Video text screening method, device and electronic equipment

Abstract

The present application discloses a video text screening method, device and electronic device, the method comprising: obtaining video texts and video text position information respectively corresponding to a plurality of video frames of a target video; performing text correction processing on the video text to obtain a video text correction with coherent text semantics; determining N video text lists for the plurality of video frames based on the video correction text and the video text position information corresponding to each video frame, where N is an integer greater than or equal to 1; screening out K target video text lists from the N video text lists based on the semantic association between the N video text lists and a video title corresponding to the target video, and generating a target video text corresponding to the target video according to the video texts corresponding to the K target video text lists, where K is an integer less than or equal to N.

Description

Technical Field

[0001] The present application belongs to the field of communication technology, and in particular, relates to a video text screening method, device and electronic equipment. Background Art

[0002] In the video classification scenario, it is often necessary to rely on the text semantic features in the video content to understand the video. For example, the subtitle text in the film and television video can directly reflect the content of the character's dialogue. At present, the text content in the video can be extracted by optical character recognition (OCR) technology, and then the text content can be semantically understood by using natural language processing (NLP) related technologies. The video visual information and audio features are combined to achieve video understanding for video classification. However, the text content contained in the video is often rich and diverse. In addition to the subtitle text that plays a positive role in video understanding, there may also be interference text such as advertising text and author nicknames. In the process of optical character recognition, it is difficult to directly judge the interference information, resulting in the final extracted video text containing a large amount of redundant text. If it is directly used for visual semantic understanding, the effect is very unsatisfactory. Therefore, how to screen the video OCR text is an important part of video understanding.

[0003] Currently, text filtering is performed using the optical features of text boxes, such as position features (coordinates, etc.), geometric features (area of ​​text boxes), etc. These solutions can usually only handle certain types of videos. For example, in film and television videos, subtitles are usually placed at the bottom of the video. In scenarios where video text is used more casually, the use of optical features of text boxes for text filtering is not universal enough. Summary of the invention

[0004] The purpose of the embodiments of the present application is to provide a video text screening method, device and electronic device to solve the problem of insufficient versatility in the prior art when using the optical features of the text box to perform text screening.

[0005] In a first aspect, an embodiment of the present application provides a video text screening method, comprising:

[0006] Obtaining video texts and video text position information corresponding to multiple video frames of a target video respectively;

[0007] Performing text correction processing on the video text to obtain a video-corrected text with coherent text semantics;

[0008] For the multiple video frames, based on the video correction text and the video text position information corresponding to each video frame, determine N video text lists, where N is an integer greater than or equal to 1;

[0009] Based on the semantic association between the N video text lists and the video titles corresponding to the target video, K target video text lists are screened out from the N video text lists, and a target video text corresponding to the target video is generated according to the video texts corresponding to the K target video text lists, where K is an integer less than or equal to N.

[0010] In a second aspect, an embodiment of the present application provides a video text screening device, comprising:

[0011] A first acquisition module is used to acquire video texts and video text position information respectively corresponding to a plurality of video frames of a target video;

[0012] A second acquisition module is used to perform text correction processing on the video text to obtain a video-corrected text with coherent text semantics;

[0013] A determination module, configured to determine, for the plurality of video frames, N video text lists based on the video correction text and the video text position information corresponding to each video frame, where N is an integer greater than or equal to 1;

[0014] A processing module is used to screen out K target video text lists from the N video text lists based on the semantic association between the N video text lists and the video titles corresponding to the target video, and generate a target video text corresponding to the target video according to the video texts corresponding to the K target video text lists, where K is an integer less than or equal to N.

[0015] In a third aspect, an embodiment of the present application provides an electronic device, which includes a processor and a memory, wherein the memory stores programs or instructions that can be run on the processor, and when the program or instructions are executed by the processor, the steps of the method described in the first aspect are implemented.

[0016] In a fourth aspect, an embodiment of the present application provides a readable storage medium, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the method described in the first aspect are implemented.

[0017] In a fifth aspect, an embodiment of the present application provides a chip, comprising a processor and a communication interface, wherein the communication interface is coupled to the processor, and the processor is used to run a program or instruction to implement the method described in the first aspect.

[0018] In a sixth aspect, an embodiment of the present application provides a computer program product, which is stored in a storage medium and is executed by at least one processor to implement the method described in the first aspect.

[0019] In an embodiment of the present application, after obtaining the video texts corresponding to multiple video frames of the target video, text correction is performed on the video text to ensure the integrity and coherence of the text semantics; after the text correction is performed, for multiple video frames, based on the video correction text and the video text position information corresponding to each video frame, N video text lists are determined, and a video text list including video texts with coherent text semantics can be obtained; by screening out the target video text list based on the semantic correlation between the video text list and the video title, and then generating the target video text corresponding to the target video, it is possible to retain video texts with high semantic correlation with the video title in a variety of video text usage scenarios, thereby achieving effective screening of video texts. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a schematic diagram of a video text screening method provided in an embodiment of the present application;

[0021] Figure 2 is an example diagram of text segmentation based on semantic coherence provided in an embodiment of the present application;

[0022] Figure 3 is an example diagram of text correction provided by an embodiment of the present application;

[0023] Figure 4a is a schematic diagram of a first video frame provided in an embodiment of the present application;

[0024] Figure 4b is a schematic diagram of a second video frame provided in an embodiment of the present application;

[0025] Figure 5 is an implementation flow chart of the video text screening method provided in the embodiment of the present application;

[0026] Figure 6 is a schematic diagram of a video text screening device provided in an embodiment of the present application;

[0027] Figure 7 It is one of the schematic block diagrams of the electronic device provided in the embodiment of the present application;

[0028] Figure 8 This is the second schematic block diagram of the electronic device provided in the embodiment of the present application. DETAILED DESCRIPTION

[0029] The following will be combined with the drawings in the embodiments of the present application to clearly describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments in the present application belong to the scope of protection of this application.

[0030] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here, and the objects distinguished by "first", "second", etc. are generally of one type, and the number of objects is not limited. For example, the first object can be one or more. In addition, "and / or" in the specification and claims represents at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0031] The video text screening method provided by the embodiment of the present application is described in detail below through specific embodiments and their application scenarios in conjunction with the accompanying drawings.

[0032] See also Figure 1 As shown, the embodiment of the present application provides a video text screening method, comprising:

[0033] Step 101: Obtain video texts and video text position information corresponding to a plurality of video frames of a target video.

[0034] The video text screening method provided in the embodiment of the present application, for a target video including a video text, first determines multiple video frames of the target video, and the multiple video frames can be at least part of all the video frames included in the target video. For each of the multiple video frames, the video text is extracted by optical character recognition technology, and the corresponding video text position information is determined for the extracted video text, so as to obtain the video text corresponding to the video frame and the video text position information corresponding to the video text. Among them, the video text position information can be understood as the position information of the text box including the video text, such as obtaining the coordinate information of the upper left corner and the coordinate information of the lower right corner of the text box including the video text, and then obtaining the position information of the text box.

[0035] When multiple video frames of the target video are determined, the target video may be subjected to frame cutting to obtain multiple video frames. For each video frame, after extracting the video text corresponding to the video frame and obtaining the video text position information corresponding to the extracted video text, the video text and the video text position information are saved.

[0036] When performing frame capture processing on the target video, the frame capture frequency can be set, such as a frame capture frequency of 0.5s per frame, that is, a frame capture processing is performed every 0.5s. After performing frame capture processing to obtain the video frame, the moment information of the video frame can be saved so that multiple video frames can be sorted based on the moment information of the video frame. By performing frame capture processing on the target video, it is convenient to extract the video text in the video frame, and by saving the video text position information, it is convenient to make reasonable use of the text position information later.

[0037] When obtaining the video text and video text position information corresponding to each video frame for each of the multiple video frames, the information can be obtained sequentially in chronological order. For example, based on the moment information of the video frames (the time information corresponding to the frame capture), the multiple video frames are sorted in the order of increasing moment information, and then the video text and video text position information corresponding to the video frames are sequentially obtained for the sorted multiple video frames. Of course, the information can also be obtained in parallel, which will not be further elaborated here.

[0038] Step 102: Perform text correction processing on the video text to obtain a video-corrected text with coherent text semantics.

[0039] When extracting video text based on optical character recognition technology, there may be problems such as video text recognition errors and video text including redundant information, which makes it difficult to form actual semantics. Therefore, after obtaining the video text and video text position information corresponding to each video frame, the extracted video text can be corrected for each video frame to obtain a video corrected text with coherent text semantics.

[0040] Step 103: for the multiple video frames, based on the video correction text and the video text position information corresponding to each video frame, determine N video text lists, where N is an integer greater than or equal to 1.

[0041] After completing the text correction process for each video frame, N video text lists can be determined for multiple video frames based on the video correction text and the video text position information corresponding to each video frame. Specifically, based on the text semantic coherence of adjacent video frames, the text semantic coherence of a single video frame, and the video text position information corresponding to each video frame, the video correction text in multiple video frames is processed to determine N video text lists, and the processing here includes but is not limited to the screening and combination of video texts. Among them, the video texts that are associated with the previous and next frames are stored in one video text list, and the video texts in different video text lists have no semantic association.

[0042] Step 104: based on the semantic association between the N video text lists and the video titles corresponding to the target video, K target video text lists are screened out from the N video text lists, and a target video text corresponding to the target video is generated according to the video texts corresponding to the K target video text lists, where K is an integer less than or equal to N.

