which are not described herein.
Fig. 3 is a schematic structural diagram of an electronic device 1 implementing a multi-mode video retrieval method according to the present invention.
The electronic device 1 may comprise a processor 10, a memory 11, a communication bus 12 and a communication interface 13, and may further comprise a computer program stored in the memory 11 and executable on the processor 10, such as a multimodal-based video retrieval program.
The processor 10 may be formed by an integrated circuit in some embodiments, for example, a single packaged integrated circuit, or may be formed by a plurality of integrated circuits packaged with the same function or different functions, including one or more central processing units (Central Processing Unit, CPU), a microprocessor, a digital processing chip, a graphics processor, a combination of various control chips, and so on. The processor 10 is a Control Unit (Control Unit) of the electronic device 1, connects the respective components of the entire electronic device 1 using various interfaces and lines, executes various functions of the electronic device 1 and processes data by running or executing programs or modules stored in the memory 11 (for example, executing a multi-modality-based video retrieval program or the like), and calls data stored in the memory 11.
The memory 11 includes at least one type of readable storage medium including flash memory, a removable hard disk, a multimedia card, a card type memory (e.g., SD or DX memory, etc.), a magnetic memory, a magnetic disk, an optical disk, etc. The memory 11 may in some embodiments be an internal storage unit of the electronic device 1, such as a removable hard disk of the electronic device 1. The memory 11 may in other embodiments also be an external storage device of the electronic device 1, such as a plug-in mobile hard disk, a Smart Media Card (SMC), a Secure Digital (SD) Card, a Flash memory Card (Flash Card) or the like, which are provided on the electronic device 1. Further, the memory 11 may also include both an internal storage unit and an external storage device of the electronic device 1. The memory 11 may be used not only for storing application software installed in the electronic device 1 and various types of data, such as codes based on a multi-modality video search program, but also for temporarily storing data that has been output or is to be output.
The communication bus 12 may be a peripheral component interconnect standard (Peripheral Component Interconnect, PCI) bus, or an extended industry standard architecture (Extended Industry Standard Architecture, EISA) bus, among others. The bus may be classified as an address bus, a data bus, a control bus, etc. The bus is arranged to enable a connection communication between the memory 11 and at least one processor 10 etc.
The communication interface 13 is used for communication between the electronic device 1 and other devices, including a network interface and an employee interface. Optionally, the network interface may comprise a wired interface and/or a wireless interface (e.g. WI-FI interface, bluetooth interface, etc.), typically used to establish a communication connection between the electronic device 1 and other electronic devices 1. The employee interface may be a Display (Display), an input unit such as a Keyboard (Keyboard), or alternatively a standard wired interface, a wireless interface. Alternatively, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode) touch, or the like. The display may also be referred to as a display screen or display unit, as appropriate, for displaying information processed in the electronic device 1 and for displaying a visual staff interface.
Fig. 3 shows only an electronic device 1 with components, it being understood by a person skilled in the art that the structure shown in fig. 3 does not constitute a limitation of the electronic device 1, and may comprise fewer or more components than shown, or may combine certain components, or may be arranged in different components.
For example, although not shown, the electronic device 1 may further include a power source (such as a battery) for supplying power to each component, and preferably, the power source may be logically connected to the at least one processor 10 through a power management device, so that functions of charge management, discharge management, power consumption management, and the like are implemented through the power management device. The power supply may also include one or more of any of a direct current or alternating current power supply, recharging device, power failure detection circuit, power converter or inverter, power status indicator, etc. The electronic device 1 may further include various sensors, bluetooth modules, wi-Fi modules, etc., which will not be described herein.
It should be understood that the embodiments described are for illustrative purposes only and are not limited in scope by this configuration.
The multimodal video retrieval program stored in the memory 11 in the electronic device 1 is a combination of a plurality of computer programs which, when run in the processor 10, may implement:
acquiring a video to be searched, inquiring the multi-modal attribute of the video to be searched, selecting a multi-modal network corresponding to the multi-modal attribute, and extracting multi-modal characteristics in the video to be searched by utilizing the multi-modal network;
Extracting a characteristic sequence of the multi-mode characteristic, constructing a fixed cluster group of the characteristic sequence, carrying out mean pooling on the fixed cluster group to obtain a pooled cluster group, and carrying out fixed-length alignment on the characteristic sequence according to the pooled cluster group to obtain an alignment sequence;
extracting feature dimensions of the multi-mode features, and aligning dimension lengths of the feature dimensions by utilizing a pre-constructed multi-layer perceptron to obtain pairs Ji Weidu;
determining an aligned multi-modal feature of the multi-modal features according to the aligned sequence and the aligned dimension, performing feature stitching on a plurality of features in the aligned multi-modal features to obtain a stitched multi-modal feature, and performing feature fusion on the stitched multi-modal feature by using a full-connection visual network to obtain a fused multi-modal feature;
and obtaining a search text, carrying out feature vector coding on the search text to obtain a coded text vector, calculating cosine similarity between the fused multi-mode features and the coded text vector, and determining a video search result of the search text according to the cosine similarity.
In particular, the specific implementation method of the processor 10 on the computer program may refer to the description of the relevant steps in the corresponding embodiment of fig. 1, which is not repeated herein.
Further, the integrated modules/units of the electronic device 1 may be stored in a non-volatile computer readable storage medium if implemented in the form of software functional units and sold or used as a stand alone product. The computer readable storage medium may be volatile or nonvolatile. For example, the computer readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a U disk, a removable hard disk, a magnetic disk, an optical disk, a computer Memory, a Read-Only Memory (ROM).
The present invention also provides a computer readable storage medium storing a computer program which, when executed by a processor of an electronic device 1, may implement:
acquiring a video to be searched, inquiring the multi-modal attribute of the video to be searched, selecting a multi-modal network corresponding to the multi-modal attribute, and extracting multi-modal characteristics in the video to be searched by utilizing the multi-modal network;
extracting a characteristic sequence of the multi-mode characteristic, constructing a fixed cluster group of the characteristic sequence, carrying out mean pooling on the fixed cluster group to obtain a pooled cluster group, and carrying out fixed-length alignment on the characteristic sequence according to the pooled cluster group to obtain an alignment sequence;
Extracting feature dimensions of the multi-mode features, and aligning dimension lengths of the feature dimensions by utilizing a pre-constructed multi-layer perceptron to obtain pairs Ji Weidu;
determining an aligned multi-modal feature of the multi-modal features according to the aligned sequence and the aligned dimension, performing feature stitching on a plurality of features in the aligned multi-modal features to obtain a stitched multi-modal feature, and performing feature fusion on the stitched multi-modal feature by using a full-connection visual network to obtain a fused multi-modal feature;
and obtaining a search text, carrying out feature vector coding on the search text to obtain a coded text vector, calculating cosine similarity between the fused multi-mode features and the coded text vector, and determining a video search result of the search text according to the cosine similarity.
In the several embodiments provided in the present invention, it should be understood that the disclosed apparatus, device and method may be implemented in other manners. For example, the above-described apparatus embodiments are merely illustrative, and for example, the division of the modules is merely a logical function division, and there may be other manners of division when actually implemented.
The modules described as separate components may or may not be physically separate, and components shown as modules may or may not be physical units, may be located in one place, or may be distributed over multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
In addition, each functional module in the embodiments of the present invention may be integrated in one processing unit, or each unit may exist alone physically, or two or more units may be integrated in one unit. The integrated units can be realized in a form of hardware or a form of hardware and a form of software functional modules.
It will be evident to those skilled in the art that the invention is not limited to the details of the foregoing illustrative embodiments, and that the present invention may be embodied in other specific forms without departing from the spirit or essential characteristics thereof.
The present embodiments are, therefore, to be considered in all respects as illustrative and not restrictive, the scope of the invention being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein. Any reference signs in the claims shall not be construed as limiting the claim concerned.
The embodiment of the invention can acquire and process the related data based on the artificial intelligence technology. Among these, artificial intelligence (Artificial Intelligence, AI) is the theory, method, technique and application system that uses a digital computer or a digital computer-controlled machine to simulate, extend and extend human intelligence, sense the environment, acquire knowledge and use knowledge to obtain optimal results.
Furthermore, it is evident that the word "comprising" does not exclude other elements or steps, and that the singular does not exclude a plurality. A plurality of units or means recited in the system claims can also be implemented by means of software or hardware by means of one unit or means. The terms second, etc. are used to denote a name, but not any particular order.
Finally, it should be noted that the above-mentioned embodiments are merely for illustrating the technical solution of the present invention and not for limiting the same, and although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that modifications and equivalents may be made to the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention.
Claims (10)
1. A multi-modality based video retrieval method, the method comprising:
Acquiring a video to be searched, inquiring the multi-modal attribute of the video to be searched, selecting a multi-modal network corresponding to the multi-modal attribute, and extracting multi-modal characteristics in the video to be searched by utilizing the multi-modal network;
extracting a characteristic sequence of the multi-mode characteristic, constructing a fixed cluster group of the characteristic sequence, carrying out mean pooling on the fixed cluster group to obtain a pooled cluster group, and carrying out fixed-length alignment on the characteristic sequence according to the pooled cluster group to obtain an alignment sequence;
extracting feature dimensions of the multi-mode features, and aligning dimension lengths of the feature dimensions by utilizing a pre-constructed multi-layer perceptron to obtain pairs Ji Weidu;
determining an aligned multi-modal feature of the multi-modal features according to the aligned sequence and the aligned dimension, performing feature stitching on a plurality of features in the aligned multi-modal features to obtain a stitched multi-modal feature, and performing feature fusion on the stitched multi-modal feature by using a full-connection visual network to obtain a fused multi-modal feature;
and obtaining a search text, carrying out feature vector coding on the search text to obtain a coded text vector, calculating cosine similarity between the fused multi-mode features and the coded text vector, and determining a video search result of the search text according to the cosine similarity.