[0043] After determining N video text lists, the semantic relevance between the video title corresponding to the target video and each video text list can be calculated, and K target video text lists can be screened out from the N video text lists based on the calculated semantic relevance. That is, K target video text lists whose semantic relevance to the video title meets a preset condition can be screened out from the N video text lists. Among them, satisfying the preset condition can be that the semantic relevance is greater than a preset relevance threshold.

[0044] K is an integer less than or equal to N. When K is equal to N, the semantic associations between the N video text lists and the video title all meet the preset conditions, and the N video text lists are all target video text lists. When the value of K is less than N, the semantic associations between K video text lists and the video title among the N video text lists meet the preset conditions, and the K video text lists are target video text lists. For example, when the value of N is 10, the semantic associations between 5 (the value of K) video text lists among the N video text lists and the video title are greater than the preset association threshold, and these 5 video text lists are determined as target video text lists.

[0045] After screening out K target video text lists, the target video text corresponding to the target video can be generated based on the video texts corresponding to the K target video text lists. The video text with a high semantic relevance to the video title can be retained to achieve effective screening of the video text.

[0046] In the above implementation process of the present application, after obtaining the video texts corresponding to multiple video frames of the target video, text correction is performed on the video texts to ensure the integrity and coherence of the text semantics; after the text correction, for multiple video frames, based on the video correction text and the video text position information corresponding to each video frame, N video text lists are determined, and a video text list including video texts with coherent text semantics can be obtained; by screening out the target video text list based on the semantic correlation between the video text list and the video title, and then generating the target video text corresponding to the target video, it is possible to retain the video text with a high semantic correlation with the video title in a variety of video text usage scenarios, thereby achieving effective screening of the video text.

[0047] Optionally, step 102 performs text correction processing on the video text to obtain a video corrected text with coherent text semantics, including:

[0048] For each video frame, segment each video text corresponding to the video frame based on semantic coherence;

[0049] In a case where the video text includes a target text segment and at least one first text segment to be corrected, taking the target text segment as the center, performing text correction on at least one first text segment to obtain a video correction text corresponding to the video text;

[0050] In a case where the video text only includes the target text segment, determining that the video text does not need to be corrected;

[0051] In a case where the video text does not include the target text segment, includes a second text segment and at least one first text segment to be corrected, performing text correction on at least one of the first text segments based on semantic coherence to obtain a video correction text corresponding to the video text;

[0052] The second text segment is a correct text segment, and the target text segment is a text segment with semantic coherence and the longest character length.

[0053] When performing text correction processing on a single video frame (each of multiple video frames), each video text corresponding to the current video frame may be firstly cleaned. In this embodiment, one video frame may include at least one video text.

[0054] Cleaning the video text at least includes deleting special symbols and word segmentation. Among them, special symbols (which may include symbols other than Chinese and English words and numbers, such as emoticons and punctuation marks) will affect the semantics of the video text. By deleting special symbols, the semantics of the video text can be ensured not to be affected; after deleting special symbols, the video text can be segmented using word segmentation software (such as Jieba word segmentation) to ensure the fluency of the subsequent calculation of the video text.

[0055] Among them, for multiple video frames, the cleaning process can be performed in parallel or sequentially during text cleaning. The text segments can be segmented after the text cleaning of multiple video frames is completed, or each video frame can be independent of each other. For a video frame, the text segments can be segmented after the text cleaning is completed without paying attention to the situation of other video frames. For a video frame, the segmentation can be performed after the cleaning of all video texts is completed, or the video text can be segmented after the cleaning of a certain video text is completed.

[0056] For each video frame, each video text corresponding to the current video frame needs to be segmented based on semantic coherence. The text segmentation principle of the video text corresponding to each video frame is the same. The following example illustrates the process of text segmentation for each video text corresponding to a video frame.

[0057] For example, the current video frame extracts 4 video texts based on optical character recognition technology, namely video text 1 "XX Satellite TV", video text 2 "Gao Qian", video text 3 "XX Food" and video text 4 "The store Yin Zi Tan should take corresponding responsibilities and make rectifications". After word segmentation of the video texts, the following content can be obtained: video text 1 "XX Satellite TV", video text 2 "Gao Qian", video text 3 "XX Food" and video text 4 "The store Yin Zi Tan should take corresponding responsibilities and make rectifications".

[0058] Then, based on semantic coherence, the four video texts were segmented into text segments, where video text 1 was segmented into one text segment: XX Satellite TV; video text 2 was segmented into two text segments: Gao, Qian; video text 3 was segmented into one text segment: XX Food; video text 4 was segmented into five text segments: store, sound, son, tank, and take corresponding responsibilities and make rectifications.

[0059] The following takes video text 4 as an example to introduce the segmentation process. Before introducing the segmentation process, the semantic coherence measurement formula is first introduced. The semantic coherence measurement formula is formula 1:

[0060]

[0061] Among them, P(S) represents the positive smoothness measurement function, represents the reverse fluency measurement function, R represents the number of characters corresponding to the text segment, and S represents the text segment. If the fluency measurement is only performed according to the forward word order, calculation deviation may occur. Therefore, the fluency measurement based on the reverse word order is also required. The expression is as follows.

[0062] P(S)=p(w1)p(w2|w1)…p(w R |w1,…w R-1 )

[0063]

[0064] Among them, p(w1) represents the probability that the first character in the text segment appears in the target text library, and p(w2|w1) represents the probability that the second character in the text segment appears in the target text library when the first character in the text segment appears in the target text library, which is equivalent to the ratio of the probability that the first character and the second character appear in the target text library at the same time to the probability that the first character appears in the target text library. R |w1,…w R-1 ) represents the probability that the Rth character in the text segment appears in the target text library when the first character to the (R-1)th character in the text segment appears in the target text library, which is equivalent to the ratio of the probability that the first character to the Rth character appear in the target text library at the same time to the probability that the first character to the (R-1)th character appear in the target text library. P(S) represents the product of the probabilities in the above formula, and the value of P(S) is less than 1.

[0065] Correspondingly, p(w R ) represents the probability of the Rth character in the text segment appearing in the target text library, p(w R-1 |w R ) represents the probability that the (R-1)th character in the text segment appears in the target text library when the Rth character in the text segment appears in the target text library, which is equivalent to the ratio of the probability that the Rth character and the (R-1)th character appear in the target text library at the same time to the probability that the Rth character appears in the target text library, p(w1|w2,…w R ) represents the probability that the first character in the text segment appears in the target text library when the second to R-th characters in the text segment appear in the target text library, which is equivalent to the ratio of the probability that the first to R-th characters appear in the target text library at the same time to the probability that the second to R-th characters appear in the target text library. represents the multiplication of the probabilities in the above formula, The value of is less than 1.

[0066] Since the value of P(S) is less than 1, If the value of is less than 1, Less than 1, then The value of is from 0 to negative infinity, and the corresponding The value of can be from 0 to positive infinity.

[0067] The following is a specific example to illustrate P(S) and The calculation process is explained.

[0068] Among them, P(extremely affectionate) = p(extremely)p(affectionate|extremely)p(love|extremely affectionate). If only the forward word order is used for smoothness measurement, calculation deviation may occur. For example, for "extremely affectionate", the probability corresponding to "extremely affectionate" is small, the smoothness is low, and the forward semantics is incomplete, so calculation deviation is likely to occur. Therefore, smoothness measurement can also be based on the reverse word order; Then merge the forward and reverse smoothness measurements.

[0069] The semantic coherence measurement formula has been introduced above. Next, taking video text 4 as an example, the segmentation process will be introduced. Starting from the first word, each text word is added one by one, and the coherence of the text segment is calculated according to the semantic coherence measurement formula. See Figure 2 As shown, the score (semantic coherence measurement value) obtained by measuring "store" based on the semantic coherence measurement formula is 89, the score obtained by measuring "store sound" based on the semantic coherence measurement formula is 32, and the score obtained by measuring "sound" based on the semantic coherence measurement formula is 78. Calculate the coherence difference based on formula 2:

[0070] D(S j ) = |L(S j-1 +w j ) - L(S j-1 )| (formula 2)

[0071] Among them, when j = 0, D(S j ) = 0. For the case of calculating the coherence difference of "store sound", L(S j-1 ) corresponds to the score (semantic coherence measurement value) of "store", and L(S j-1 +w j ) corresponds to the score (semantic coherence measurement value) of "store sound". Since the absolute value difference between the score corresponding to "store sound" and the score corresponding to "store" is 57, it indicates that the coherence difference between "store" and "sound" is large. Therefore, "store" and "sound" are segmented.

[0072] The score obtained by measuring "sound son" based on the semantic coherence measurement formula is 41, the score obtained by measuring "son" based on the semantic coherence measurement formula is 82. Calculate the coherence difference of "sound son" based on formula 2. Since the absolute value difference between the score corresponding to "sound son" and the score corresponding to "sound" is 37, "sound" and "son" are segmented. The score obtained by measuring "son flat" based on the semantic coherence measurement formula is 47, the score obtained by measuring "flat" based on the semantic coherence measurement formula is 82. Calculate the coherence difference of "son flat" based on formula 2. Since the absolute value difference between the score corresponding to "son flat" and the score corresponding to "son" is 35, "son" and "flat" are segmented.

[0073] For Figure 2 For the part corresponding to the ellipsis in , the corresponding processing principle is the same as the above-mentioned processing principle. Finally, the score of "assuming corresponding responsibilities and making rectifications" is measured as 76, with a difference of 8 from the score of "assuming corresponding responsibilities and making". The coherence of "assuming corresponding responsibilities and making rectifications" meets the coherence condition, so it does not need to be segmented. So far, the video text 4 can be segmented into multiple text segments: {[store], [sound], [son], [tan], [assuming corresponding responsibilities and making rectifications]}, completing the segmentation of the text segments of the video text 4.

[0074] After segmenting the text segments based on semantic coherence, for the case where the video text includes different types of text segments, corresponding processing strategies need to be adopted for text correction. For the case where the video text includes a target text segment and at least one first text segment to be corrected, the at least one first text segment can be corrected with the target text segment as the center; for the case where the video text only includes the target text segment, since there is no text segment to be corrected, it is determined that the video text does not need to be corrected; for the case where the video text does not include the target text segment, includes a second text segment and at least one first text segment to be corrected, the at least one first text segment can be corrected based on the semantic coherence of the context. For each video frame in multiple videos, after completing the segmentation of the video text, corresponding processing strategies can be adopted for text correction.

[0075] In the above implementation process of this application, by segmenting the video text based on semantic coherence and then adopting corresponding processing strategies for text correction based on the type of text segments, text correction can be achieved using a matching correction method, improving the correction efficiency.

[0076] Optionally, the correcting the at least one first text segment with the target text segment as the center includes:

[0077] With the target text segment as the center, based on the processing methods of character retention, character replacement, and character deletion, the first text segment located before the target text segment is corrected in reverse word order; and / or

[0078] With the target text segment as the center, based on the processing methods of character retention, character replacement, and character deletion, the first text segment located after the target text segment is corrected in forward word order.

[0079] In the case where the video text includes a target text segment and at least one first text segment, when correcting the first text segment, the target text segment can be used as the center, and the first text segment can be corrected based on character retention, character replacement, and character deletion processing methods. When the first text segment is located after the target text segment, text correction is performed based on the forward word order. When the first text segment is located before the target text segment, text correction is performed based on the reverse word order.

[0080] In the case where there are multiple first text segments, and some of the first text segments are located after the target text segment and some are located before the target text segment, the first text segment on the side with the shorter length can be corrected first. For example, taking "[take corresponding responsibilities and make rectifications]" as the central segment, comparing the text lengths on both sides, first correct the text on the right side. Since the length is 0, it is directly skipped, and then correct the text on the left side. That is, for this case, the correction priority of the first text segment is negatively correlated with the character length.

[0081] When correcting the first text segment, based on character retention, character replacement, and character deletion processing methods, word-by-word correction is performed on the first text segment. During the word-by-word correction process, multiple predicted texts are constructed. Since constructing a single predicted text cannot guarantee the accuracy of the constructed predicted text, multiple predicted texts can be constructed, and the smoothness of the predicted texts can be measured by the semantic coherence metric formula to screen out several predicted texts with higher rankings.

[0082] Taking "[store][sound][son][tan][take corresponding responsibilities and make rectifications]" as an example, the correction process is described below. See Figure 3As shown, first, the character "坦" is MASKed (the "坦" is replaced with [MASK], and MASK is a prediction indicator), and then predictions are made through a masked language model (MLM). For the character "坦", multiple predicted texts are obtained. Based on the fluency metric, the top 4 predicted texts "负, 担, 的, 有" are obtained. If the character "坦" is retained, it is: 店音子坦; if the character "坦" is deleted, it is 店音子…; if the character "坦" is replaced, it is: 店音子负, 店音子担, 店音子的, 店音子有. Then the character "子" is MASKed (the "子" is replaced with [MASK]), and then predictions are made through MLM. For the character "子", multiple predicted texts are obtained. After combining with the predicted texts corresponding to the character "坦", the following content is obtained: 店音<承, 追, 究, 应, 认>坦; 店音<负, 担, 究, 有, 的>…; 店音<将, 应, 要, 会, 乐>负; 店音<承, 应, 负, 自, 要>担; 店音<乐, 部, 社, 响, 队>的; 店音<负, 承, 会, 也, 应>有. For "子坦", based on the fluency metric, the top 4 predicted texts in the ranking are obtained, and the following content is obtained: 店音<将负, 应负, 承担, 负有>相应责任并作出整改. Based on the character retention, character replacement, and character deletion processing methods, the correction is carried out, and then the pronunciation and the store are corrected in turn according to the above process to complete the correction.

[0083] In the above implementation process of this application, by determining the target text segment, taking the target text segment as the center, and using the forward word order and / or reverse word order for text correction, reasonable correction can be carried out based on the semantics of the context with the target text segment as the basis.

[0084] Optionally, the text correction of at least one of the first text segments based on semantic coherence includes:

[0085] For each of the first text segments, based on semantic coherence, multiple correction segments are constructed for the first text segment by using the character retention, character replacement, and character deletion processing methods;

[0086] Based on semantic coherence, the fluency of the multiple correction segments with the context before and after is measured;

[0087] According to the measurement result, a target correction segment is determined among the multiple correction segments, and the first text segment is text-corrected by using the target correction segment.

[0088] In the case where the video text includes a second text segment and at least one first text segment to be corrected, when correcting the first text segment, since there is a second text segment in the current situation, multiple correction segments can be constructed for the first text segment by using character retention, character replacement, and character deletion processing methods based on the semantic coherence of the context. Then, the smoothness of the multiple correction segments is measured using the semantic coherence metric formula, and the target correction segment is selected to correct the first text segment with the target correction segment.

[0089] The following is an example to introduce this situation. The video text recognized from "情深万分恩爱" is: 情深万会恩爱. After segmenting the video text, multiple text segments are obtained: [情深][万][会][恩爱]. At this time, there are only the second text segments [情深] and [恩爱], and the first text segments [万] and [会]. First, correct "万". Mask the character "万", and then perform prediction through MLM. For the character "万", multiple predicted texts are obtained. Based on the smoothness measurement, the top 5 predicted texts are "才, 也, 很, 更, 又". If the character "万" is retained, it is: 万会恩爱; if the character "万" is deleted, it is:...会恩爱; if the character "万" is replaced, it is: 才会恩爱, 也会恩爱, 很会恩爱, 更会恩爱, 又会恩爱. Then mask the character "会", and then perform prediction through MLM. For the character "会", multiple predicted texts are obtained. After combining with the predicted texts corresponding to the character "万", the following content is obtained: 万<般, 分, 种, 世, 家>恩爱;...<恩, 亲, 相, 是, 恋>恩爱; 才<会>恩爱; 也<很, 是, 有, 算, 挺>恩爱; 很<是, 有, 不, 的, 感>恩爱; 更<是, 加, 有, 多, 显>恩爱; 又<很, 秀, 是, 又, 最>恩爱. For "万会", based on the smoothness measurement, the top 5 predicted texts in the ranking are obtained: 情深<万般, 万分, 也很, 很是, 更是>恩爱, and then the video text is corrected to get the corrected video text: 情深万般恩爱.

[0090] In the above implementation process of this application, by constructing multiple correction segments and measuring the smoothness of the multiple correction segments with the context based on semantic coherence to select the target correction segment, the first text segment can then be corrected with the target correction segment.

[0091] Optionally, in step 103, for the multiple video frames, based on the video correction text and the video text position information corresponding to each video frame, determining N video text lists includes:

[0092] For a first video frame among the multiple video frames, generating a list set including L first text lists and list position information of each first text list according to the text semantic coherence of the video-corrected text corresponding to the first video frame and the video text position information, where L is an integer greater than or equal to 1;

[0093] For the Mth video frame, according to the video text position information corresponding to the Mth video frame and the list set updated based on the (M-1)th video frame, determine the second text list corresponding to the Mth video frame, and dynamically update the list set, wherein M is an integer greater than or equal to 2 and less than or equal to G, G is the number of video frames corresponding to the multiple video frames, and the (G-1)th video frame updates the list set in sequence according to the time sequence corresponding to the video frames;

[0094] For the Gth video frame, after determining the second text list corresponding to the Gth video frame and dynamically updating the list set, determining the N video text lists according to the N text lists in the list set, where N is greater than or equal to L;

[0095] Wherein, for video frames other than the first video frame, the list set is updated once based on each video frame.

[0096] When determining N video text lists based on the video correction text and the video text position information corresponding to each video frame, first determine L first text lists for the first video frame among multiple video frames (the video frame obtained by first performing frame capture processing on the target video), based on the text semantic coherence and video text position information of the video correction text corresponding to the first video frame, and obtain the list position information corresponding to each first text list. For the L first text lists, different first text lists have no semantic association between video texts. After determining the L first text lists and the list position information corresponding to each first text list, create a list set. When the number of first text lists is updated and the content in the first text list is updated, the list set is dynamically updated. In the case of content update in the first text list, the corresponding list position information will also be updated.

[0097] For video frames other than the first video frame, the list set needs to be updated once based on each video frame. After updating based on the second video frame, when processing the third video frame, it is necessary to base it on the list set updated based on the second video frame. Correspondingly, when processing the fourth video frame, it is necessary to base it on the list set updated based on the third video frame. That is, the previous update needs to be used as the basis for the next processing.

[0098] After creating a list set based on the first video frame, the remaining G-1 video frames can be processed in sequence according to the time sequence from early to late (the time information corresponding to the video frame). For any video frame, it is necessary to process the video text corresponding to the current video frame according to the video text position information corresponding to the current video frame and the list set in the latest state (the list set updated according to the previous video frame), determine the second text list corresponding to the current video frame, and dynamically update the list set.

[0099] Among them, the second text list corresponding to the current video frame can be a text list in the list set, or a newly created text list storing video correction text, and the current video frame can correspond to at least one second text list. In the case where the second text list corresponding to the current video frame is a list in the list set, the video correction text in the current video frame is added to the matching first text list in the list set to update the content in the first text list, and the list position information of the first text list is updated, and the list set is updated while the first text list is updated. In the case where the second text list corresponding to the current video frame is a newly created text list, after the new text list is created, the video correction text is added to the newly created text list, and the newly created text list storing the video correction text is the second text list. The second text list and the list position information are added to the list set to update the list set.

[0100] In the case where the current video frame corresponds to at least two second text lists, the at least two second text lists may both be text lists in the list set, or both be newly created text lists, or some may be text lists in the list set and the other may be newly created text lists.

[0101] After performing video text processing on the Gth (the last of multiple video frames) video frame, determining the second text list corresponding to the Gth video frame, and dynamically updating the list set, the text list in the list set can be determined as a video text list. At this time, the number of text lists in the list set is N. Due to the existence of a new text list, the value of N can be greater than L. When no new text list is created, the value of N is equal to L.

[0102] The above-mentioned implementation process of the present application creates a list set based on the first video frame, determines the second text list corresponding to any video frame different from the first video frame and dynamically updates the list set, thereby updating the list set to determine the video text list based on the updated list set.

[0103] Optionally, for the Mth video frame, determining the second text list corresponding to the Mth video frame according to the video text position information corresponding to the Mth video frame and the list set updated based on the (M-1)th video frame, and dynamically updating the list set, comprises:

[0104] According to the video text position information corresponding to the video correction text in the Mth video frame and the list position information corresponding to the list set updated based on the (M-1)th video frame, detecting whether there is a target text list matching the video correction text in the Mth video frame in the list set;

[0105] If it does not exist, create a new third text list, add the video correction text in the Mth video frame to the third text list to determine the second text list corresponding to the Mth video frame, and add the second text list corresponding to the Mth video frame and the corresponding list position information to the list set;

[0106] If so, the video correction text in the Mth video frame is processed according to the semantic relationship between the video correction text in the Mth video frame and the video text in the target text list to determine the second text list corresponding to the Mth video frame and dynamically update the list set.

[0107] When performing video text processing on the Mth video frame, the video text position information corresponding to the video correction text in the Mth video frame and the list position information of each text list in the latest state list set can be obtained, and based on the position relationship, it is detected whether there is a target text list in the list set that matches the video correction text in the Mth video frame. The Mth video frame may include multiple video correction texts, and a detection process needs to be performed for each video correction text to determine the corresponding second text list.

[0108] The following describes the process of detecting whether there is a target text list in the list set that matches the video correction text in the Mth video frame. Assume that all the text lists in the current list set are {P1, P2, ..., P l}, the video correction text in the Mth video frame is T, and the position coincidence relationship between the video correction text T and all text lists in the current list set is determined. The position information of the video correction text T includes the coordinate information of the upper left corner and the lower right corner of the corresponding text box. For example, the coordinate form of the video correction text T can be expressed as The following formula is used to calculate the overlap degree between the video correction text T and the text boxes corresponding to each text list, and the threshold β is set.

[0109]

[0110] In the above formula, Pe q represents the position overlap between the qth text list in the list set and the video correction text T, and the value of q ranges from 1 to l. q ∩T represents the intersection of the qth text list and the position information of the video correction text T, P q ∪T represents the union of the position information of the qth text list and the video correction text T. For {Pe1,Pe2,...,Pe l}, the largest one is selected as Pe, and the corresponding expression of Pe is Pe=max({Pe1,Pe2,...,Pe l}). If Pe<β, it means that the video correction text T has no overlap with all the text lists in the list set. If Pe>β, it means that the video correction text T has an overlap with at least one text list in the list set.

[0111] For the case of Pe>β, the text list with the largest position overlap with the video correction text T is the target text list that matches the video correction text T. At this time, it is necessary to process the video correction text T based on the semantic relationship between the video correction text T and the video text in the target text list to determine the second text list corresponding to the Mth video frame and dynamically update the list set. For example, the existing first text lists in the list set are first text list 1, first text list 2, first text list 3, first text list 4, and first text list 5. The five first text lists are determined based on the first video frame. See Figure 4a As shown, it is a schematic diagram of the first video frame, the first text list 1 includes the following video text: XX Satellite TV, the first text list 2 includes the following video text: HD, the first text list 3 includes the following video text: XX Food, the first text list 4 includes the following video text: XX Mediation, the first text list 5 includes the following video text: His foundation is mine (with a share), Dad didn't leave it to him alone. The video correction text T is "Brother Zhu left this time", and the video correction text T is the content in the second video frame, see Figure 4b As shown, it is a schematic diagram of the second video frame, where Figure 4a In Figure 4b In the figure, the text is marked with a dotted box. First, the position overlap relationship between the video correction text T and all the first text lists is calculated. If the video correction text T overlaps with the first text list 5, a semantic analysis is required based on the content of the video correction text T and the content corresponding to the first text list 5 to determine the relationship between the two, and then determine the text list corresponding to the video correction text T.

[0112] It should be noted that the position information of the video correction text in the embodiment of the present application refers to the position information corresponding to the text box corresponding to the video correction text, and correspondingly, the position information of the text list is the position information of the text box corresponding to the text list.

[0113] For the case of Pe<β, since the video correction text T has no overlapping relationship with the text lists in all list sets, it is necessary to create a third text list to store the video correction text T. After the third text list is created, the video correction text T is added to the newly created third text list. The newly created third text list to store the video correction text T is the second text list corresponding to the Mth video frame. The second text list and the corresponding list position information are added to the list set to update the list set.

[0114] The above implementation process of the present application detects whether there is a target text list in the list set based on the position information of the video correction text and the list position information of the text list in the list set, and creates a new text list if it does not exist. If it does exist, based on the text semantics, the relationship between the video correction text and the target text list is further determined, thereby realizing the update of the list set based on the two dimensions of position information and text semantics.

[0115] Optionally, the processing the video-corrected text in the M-th video frame according to the semantic relationship between the video-corrected text in the M-th video frame and the video text in the target text list to determine a second text list corresponding to the M-th video frame and dynamically update the list set includes:

[0116] When the video correction text in the Mth video frame is identical to the video text in the target text list, ignoring the video correction text in the Mth video frame, wherein the target text list is a second text list corresponding to the Mth video frame;

[0117] In the case where there is a semantic association between the video correction text in the Mth video frame and the video text in the target text list, the video correction text in the Mth video frame is added to the target text list, and the list position information of the target text list is updated to update the list set, and the target text list is a second text list corresponding to the Mth video frame;

[0118] In the case that there is no semantic association between the video correction text in the Mth video frame and the video text in the target text list, a fourth text list is newly created, the video correction text in the Mth video frame is added to the fourth text list to determine the second text list corresponding to the Mth video frame, and the second text list corresponding to the Mth video frame and the corresponding list position information are added to the list set to update the list set.

[0119] In the case where the video correction text in the Mth video frame is identical to the video text in the target text list, due to the identical contents, if the video correction text in the Mth video frame is added to the target text list, the content will be repeated and the repeated part needs to be deleted. Therefore, in this case, the video correction text in the Mth video frame can be directly ignored. At this time, the target text list is the second text list corresponding to the Mth video frame, and the list set remains unchanged before and after the update, that is, the list set is not updated.

[0120] In the case where the video correction text in the Mth video frame has a semantic association with the video text in the target text list, the video correction text in the Mth video frame can be directly added to the target text list to update the content of the target text list. While updating the content of the target text list, the list position information of the target text list is updated. The list set is updated by updating the content and the list position information. At this time, the target text list is the second text list corresponding to the Mth video frame. For example, the video correction text T0 is: Zhu's brother left this time, and the video text T1 corresponding to the target text list P is: His foundation is mine (with a share), and his father did not leave it to him alone. The spliced ​​video text T2 is: His foundation is mine (with a share), and his father did not leave it to him alone. Zhu's brother left this time. Calculate the coherence of T0, T1 and T2 respectively. If the relationship between T0, T1 and T2 satisfies the following formula, add the video correction text T0 to the target text list P, and update the position information of the target text list P.

[0121]

[0122] Among them, the position update process of the target text list P is:

[0123]

[0124]

[0125]

[0126]

[0127] In the process of updating the position of the target text list P, for the coordinates in the upper left corner, select the left coordinate in the X-axis direction, and select the upper coordinate in the Y-axis direction; for the coordinates in the lower right corner, select the right coordinate in the X-axis direction, and select the lower coordinate in the Y-axis direction. This position update method can ensure that the video correction text T0 is accommodated.

[0128] When there is no semantic association between the video correction text in the Mth video frame and the video text in the target text list, a fourth text list is created, and then the video correction text in the Mth video frame is added to the newly created fourth text list, and the fourth text list storing the video correction text in the Mth video frame is the second text list corresponding to the Mth video frame, and then the second text list and the corresponding list position information are added to the list set to realize the update of the list set.

[0129] In the above implementation process of the present application, when processing the video correction text based on semantic relationship, the video correction text can be processed by ignoring the video correction text, adding the video correction text to the target text list, or creating a new text list according to different situations, so as to update the list set according to different situations.

[0130] Optionally, the step 104 of selecting K target video text lists from the N video text lists based on the semantic association between the N video text lists and the video titles corresponding to the target video includes:

[0131] Obtain text vectors corresponding to the N video text lists and a title vector corresponding to the video title;

[0132] For each of the text vectors, a similarity weight between the text vector and the title vector is calculated, and a similarity score is obtained according to the similarity weight;

[0133] According to the similarity scores respectively corresponding to the N video text lists, K target video text lists whose semantic relevance to the video title meets preset conditions are screened out.

[0134] After determining N video text lists, the text vectors corresponding to each of the N video text lists can be obtained, and at the same time, the title vector corresponding to the video title corresponding to the target video can be obtained.

[0135] The following is a brief description of the process of obtaining text vectors. For the video text in the video text list, the text words are represented by the deep bidirectional pre-trained converter (Bidirectional Encoder Representation from Transformers, Bert) based on semantic understanding or Tencent word vectors, and then the text word weights are calculated by term frequency-inverse document frequency (Term Frequency-Inverse Document Frequency, TF-IDF) to perform vector fusion, as shown below:

[0136] vec(S)=∑ j TFIDF(w j )*vec(w j )

[0137] Among them, vec(w j ) is the vector representation of the text words in the video text list, TFIDF(w j ) represents the text word weight corresponding to the text word in the video text list. By multiplying the vector form of the text word by the corresponding text word weight for each text word in the video text list, and then accumulating, the text vector corresponding to the video text list can be obtained. For example, the video text corresponding to the video text list is: XX Satellite TV. The four characters of XX Satellite TV correspond to four text words. The value of j ranges from 1 to 4. Based on the vectors of the four text words and the corresponding text word weights, the text vector corresponding to the video text list is determined.

[0138] For the video title, the method for determining the title vector corresponding to the video title is the same as the method for determining the text vector described above, which will not be repeated here. After determining the text vectors corresponding to the N video text lists and the title vector corresponding to the video title, the similarity weight between the current text vector and the title vector can be calculated for each text vector, and the similarity score can be obtained based on the similarity weight.

[0139] The following is an explanation of how to calculate the similarity weight and similarity score between the text vector and the title vector. See the following formula.

[0140]

[0141] Among them, p(title, content i ) represents the similarity weight between the text vector and the title vector, vec(title) represents the title vector, vec(content i ) represents the text vector corresponding to the i-th video text list, and the value of i is a number between 1 and N.

[0142] After obtaining the similarity weights of the text vector and the title vector, the similarity score may be calculated based on the similarity weights. When calculating the similarity score, please refer to the following formula for details.

[0143]

[0144] The value of i is a number between 1 and N. After calculating the similarity scores corresponding to the N video text lists, the N similarity scores can be compared with a threshold value γ, and the corresponding video text lists with similarity scores greater than the threshold value γ can be screened out, and the screened video text lists are determined as target video text lists whose semantic association with the video title meets a preset condition.

[0145] The above implementation process of the present application calculates the similarity weights between the text vector and the title vector, obtains the similarity score based on the similarity weight, and screens out the target video text list whose semantic relevance to the video title meets the preset conditions according to the similarity scores corresponding to N video text lists, thereby realizing effective screening of video texts.

[0146] The overall implementation process of the video text screening method provided in the embodiment of the present application is introduced below. Figure 5 As shown, including:

[0147] Step 501: Capture a target video to obtain multiple video frames.

[0148] Step 502: Obtain video texts and video text position information corresponding to a plurality of video frames.

[0149] Step 503: Perform single-frame text correction on each video frame.

[0150] Step 504: Generate a list set according to the text semantic coherence and video text position information of the video-corrected text corresponding to the first video frame.

[0151] Step 505 : According to the order of the video frames, the list set of the remaining video frames is updated in sequence based on the two dimensions of video text position information and text semantics, so as to determine a video text list based on the updated list set.

[0152] Step 506: based on the semantic association between the video text list and the video title, a target video text list is screened out from the video text list, and a target video text corresponding to the target video is generated according to the video text corresponding to the target video text list.

[0153] The above implementation process can ensure the integrity and coherence of text semantics by performing text correction, and can ensure the relevance of video text by judging the positional relationship of video texts of previous and next frame videos and using semantic understanding to determine the coherence of video texts of previous and next frame videos. It can also screen video texts based on semantic relevance and remove irrelevant texts, thereby facilitating rapid understanding of video texts.

[0154] The above is the overall implementation process of the video text screening method provided in the embodiment of the present application. After obtaining the video texts corresponding to multiple video frames of the target video, text correction is performed on the video text to ensure the integrity and coherence of the text semantics; after text correction, for multiple video frames, based on the video correction text and the video text position information corresponding to each video frame, N video text lists are determined, and a video text list including video texts with coherent text semantics can be obtained; by screening out the target video text list based on the semantic correlation between the video text list and the video title, and then generating the target video text corresponding to the target video, video texts with high semantic correlation with the video title can be retained in a variety of video text usage scenarios, thereby achieving effective screening of video texts.

[0155] Furthermore, by segmenting the video text into text segments based on semantic coherence, after segmentation, the text is corrected by adopting corresponding processing strategies based on the types of text segments, so that text correction can be performed using a matching correction method to improve correction efficiency.

[0156] By creating a list set based on the first video frame, determining the text list corresponding to the video correction text in any video frame different from the first video frame and dynamically updating the list set, it is possible to update the list set based on two dimensions: position information and text semantics, so as to determine the video text list based on the updated list set.

[0157] By calculating the similarity weights between the text vector and the title vector, obtaining the similarity score based on the similarity weight, and filtering out the target video text list whose semantic relevance to the video title meets the preset conditions according to the similarity scores corresponding to the N video text lists, effective screening of video texts can be achieved.

[0158] The video text screening method provided in the embodiment of the present application can be executed by a video text screening device. In the embodiment of the present application, the video text screening device performs a method of video text screening as an example to illustrate the video text screening device provided in the embodiment of the present application.

[0159] Figure 6 is a schematic block diagram of a video text screening device provided by an embodiment of the present application. Figure 6 As shown, the device comprises:

[0160] The first acquisition module 601 is used to acquire video texts and video text position information respectively corresponding to a plurality of video frames of a target video;

[0161] The second acquisition module 602 is used to perform text correction processing on the video text to obtain a video corrected text with coherent text semantics;

[0162] A determination module 603 is used to determine N video text lists for the multiple video frames based on the video correction text and the video text position information corresponding to each video frame, where N is an integer greater than or equal to 1;

[0163] Processing module 604 is used to screen out K target video text lists from the N video text lists based on the semantic association between the N video text lists and the video titles corresponding to the target video, and generate a target video text corresponding to the target video according to the video texts corresponding to the K target video text lists, where K is an integer less than or equal to N.

[0164] Optionally, the second acquisition module includes:

[0165] A segmentation submodule, for performing text segmentation on each video text corresponding to each video frame based on semantic coherence;

[0166] A first correction submodule is used for, when the video text includes a target text segment and at least one first text segment to be corrected, taking the target text segment as the center, performing text correction on at least one first text segment to obtain a video correction text corresponding to the video text;

[0167] A determination submodule, configured to determine that the video text does not need to be corrected when the video text only includes the target text segment;

[0168] A second correction submodule is used for, when the video text does not include the target text segment, includes a second text segment and at least one first text segment to be corrected, performing text correction on at least one of the first text segments based on semantic coherence to obtain a video correction text corresponding to the video text;

[0169] The second text segment is a correct text segment, and the target text segment is a text segment with semantic coherence and the longest character length.

[0170] Optionally, the first correction submodule includes:

[0171] A first correction unit is configured to correct the first text segment preceding the target text segment by using a reverse word order based on character retention, character replacement, and character deletion processing modes with respect to the target text segment; and / or

[0172] The second correction unit is used to perform text correction on the first text segment located after the target text segment by using a positive word order based on character retention, character replacement and character deletion processing methods with the target text segment as the center.

[0173] Optionally, the second correction submodule includes:

[0174] a construction unit, configured to construct a plurality of correction segments for each of the first text segments based on semantic coherence by adopting character retention, character replacement and character deletion processing methods;

[0175] A measuring unit, used for measuring the coherence between the plurality of corrected segments and the context based on semantic coherence;

[0176] The third correction unit is used to determine a target correction segment from the multiple correction segments according to the measurement result, and use the target correction segment to perform text correction on the first text segment.

[0177] Optionally, the determining module includes:

[0178] A generating submodule, for generating, for a first video frame among the plurality of video frames, a list set including L first text lists and list position information of each first text list according to the text semantic coherence of the video-corrected text corresponding to the first video frame and the video text position information, where L is an integer greater than or equal to 1;

[0179] A first processing submodule is used to determine, for an M-th video frame, a second text list corresponding to the M-th video frame according to the video text position information corresponding to the M-th video frame and the list set updated based on the (M-1)-th video frame, and dynamically update the list set, wherein M is an integer greater than or equal to 2 and less than or equal to G, G is the number of video frames corresponding to the multiple video frames, and the (G-1)-th video frame updates the list set in sequence according to the time sequence corresponding to the video frames;

[0180] A second processing submodule is used for determining, for a G-th video frame, a second text list corresponding to the G-th video frame and dynamically updating the list set, and then determining the N video text lists according to the N text lists in the list set, where N is greater than or equal to L;

[0181] Wherein, for video frames other than the first video frame, the list set is updated once based on each video frame.

[0182] Optionally, the first processing submodule includes:

[0183] a detection unit, configured to detect whether there is a target text list matching the video correction text in the Mth video frame in the list set according to the video text position information corresponding to the video correction text in the Mth video frame and the list position information corresponding to the list set updated based on the (M-1)th video frame;

[0184] A first processing unit is configured to create a new third text list if it does not exist, add the video correction text in the Mth video frame to the third text list to determine a second text list corresponding to the Mth video frame, and add the second text list corresponding to the Mth video frame and corresponding list position information to the list set;

[0185] The second processing unit is used to process the video correction text in the Mth video frame according to the semantic relationship between the video correction text in the Mth video frame and the video text in the target text list, if any, so as to determine a second text list corresponding to the Mth video frame and dynamically update the list set.

[0186] Optionally, the second processing unit includes:

[0187] A first processing subunit is configured to ignore the video correction text in the Mth video frame when the video correction text in the Mth video frame has the same content as the video text in the target text list, wherein the target text list is a second text list corresponding to the Mth video frame;

[0188] a second processing subunit, configured to add the video correction text in the Mth video frame to the target text list and update the list position information of the target text list to update the list set, when there is a semantic association between the video correction text in the Mth video frame and the video text in the target text list, wherein the target text list is a second text list corresponding to the Mth video frame;

[0189] The third processing sub-unit is used to create a fourth text list when there is no semantic association between the video correction text in the Mth video frame and the video text in the target text list, add the video correction text in the Mth video frame to the fourth text list to determine a second text list corresponding to the Mth video frame, and add the second text list corresponding to the Mth video frame and the corresponding list position information to the list set to update the list set.

[0190] Optionally, the processing module includes:

[0191] An acquisition submodule, used to acquire text vectors corresponding to the N video text lists and a title vector corresponding to the video title;

[0192] A calculation and acquisition submodule, used for calculating the similarity weight between the text vector and the title vector for each of the text vectors, and obtaining a similarity score according to the similarity weight;

[0193] The screening submodule is used to screen out K target video text lists whose semantic relevance to the video title meets preset conditions according to the similarity scores corresponding to the N video text lists.

[0194] The above is an embodiment of the video text screening device provided in the embodiment of the present application. After obtaining the video texts corresponding to multiple video frames of the target video, the video text is corrected to ensure the integrity and coherence of the text semantics; after the text correction is performed, for multiple video frames, based on the video correction text and the video text position information corresponding to each video frame, N video text lists are determined, and a video text list including video texts with coherent text semantics can be obtained; by screening out the target video text list based on the semantic correlation between the video text list and the video title, and then generating the target video text corresponding to the target video, the video text with a high semantic correlation with the video title can be retained in a variety of video text usage scenarios, thereby realizing effective screening of the video text.

[0195] Furthermore, by segmenting the video text into text segments based on semantic coherence, after segmentation, the text is corrected by adopting corresponding processing strategies based on the types of text segments, so that text correction can be performed using a matching correction method to improve correction efficiency.

[0196] By creating a list set based on the first video frame, determining the text list corresponding to the video correction text in any video frame different from the first video frame and dynamically updating the list set, it is possible to update the list set based on two dimensions: position information and text semantics, so as to determine the video text list based on the updated list set.

[0197] By calculating the similarity weights between the text vector and the title vector, obtaining the similarity score based on the similarity weight, and filtering out the target video text list whose semantic relevance to the video title meets the preset conditions according to the similarity scores corresponding to the N video text lists, effective screening of video texts can be achieved.

[0198] The video text screening device in the embodiment of the present application can be an electronic device, or a component in the electronic device, such as an integrated circuit or a chip. The electronic device can be a terminal, or other devices other than the terminal. Exemplarily, the electronic device can be a mobile phone, a tablet computer, a laptop computer, a palmtop computer, a vehicle-mounted electronic device, a mobile Internet device (Mobile Internet Device, MID), an augmented reality (augmented reality, AR) / virtual reality (virtual reality, VR) device, a robot, a wearable device, an ultra-mobile personal computer (ultra-mobile personal computer, UMPC), a netbook or a personal digital assistant (personal digital assistant, PDA), etc., and can also be a server, a network attached storage (Network Attached Storage, NAS), a personal computer (personal computer, PC), a television (television, TV), a teller machine or a self-service machine, etc., which is not specifically limited in the embodiment of the present application.

[0199] The video text screening device in the embodiment of the present application may be a device having an operating system. The operating system may be an Android operating system, an iOS operating system, or other possible operating systems, which are not specifically limited in the embodiment of the present application.

[0200] The video text screening device provided in the embodiment of the present application can achieve Figure 1 To avoid repetition, the various processes implemented in the video text screening method embodiment shown will not be described here.

[0201] Alternatively, if Figure 7 As shown, an embodiment of the present application also provides an electronic device 700, including a processor 701, a memory 702, and a program or instruction stored in the memory 702 and executable on the processor 701. When the program or instruction is executed by the processor 701, each step of the above-mentioned video text screening method embodiment is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0202] It should be noted that the electronic devices in the embodiments of the present application include the mobile electronic devices and non-mobile electronic devices mentioned above.

[0203] Figure 8 A schematic diagram of the hardware structure of an electronic device to implement an embodiment of the present application.

[0204] The electronic device 800 includes but is not limited to: a radio frequency unit 801, a network module 802, an audio output unit 803, an input unit 804, a sensor 805, a display unit 806, a user input unit 807, an interface unit 808, a memory 809 and a processor 810, and the electronic device also includes multiple cameras.

[0205] Those skilled in the art will appreciate that the electronic device 800 may also include a power source (such as a battery) for supplying power to each component, and the power source may be logically connected to the processor 810 through a power management system, thereby implementing functions such as managing charging, discharging, and power consumption through the power management system. Figure 8 The electronic device structure shown in the figure does not constitute a limitation on the electronic device. The electronic device may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently, which will not be described in detail here.

[0206] Among them, the processor 810 is used to: obtain video texts and video text position information corresponding to multiple video frames of the target video respectively; perform text correction processing on the video text to obtain video correction text with semantically coherent text; for the multiple video frames, based on the video correction text and the video text position information corresponding to each video frame, determine N video text lists, N is an integer greater than or equal to 1; based on the semantic association between the N video text lists and the video title corresponding to the target video, screen out K target video text lists from the N video text lists, and generate the target video text corresponding to the target video according to the video texts corresponding to the K target video text lists, K is an integer less than or equal to N.

[0207] Optionally, when performing text correction processing on the video text to obtain a video correction text with semantic coherence, the processor 810 is also used to: for each video frame, for each video text corresponding to the video frame, perform text segmentation based on semantic coherence; when the video text includes a target text segment and at least one first text segment to be corrected, perform text correction on at least one of the first text segments with the target text segment as the center to obtain a video correction text corresponding to the video text; when the video text only includes the target text segment, determine that the video text does not need to be text corrected; when the video text does not include the target text segment but includes a second text segment and at least one first text segment to be corrected, perform text correction on at least one of the first text segments based on semantic coherence to obtain a video correction text corresponding to the video text; wherein the second text segment is a correct text segment, and the target text segment is a text segment with semantic coherence and the longest character length.

[0208] Optionally, when performing text correction on at least one of the first text segments with the target text segment as the center, the processor 810 is also used to: with the target text segment as the center, based on character retention, character replacement and character deletion processing methods, perform text correction on the first text segment before the target text segment using reverse word order; and / or with the target text segment as the center, based on character retention, character replacement and character deletion processing methods, perform text correction on the first text segment after the target text segment using forward word order.

[0209] Optionally, when performing text correction on at least one of the first text segments based on semantic coherence, the processor 810 is also used to: for each of the first text segments, based on semantic coherence, use character retention, character replacement and character deletion processing methods to construct multiple correction segments for the first text segment; based on semantic coherence, measure the coherence of the multiple correction segments with the context before and after; according to the measurement results, determine a target correction segment from the multiple correction segments, and use the target correction segment to perform text correction on the first text segment.

[0210] Optionally, when determining N video text lists for the multiple video frames based on the video correction text and the video text position information corresponding to each video frame, the processor 810 is further used to: for the first video frame among the multiple video frames, generate a list set including L first text lists and list position information of each first text list according to the text semantic coherence of the video correction text corresponding to the first video frame and the video text position information, where L is an integer greater than or equal to 1; for the Mth video frame, generate a list set including L first text lists and list position information of each first text list according to the video text position information corresponding to the Mth video frame and the list set updated based on the (M-1)th video frame. The method comprises the steps of: determining a second text list corresponding to the M-th video frame, and dynamically updating the list set, wherein M is an integer greater than or equal to 2 and less than or equal to G, G is the number of video frames corresponding to the multiple video frames, and updating the list set in sequence for (G-1) video frames according to the time sequence corresponding to the video frames; for the G-th video frame, after determining the second text list corresponding to the G-th video frame and dynamically updating the list set, determining the N video text lists according to the N text lists in the list set, where N is greater than or equal to L; wherein, for video frames other than the first video frame, the list set is updated once based on each video frame.

[0211] Optionally, for the Mth video frame, according to the video text position information corresponding to the Mth video frame and the list set updated based on the (M-1)th video frame, a second text list corresponding to the Mth video frame is determined, and the list set is dynamically updated, the processor 810 is further used to: detect whether there is a target text matching the video correction text in the Mth video frame in the list set according to the video text position information corresponding to the video correction text in the Mth video frame and the list position information corresponding to the list set updated based on the (M-1)th video frame. this list; if it does not exist, a new third text list is created, and the video correction text in the Mth video frame is added to the third text list to determine the second text list corresponding to the Mth video frame, and the second text list corresponding to the Mth video frame and the corresponding list position information are added to the list set; if it exists, according to the semantic relationship between the video correction text in the Mth video frame and the video text in the target text list, the video correction text in the Mth video frame is processed to determine the second text list corresponding to the Mth video frame and dynamically update the list set.

[0212] Optionally, when processing the video correction text in the Mth video frame according to the semantic relationship between the video correction text in the Mth video frame and the video text in the target text list to determine the second text list corresponding to the Mth video frame and dynamically update the list set, the processor 810 is also used to: when the video correction text in the Mth video frame is the same as the content of the video text in the target text list, ignore the video correction text in the Mth video frame, wherein the target text list is the second text list corresponding to the Mth video frame; the video correction text in the Mth video frame has a semantic association with the video text in the target text list In the case where the video correction text in the Mth video frame is added to the target text list, and the list position information of the target text list is updated to update the list set, the target text list being the second text list corresponding to the Mth video frame; in the case where there is no semantic association between the video correction text in the Mth video frame and the video text in the target text list, a fourth text list is newly created, the video correction text in the Mth video frame is added to the fourth text list to determine the second text list corresponding to the Mth video frame, and the second text list corresponding to the Mth video frame and the corresponding list position information are added to the list set to update the list set.

[0213] Optionally, when K target video text lists are screened out from the N video text lists based on the semantic association between the N video text lists and the video title corresponding to the target video, the processor 810 is further used to: obtain text vectors corresponding to the N video text lists and title vectors corresponding to the video titles; for each of the text vectors, calculate the similarity weight between the text vector and the title vector, and obtain a similarity score based on the similarity weight; and based on the similarity scores corresponding to the N video text lists, screen out K target video text lists whose semantic association with the video title meets preset conditions.

[0214] In this way, after obtaining the video texts corresponding to multiple video frames of the target video, text correction is performed on the video texts to ensure the integrity and coherence of the text semantics; after text correction, for multiple video frames, based on the video correction text and the video text position information corresponding to each video frame, N video text lists are determined, and a video text list including video texts with coherent text semantics can be obtained; by screening out the target video text list based on the semantic correlation between the video text list and the video title, and then generating the target video text corresponding to the target video, video texts with high semantic correlation with the video title can be retained in a variety of video text usage scenarios, thereby achieving effective screening of video texts.

[0215] Furthermore, by segmenting the video text into text segments based on semantic coherence, after segmentation, the text is corrected by adopting corresponding processing strategies based on the types of text segments, so that text correction can be performed using a matching correction method to improve correction efficiency.

[0216] By creating a list set based on the first video frame, determining the text list corresponding to the video correction text in any video frame different from the first video frame and dynamically updating the list set, it is possible to update the list set based on two dimensions: position information and text semantics, so as to determine the video text list based on the updated list set.

[0217] By calculating the similarity weights between the text vector and the title vector, obtaining the similarity score based on the similarity weight, and filtering out the target video text list whose semantic relevance to the video title meets the preset conditions according to the similarity scores corresponding to the N video text lists, effective screening of video texts can be achieved.

[0218] It should be understood that in an embodiment of the present application, the input unit 804 may include a graphics processor (GPU) 8041 and a microphone 8042, and the graphics processor 8041 processes the image data of a static picture or video obtained by an image capture device (such as a camera) in a video capture mode or an image capture mode. The display unit 806 may include a display panel 8061, and the display panel 8061 may be configured in the form of a liquid crystal display, an organic light emitting diode, etc. The user input unit 807 includes a touch panel 8071 and at least one of other input devices 8072. The touch panel 8071 is also called a touch screen. The touch panel 8071 may include two parts: a touch detection device and a touch controller. Other input devices 8072 may include, but are not limited to, a physical keyboard, function keys (such as a volume control button, a switch button, etc.), a trackball, a mouse, and a joystick, which will not be repeated here.

[0219] The memory 809 can be used to store software programs and various data. The memory 809 may mainly include a first storage area for storing programs or instructions and a second storage area for storing data, wherein the first storage area may store an operating system, an application program or instructions required for at least one function (such as a sound playback function, an image playback function, etc.), etc. In addition, the memory 809 may include a volatile memory or a non-volatile memory, or the memory 809 may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), a static random access memory (SRAM), a dynamic random access memory (DRAM), a synchronous dynamic random access memory (SDRAM), a double data rate synchronous dynamic random access memory (DDRSDRAM), an enhanced synchronous dynamic random access memory (ESDRAM), a synchronous link dynamic random access memory (SLDRAM) and a direct memory bus random access memory (DRRAM). The memory 809 in the embodiment of the present application includes but is not limited to these and any other suitable types of memory.

[0220] The processor 810 may include one or more processing units; optionally, the processor 810 integrates an application processor and a modem processor, wherein the application processor mainly processes operations related to an operating system, a user interface, and application programs, and the modem processor mainly processes wireless communication signals, such as a baseband processor. It is understandable that the modem processor may not be integrated into the processor 810.

[0221] An embodiment of the present application also provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, each process of the above-mentioned video text screening method embodiment is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0222] The processor is the processor in the electronic device described in the above embodiment. The readable storage medium includes a computer readable storage medium, such as a computer read-only memory ROM, a random access memory RAM, a magnetic disk or an optical disk.

[0223] An embodiment of the present application further provides a chip, which includes a processor and a communication interface, wherein the communication interface is coupled to the processor, and the processor is used to run programs or instructions to implement the various processes of the above-mentioned video text screening method embodiment, and can achieve the same technical effect. To avoid repetition, it will not be repeated here.

[0224] It should be understood that the chip mentioned in the embodiments of the present application can also be called a system-level chip, a system chip, a chip system or a system-on-chip chip, etc.

[0225] An embodiment of the present application provides a computer program product, which is stored in a storage medium. The program product is executed by at least one processor to implement the various processes of the above-mentioned video text screening method embodiment and can achieve the same technical effect. To avoid repetition, it will not be repeated here.

[0226] It should be noted that, in this article, the terms "comprise", "include" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises one..." does not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the method and device in the embodiment of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be added, omitted or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0227] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, a disk, or an optical disk), and includes a number of instructions for a terminal (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0228] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.

Claims

1. A video text screening method, characterized in that: include: Obtaining video texts and video text position information corresponding to multiple video frames of a target video respectively; Performing text correction processing on the video text to obtain a video-corrected text with coherent text semantics; For the multiple video frames, based on the video correction text and the video text position information corresponding to each video frame, determine N video text lists, where N is an integer greater than or equal to 1; Based on the semantic association between the N video text lists and the video title corresponding to the target video, K target video text lists are screened out from the N video text lists, and a target video text corresponding to the target video is generated according to the video texts corresponding to the K target video text lists, where K is an integer less than or equal to N; Wherein, determining N video text lists based on the video correction text and the video text position information corresponding to each video frame includes: Based on the text semantic coherence of adjacent video frames, the text semantic coherence of a single video frame, and the video text position information corresponding to each video frame, the video correction texts in the multiple video frames are screened and combined to determine N video text lists, wherein video texts associated with previous and next frames are stored in one text list, and video texts in different video text lists have no semantic association.

2. The video text screening method according to claim 1, characterized in that: The performing text correction processing on the video text to obtain a video-corrected text with coherent text semantics includes: For each video frame, segment each video text corresponding to the video frame based on semantic coherence; In a case where the video text includes a target text segment and at least one first text segment to be corrected, taking the target text segment as the center, performing text correction on at least one first text segment to obtain a video correction text corresponding to the video text; In a case where the video text only includes the target text segment, determining that the video text does not need to be corrected; In a case where the video text does not include the target text segment, includes a second text segment and at least one first text segment to be corrected, performing text correction on at least one of the first text segments based on semantic coherence to obtain a video correction text corresponding to the video text; The second text segment is a correct text segment, and the target text segment is a text segment with semantic coherence and the longest character length.

3. The video text screening method according to claim 2, characterized in that: The step of performing text correction on at least one of the first text segments with the target text segment as the center includes: Taking the target text segment as the center, based on character retention, character replacement and character deletion processing methods, the first text segment before the target text segment is corrected by using reverse word order; and / or Taking the target text segment as the center, based on character retention, character replacement and character deletion processing methods, the first text segment located after the target text segment is corrected using positive word order.

4. The video text screening method according to claim 2, characterized in that: The performing text correction on at least one of the first text segments based on semantic coherence includes: For each of the first text segments, based on semantic coherence, a plurality of correction segments are constructed for the first text segment by adopting character retention, character replacement and character deletion processing methods; Based on semantic coherence, measuring the coherence between the plurality of correction segments and the context; According to the measurement result, a target correction segment is determined from the multiple correction segments, and the target correction segment is used to perform text correction on the first text segment.

5. The video text screening method according to claim 1, characterized in that: The step of determining N video text lists for the multiple video frames based on the video correction text and the video text position information corresponding to each video frame includes: For a first video frame among the multiple video frames, generating a list set including L first text lists and list position information of each first text list according to the text semantic coherence of the video-corrected text corresponding to the first video frame and the video text position information, where L is an integer greater than or equal to 1; For the Mth video frame, according to the video text position information corresponding to the Mth video frame and the list set updated based on the (M-1)th video frame, determine the second text list corresponding to the Mth video frame, and dynamically update the list set, wherein M is an integer greater than or equal to 2 and less than or equal to G, G is the number of video frames corresponding to the multiple video frames, and the (G-1)th video frame updates the list set in sequence according to the time sequence corresponding to the video frames; For the Gth video frame, after determining the second text list corresponding to the Gth video frame and dynamically updating the list set, determining the N video text lists according to the N text lists in the list set, where N is greater than or equal to L; Wherein, for video frames other than the first video frame, the list set is updated once based on each video frame.

6. The video text screening method according to claim 5, characterized in that: The step of determining, for the Mth video frame, a second text list corresponding to the Mth video frame according to the video text position information corresponding to the Mth video frame and the list set updated based on the (M-1)th video frame, and dynamically updating the list set comprises: According to the video text position information corresponding to the video correction text in the Mth video frame and the list position information corresponding to the list set updated based on the (M-1)th video frame, detecting whether there is a target text list matching the video correction text in the Mth video frame in the list set; If it does not exist, create a new third text list, add the video correction text in the Mth video frame to the third text list to determine the second text list corresponding to the Mth video frame, and add the second text list corresponding to the Mth video frame and the corresponding list position information to the list set; If so, the video correction text in the Mth video frame is processed according to the semantic relationship between the video correction text in the Mth video frame and the video text in the target text list to determine the second text list corresponding to the Mth video frame and dynamically update the list set.

7. The video text screening method according to claim 6, characterized in that: The processing of the video correction text in the Mth video frame according to the semantic relationship between the video correction text in the Mth video frame and the video text in the target text list to determine a second text list corresponding to the Mth video frame and dynamically update the list set includes: When the video correction text in the Mth video frame is identical to the video text in the target text list, ignoring the video correction text in the Mth video frame, wherein the target text list is a second text list corresponding to the Mth video frame; In the case where there is a semantic association between the video correction text in the Mth video frame and the video text in the target text list, the video correction text in the Mth video frame is added to the target text list, and the list position information of the target text list is updated to update the list set, and the target text list is a second text list corresponding to the Mth video frame; In the case that there is no semantic association between the video correction text in the Mth video frame and the video text in the target text list, a fourth text list is newly created, the video correction text in the Mth video frame is added to the fourth text list to determine the second text list corresponding to the Mth video frame, and the second text list corresponding to the Mth video frame and the corresponding list position information are added to the list set to update the list set.

8. The video text screening method according to claim 1, characterized in that: The step of selecting K target video text lists from the N video text lists based on the semantic association between the N video text lists and the video title corresponding to the target video includes: Obtain text vectors corresponding to the N video text lists and a title vector corresponding to the video title; For each of the text vectors, calculating a similarity weight between the text vector and the title vector, and obtaining a similarity score according to the similarity weight; According to the similarity scores respectively corresponding to the N video text lists, K target video text lists whose semantic relevance to the video title meets preset conditions are screened out.

9. A video text screening device, characterized in that: include: A first acquisition module is used to acquire video texts and video text position information respectively corresponding to a plurality of video frames of a target video; A second acquisition module is used to perform text correction processing on the video text to obtain a video-corrected text with coherent text semantics; A determination module, configured to determine, for the plurality of video frames, N video text lists based on the video correction text and the video text position information corresponding to each video frame, where N is an integer greater than or equal to 1; A processing module, configured to screen out K target video text lists from the N video text lists based on the semantic association between the N video text lists and the video title corresponding to the target video, and generate a target video text corresponding to the target video according to the video texts corresponding to the K target video text lists, where K is an integer less than or equal to N; Among them, the determination module is specifically used to screen and combine the video correction texts in the multiple video frames based on the text semantic coherence of adjacent video frames, the text semantic coherence of a single video frame, and the video text position information corresponding to each video frame to determine N video text lists, wherein the video texts that are associated with the previous and next frames are stored in one text list, and the video texts in different video text lists have no semantic association.

10. The device according to claim 9, characterized in that The second acquisition module includes: A segmentation submodule, for performing text segmentation on each video text corresponding to each video frame based on semantic coherence; A first correction submodule is used for, when the video text includes a target text segment and at least one first text segment to be corrected, taking the target text segment as the center, performing text correction on at least one first text segment to obtain a video correction text corresponding to the video text; A determination submodule, configured to determine that the video text does not need to be corrected when the video text only includes the target text segment; A second correction submodule is used for, when the video text does not include the target text segment, includes a second text segment and at least one first text segment to be corrected, performing text correction on at least one of the first text segments based on semantic coherence to obtain a video correction text corresponding to the video text; The second text segment is a correct text segment, and the target text segment is a text segment with semantic coherence and the longest character length.

11. The device according to claim 10, characterized in that The first correction submodule comprises: A first correction unit is configured to correct the first text segment preceding the target text segment by using a reverse word order based on character retention, character replacement, and character deletion processing modes with respect to the target text segment; and / or The second correction unit is used to perform text correction on the first text segment located after the target text segment by using a positive word order based on character retention, character replacement and character deletion processing methods with the target text segment as the center.

12. The device according to claim 10, characterized in that The second correction submodule comprises: a construction unit, configured to construct a plurality of correction segments for each of the first text segments based on semantic coherence by adopting character retention, character replacement and character deletion processing methods; A measuring unit, used for measuring the coherence between the plurality of corrected segments and the context based on semantic coherence; The third correction unit is used to determine a target correction segment from the multiple correction segments according to the measurement result, and use the target correction segment to perform text correction on the first text segment.

13. The device according to claim 9, characterized in that The determination module comprises: A generating submodule, for generating, for a first video frame among the plurality of video frames, a list set including L first text lists and list position information of each first text list according to the text semantic coherence of the video-corrected text corresponding to the first video frame and the video text position information, where L is an integer greater than or equal to 1; A first processing submodule is used to determine, for an M-th video frame, a second text list corresponding to the M-th video frame according to the video text position information corresponding to the M-th video frame and the list set updated based on the (M-1)-th video frame, and dynamically update the list set, wherein M is an integer greater than or equal to 2 and less than or equal to G, G is the number of video frames corresponding to the multiple video frames, and the (G-1)-th video frame updates the list set in sequence according to the time sequence corresponding to the video frames; A second processing submodule is used for determining, for a G-th video frame, a second text list corresponding to the G-th video frame and dynamically updating the list set, and then determining the N video text lists according to the N text lists in the list set, where N is greater than or equal to L; Wherein, for video frames other than the first video frame, the list set is updated once based on each video frame.

14. The device according to claim 13, characterized in that The first processing submodule includes: a detection unit, configured to detect whether there is a target text list matching the video correction text in the Mth video frame in the list set according to the video text position information corresponding to the video correction text in the Mth video frame and the list position information corresponding to the list set updated based on the (M-1)th video frame; A first processing unit is configured to create a new third text list if it does not exist, add the video correction text in the Mth video frame to the third text list to determine a second text list corresponding to the Mth video frame, and add the second text list corresponding to the Mth video frame and corresponding list position information to the list set; The second processing unit is used to process the video correction text in the Mth video frame according to the semantic relationship between the video correction text in the Mth video frame and the video text in the target text list, if any, so as to determine a second text list corresponding to the Mth video frame and dynamically update the list set.

15. The device according to claim 14, characterized in that The second processing unit comprises: A first processing subunit is configured to ignore the video correction text in the Mth video frame when the video correction text in the Mth video frame has the same content as the video text in the target text list, wherein the target text list is a second text list corresponding to the Mth video frame; a second processing subunit, configured to add the video correction text in the Mth video frame to the target text list and update the list position information of the target text list to update the list set, when there is a semantic association between the video correction text in the Mth video frame and the video text in the target text list, wherein the target text list is a second text list corresponding to the Mth video frame; The third processing sub-unit is used to create a fourth text list when there is no semantic association between the video correction text in the Mth video frame and the video text in the target text list, add the video correction text in the Mth video frame to the fourth text list to determine a second text list corresponding to the Mth video frame, and add the second text list corresponding to the Mth video frame and the corresponding list position information to the list set to update the list set.

16. The device according to claim 9, characterized in that The processing module comprises: An acquisition submodule, used to acquire text vectors corresponding to the N video text lists and a title vector corresponding to the video title; A calculation and acquisition submodule, used for calculating the similarity weight between the text vector and the title vector for each of the text vectors, and obtaining a similarity score according to the similarity weight; The screening submodule is used to screen out K target video text lists whose semantic relevance to the video title meets preset conditions according to the similarity scores corresponding to the N video text lists.

17. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores a program or instruction that can be run on the processor, and when the program or instruction is executed by the processor, the steps of the video text screening method as described in any one of claims 1 to 8 are implemented.

Images