2. The multi-modality-based video retrieval method as recited in claim 1, wherein said querying the multi-modality attributes of the video to be retrieved includes:
performing playing processing on the video to be searched to obtain a playing video;
carrying out integrity check on the video elements of the playing video;
when the integrity check of the video elements of the played video is successful, determining the multi-modal attribute of the video to be retrieved;
when the integrity check of the video element of the played video fails, extracting a video failure part of which the integrity check of the video fails;
and determining the multi-modal attribute of the video to be retrieved according to the video failure part.
3. The method for retrieving video based on multiple modes as recited in claim 1, wherein said selecting a multiple mode network corresponding to said multiple mode attribute comprises:
identifying a feature objective of the multimodal property;
acquiring a neural network structure, and inquiring a network purpose of the neural network structure;
matching the network purpose with the characteristic purpose;
and when the network purpose is successfully matched with the characteristic purpose, determining the multi-mode network corresponding to the multi-mode attribute.
4. The multi-modal based video retrieval method as set forth in claim 1, wherein said aligning the dimension length of the feature dimension with a pre-built multi-layer perceptron to obtain an aligned dimension includes:
configuring a target length of the feature dimension;
selecting a multi-layer neuron in the pre-constructed multi-layer perceptron according to the target length and the characteristic dimension;
and performing dimension length alignment operation of the characteristic dimension by using the multi-layer neuron to obtain the pair Ji Weidu.
5. The multi-modal-based video retrieval method as set forth in claim 1, wherein the feature fusion of the spliced multi-modal features using a fully connected visual network to obtain fused multi-modal features includes:
and carrying out feature column fusion on the spliced multi-mode features by using the following formula to obtain column fusion features:
U *，i ＝X *，i +W 2 σ(W 1 LayerNorm(X) *，i )
wherein U is *，i Representing the column fusion feature, X *，i A column vector set representing a feature vector set in the stitched multimodal feature, i representing a column of the feature vector set in the stitched multimodal feature, W 1 And W is equal to 2 Respectively representing the weights of front and rear layers NLP in the fully connected visual network when feature sequence fusion is carried out, sigma represents the bias in the fully connected visual network, and LayerNorm represents the splicing Algorithm for normalizing multimodal features LayerNorm (X) *，i Representing a process of normalizing a column vector set of a feature vector set in the spliced multi-modal feature, and X represents the feature vector set in the spliced multi-modal feature;
and carrying out feature row fusion on the column fusion features by using the following formula to obtain the fusion multi-modal features:
Y j，* ＝U j，* +W 4 σ(W 3 LayerNorm(U) j，* )
wherein Y is j，* Representing the fused multi-modal feature, U j，* Representing a row feature vector, W, in the column fusion feature 3 And W is equal to 4 Representing weights of front and rear layers NLP in the fully connected visual network when feature lines are fused, respectively, sigma representing bias in the fully connected visual network, layerNorm representing an algorithm for normalizing the spliced multi-modal features, layerNorm (U) j，* Representing the process of normalizing the row feature vectors in the column fusion feature, and U represents the column fusion feature.
6. The multi-modal based video retrieval method as claimed in claim 1, wherein said feature vector encoding of the retrieved text to obtain an encoded text vector includes:
determining a text category of the retrieved text;
constructing an initial code of the text category;
matching the search text with the text category;
And when the search text is successfully matched with the text category, carrying out feature vector coding on the search text according to the initial coding to obtain a coded text vector.
7. The multi-modality based video retrieval method as recited in any of claims 1 to 5, wherein said calculating cosine similarity between said fused multi-modality features and said encoded text vector, includes:
calculating cosine similarity between the fused multi-modal feature and the encoded text vector using the formula
Wherein W is uv And (3) representing cosine similarity between the fused multi-modal feature and the coded text vector, wherein N (u) represents a feature corresponding to u in the fused multi-modal feature set N, and M (v) represents a text vector corresponding to v in the coded text vector set M.
8. A multi-modality based video retrieval apparatus, the apparatus comprising:
the feature extraction module is used for acquiring the video to be searched, inquiring the multi-modal attribute of the video to be searched, selecting a multi-modal network corresponding to the multi-modal attribute, and extracting multi-modal features in the video to be searched by utilizing the multi-modal network;
The fixed-length alignment module is used for extracting the characteristic sequences of the multi-mode characteristics, constructing fixed cluster groups of the characteristic sequences, carrying out mean pooling on the fixed cluster groups to obtain pooled cluster groups, and carrying out fixed-length alignment on the characteristic sequences according to the pooled cluster groups to obtain alignment sequences;
the dimension alignment module is used for extracting the characteristic dimension of the multi-mode characteristic, and aligning the dimension length of the characteristic dimension by utilizing a pre-constructed multi-layer perceptron to obtain a pair Ji Weidu;
the feature fusion module is used for determining the aligned multi-modal features in the multi-modal features according to the aligned sequences and the aligned dimensions, performing feature stitching on a plurality of features in the aligned multi-modal features to obtain stitched multi-modal features, and performing feature fusion on the stitched multi-modal features by using a full-connection visual network to obtain fused multi-modal features;
the retrieval determining module is used for obtaining a retrieval text, carrying out feature vector coding on the retrieval text to obtain a coded text vector, calculating cosine similarity between the fused multi-mode features and the coded text vector, and determining a video retrieval result of the retrieval text according to the cosine similarity.
9. An electronic device, the electronic device comprising:
at least one processor; the method comprises the steps of,
a memory communicatively coupled to the at least one processor; wherein,,
the memory stores a computer program executable by the at least one processor to enable the at least one processor to perform the multimodal video retrieval method of any of claims 1 to 7.
10. A computer readable storage medium storing a computer program, wherein the computer program when executed by a processor implements the multi-modality based video retrieval method according to any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310672336.1A CN116628263A (en) | 2023-06-07 | 2023-06-07 | Video retrieval method and device based on multiple modes, electronic equipment and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310672336.1A CN116628263A (en) | 2023-06-07 | 2023-06-07 | Video retrieval method and device based on multiple modes, electronic equipment and storage medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116628263A true CN116628263A (en) | 2023-08-22 |
Family
ID=87591894
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310672336.1A Pending CN116628263A (en) | 2023-06-07 | 2023-06-07 | Video retrieval method and device based on multiple modes, electronic equipment and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116628263A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117370933A (en) * | 2023-10-31 | 2024-01-09 | 中国人民解放军总医院 | Multi-mode unified feature extraction method, device, equipment and medium |
-
2023
- 2023-06-07 CN CN202310672336.1A patent/CN116628263A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117370933A (en) * | 2023-10-31 | 2024-01-09 | 中国人民解放军总医院 | Multi-mode unified feature extraction method, device, equipment and medium |
| CN117370933B (en) * | 2023-10-31 | 2024-05-07 | 中国人民解放军总医院 | Multi-mode unified feature extraction method, device, equipment and medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112199375B (en) | Cross-modal data processing method and device, storage medium and electronic device | |
| US9280560B1 (en) | Scalable image matching | |
| CN110502664A (en) | Video tab indexes base establishing method, video tab generation method and device | |
| CN113590854B (en) | Data processing method, data processing equipment and computer readable storage medium | |
| CN113656690B (en) | Product recommendation method and device, electronic equipment and readable storage medium | |
| CN113033507B (en) | Scene recognition method and device, computer equipment and storage medium | |
| CN116628263A (en) | Video retrieval method and device based on multiple modes, electronic equipment and storage medium | |
| CN115018588A (en) | Product recommendation method and device, electronic equipment and readable storage medium | |
| CN114419391A (en) | Target image identification method and device, electronic equipment and readable storage medium | |
| CN116450829A (en) | Medical text classification method, device, equipment and medium | |
| CN113821602B (en) | Automatic answering method, device, equipment and medium based on image-text chat record | |
| CN116975340A (en) | Information retrieval method, apparatus, device, program product, and storage medium | |
| CN113591881B (en) | Intention recognition method and device based on model fusion, electronic equipment and medium | |
| CN114168768A (en) | Image retrieval method and related equipment | |
| CN114329004A (en) | Digital fingerprint generation method, digital fingerprint generation device, data push method, data push device and storage medium | |
| CN113254687A (en) | Image retrieval and image quantification model training method, device and storage medium | |
| Kapadia et al. | Improved CBIR system using Multilayer CNN | |
| CN116630712A (en) | Information classification method and device based on modal combination, electronic equipment and medium | |
| CN116303677A (en) | Measurement method, device, equipment and storage medium based on data distribution balance degree | |
| CN114219367A (en) | User scoring method, device, equipment and storage medium | |
| CN114677526A (en) | Image classification method, device, equipment and medium | |
| CN113505698A (en) | Face recognition method, device, equipment and medium based on counterstudy | |
| CN113361274B (en) | Intent recognition method and device based on label vector, electronic equipment and medium | |
| CN116737996A (en) | Multi-mode video retrieval method, device, equipment and medium based on multi-encoder | |
| CN116485949A (en) | Image-text video method, device and equipment based on multi-modal neural network model |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |