Intelligent auditing method and device based on target detection and multi-modal large model

Abstract

The invention discloses an intelligent auditing method and device based on target detection and a multi-modal large model, and relates to the technical field of cloud resource evaluation. Comprising the following steps: 1, collecting a data set; 2, setting up a text auditing service; step 3, setting up an audio auditing service; 4, training a target detection auditing model; 5, setting up an image and video auditing service; by fusing computer vision and natural language processing technologies, efficient and accurate auditing of multi-modal data such as texts, images, videos and the like is realized.

Description

Technical Field

[0001] The present invention discloses an intelligent audit method and device based on target detection and a multimodal large model, and relates to the technical field of cloud resource evaluation. Background Art

[0002] With the acceleration of digital transformation, data such as text, images, audio, and video is exploding, and the drawbacks of existing manual review methods in terms of efficiency, cost, and accuracy are becoming increasingly prominent. In the field of content review, manual review is not only time-consuming and labor-intensive, but also susceptible to subjective factors, making it difficult to meet the real-time review needs of massive amounts of data. Existing intelligent review technologies also have limitations. Rule-based review systems lack flexibility and are difficult to adapt to complex and ever-changing review scenarios. Existing machine learning models rely on large amounts of labeled data and have weak generalization capabilities. Early deep learning models were ineffective in processing small samples and multimodal data, and were unable to adapt to the diverse data types and complex scenarios in real business operations. Summary of the Invention

[0003] In response to the problems of the existing technology, the present invention provides an intelligent audit method and device based on target detection and multimodal large models. By integrating computer vision and natural language processing technology, it can achieve efficient and accurate audit of multimodal data such as text, images, and videos.

[0004] The specific scheme proposed by the present invention is: The present invention provides an intelligent audit method based on target detection and a multimodal large model, comprising: Step 1: Collect data sets: Use online searches to collect text, audio, image, and video data containing illegal elements to form a data set for training the illegal review model; Step 2: Build a text review service: This service is based on a text error correction model and database. The text error correction model is used to find errors in text content and provide correct results. The database is used for sensitive word query matching. Ultimately, a text review service is formed that includes text error correction, ad detection, and sensitive word query. Step 3: Build an audio review service: This service converts audio to text based on an audio recognition model, then searches for errors in the text and presents the correct answer. It also performs sensitive word search and matching, ultimately forming an audio review service that includes text error correction, ad detection, and sensitive word search. Step 4: Train the object detection audit model: For the collected images, use image annotation tools to classify and annotate the illegal elements in the images. Based on the YOLO object detection pre-trained model, train and fine-tune multiple types of audit models to obtain the illegal element audit model for image and video auditing. Step 5: Build image and video review services: Use the inference server to deploy an internal visual service for the review model of illegal elements. The internal visual service detects illegal behaviors and builds image and video review services based on the visual service. Face detection involves database information comparison, and OCR detection involves text review. During video review, the video is first decomposed into audio stream and image stream. Audio review is performed on the audio stream, and frame extraction is performed on the image stream at a fixed frequency. Image review is performed on the extracted image frames. Finally, the audio review and image review results are combined as the result of the video review.

[0005] Furthermore, in step 1 of the intelligent audit method based on target detection and multimodal large models, data is screened from existing public data resources, using Hate Speech Dataset and Twitter Sentiment140 data resources in the text field; COCO and ImageNet data resources in the image field; YouTube-8M data resources in the video field; and LibriSpeech data resources in the audio field. Relevant data is extracted based on the violation category requirements. Text data contains sensitive words, typos and grammatical errors, advertising content, and other violation information. Audio data contains voice content violations and audio feature violations. Image or video data contains prohibited logos, advertisements, watermarks, and other violation scenarios.

[0006] Furthermore, in step 2 of the intelligent audit method based on target detection and multimodal large models, when constructing a text audit service based on a text error correction model and a database, text error detection and correction are implemented based on the text error correction model: grammatical errors and spelling errors in Chinese text are detected; inappropriate word usage and collocation errors are identified and corrected; the original error text, the corrected text and the error position mark are output; and error type classification is provided, which includes spelling errors, grammatical errors and inappropriate word usage; Use the database's full-text indexing function to match sensitive words; use prefix, suffix, and inclusion matching modes for matching; classify sensitive words and adopt different handling strategies based on risk levels; provide batch import and export of sensitive word libraries and real-time update functions; Identify explicit advertising features of URLs, phone numbers, and email addresses based on regular expressions; train classification models to identify marketing language and implicit advertising content that induces sharing; customize advertising keywords and blocking rules; and identify variant ads.

[0007] Furthermore, step 3 of the intelligent audit method based on target detection and multimodal large model specifically includes: using an audio recognition model to perform voice activity detection: converting the audio format into a mono WAV format with a 16kHz sampling rate and a 16-bit depth; Audio framing: split the audio into frames of fixed length; Perform model inference: Perform voice activity detection on each frame to determine whether it contains valid speech; Merge voice segments: Merge adjacent valid voice segments and filter out the silent parts; Segment caching and batch processing: Speech segments are cached sequentially and batch processed when they reach a certain length; Perform speech recognition and text generation, and use audio recognition models to achieve speech-to-text conversion: Feature extraction: extract acoustic features from audio clips; perform speech recognition: use audio recognition models to convert acoustic features into text sequences; Punctuation recovery: use audio recognition models to add punctuation to recognized text; perform digital normalization post-processing on the text.

[0008] Furthermore, step 4 of the intelligent audit method based on target detection and multimodal large model specifically includes: Perform image data preprocessing: After collecting sensitive images, uniformly scale the image pixel values ​​to the range of 0-1; adopt a unified scaling strategy for images with large resolution differences; and perform data enhancement operations on the images; Use YOLO's TXT format to store annotation information. When using the annotation tool to draw a rectangular box to select the offending element, record the coordinates of the upper left and lower right corners of the rectangular box and bind the coordinates to the corresponding offending element category label. When dividing the labeled violation image dataset into training, validation, and test sets, a ratio of 7:1:2 or 8:1:1 is used; During the target detection audit model training phase, stochastic gradient descent (SGD) and its variants are used as optimizers to update model parameters by minimizing the loss function.

[0009] Fine-tune the object detection audit model by unfreezing some layers of the pre-trained model for specific violation element categories, allowing the parameters of these layers to be updated during the training process; In image review applications, the image to be reviewed is pre-processed and input into a trained target detection review model. The target detection review model outputs the category, location, and confidence score of the detected sensitive elements. According to the preset confidence threshold, the detection results with low confidence are filtered out, and the detection results that meet the threshold requirements are visualized on the image, marking the location and category of the illegal elements.

[0010] The present invention also provides an intelligent audit device based on target detection and multimodal large models, including an acquisition module, a text service building module, an audio service building module, a model training module and an image service building module. The acquisition module collects data sets: using online searches to collect text, audio, image, and video data containing illegal elements to form a data set for training the violation review model; The text service building module builds a text review service: This service is built based on a text error correction model and a database. The text error correction model is used to find errors in text content and provide correct results. The database is used for sensitive word query matching, ultimately forming a text review service that includes text error correction, advertisement detection, and sensitive word query. The audio service building module builds an audio review service: it converts audio into text based on the audio recognition model, then finds errors in the text content and gives the correct results, performs sensitive word query matching, and ultimately forms an audio review service that includes text error correction, advertisement detection, and sensitive word query; The model training module trains the object detection audit model: for the collected images, it uses image annotation tools to classify and annotate the illegal elements in the images. Based on the YOLO object detection pre-trained model, it trains and fine-tunes multiple types of audit models to obtain the audit model for illegal elements for image and video auditing. The image service building module builds image and video review services: the review model of illegal elements uses an inference server to deploy a set of internal visual services, the internal visual services detect illegal behaviors, and build image and video review services based on the visual services. Among them, face detection involves database information comparison, and OCR detection involves text review; during video review, the video is first disassembled into audio stream and image stream, audio review is performed on the audio stream, fixed-frequency frame extraction is performed on the image stream, and image review is performed on the extracted image frames. Finally, the audio review and image review results are combined as the result of the video review.

[0011] Furthermore, the acquisition module of the intelligent audit device based on target detection and multimodal large model filters data from existing public data resources, using Hate Speech Dataset and Twitter Sentiment140 data resources in the text field; COCO and ImageNet data resources in the image field; YouTube-8M data resources in the video field; and LibriSpeech data resources in the audio field. Relevant data is extracted based on the violation category requirements. Text data contains sensitive words, typos and grammatical errors, advertising content, and other violation information. Audio data contains voice content violations and audio feature violations. Image or video data contains prohibited logos, advertisements, watermarks, and other violation scenarios.

[0012] Furthermore, when the text service building module of the intelligent audit device based on target detection and multimodal large model builds a text audit service based on a text error correction model and a database, text error detection and correction are realized based on the text error correction model: grammatical errors and spelling errors of Chinese text are detected; inappropriate word usage and collocation errors are identified and corrected; the original error text, the corrected text and the error position mark are output; and error type classification is provided, which includes spelling errors, grammatical errors and inappropriate word usage. Use the database's full-text indexing function to match sensitive words; use prefix, suffix, and inclusion matching modes for matching; classify sensitive words and adopt different handling strategies based on risk levels; provide batch import and export of sensitive word libraries and real-time update functions; Identify explicit advertising features of URLs, phone numbers, and email addresses based on regular expressions; train classification models to identify marketing language and implicit advertising content that induces sharing; customize advertising keywords and blocking rules; and identify variant ads.

[0013] Furthermore, the audio service building module of the intelligent audit device based on target detection and multimodal large model uses an audio recognition model to perform voice activity detection: the audio format is uniformly converted into a mono WAV format with a 16kHz sampling rate and a 16-bit depth; Audio framing: split the audio into frames of fixed length; Perform model inference: Perform voice activity detection on each frame to determine whether it contains valid speech; Merge voice segments: Merge adjacent valid voice segments and filter out the silent parts; Segment caching and batch processing: Speech segments are cached sequentially and batch processed when they reach a certain length; Perform speech recognition and text generation, and use audio recognition models to achieve speech-to-text conversion: Feature extraction: extract acoustic features from audio clips; perform speech recognition: use audio recognition models to convert acoustic features into text sequences; Punctuation recovery: use audio recognition models to add punctuation to recognized text; perform digital normalization post-processing on the text.

[0014] Furthermore, the model training module of the intelligent audit device based on target detection and multimodal large model performs image data preprocessing: after collecting sensitive images, the images are normalized and the pixel values ​​of the images are uniformly scaled to the range of 0-1; for images with large resolution differences, a unified scaling strategy is adopted; and data enhancement operations are performed on the images; Use YOLO's TXT format to store annotation information. When using the annotation tool to draw a rectangular box to select the offending element, record the coordinates of the upper left and lower right corners of the rectangular box and bind the coordinates to the corresponding offending element category label. When dividing the labeled violation image dataset into training, validation, and test sets, a ratio of 7:1:2 or 8:1:1 is used; During the target detection audit model training phase, stochastic gradient descent (SGD) and its variants are used as optimizers to update model parameters by minimizing the loss function.

[0015] Fine-tune the object detection audit model by unfreezing some layers of the pre-trained model for specific violation element categories, allowing the parameters of these layers to be updated during the training process; In image review applications, the image to be reviewed is pre-processed and input into a trained target detection review model. The target detection review model outputs the category, location, and confidence score of the detected sensitive elements. According to the preset confidence threshold, the detection results with low confidence are filtered out, and the detection results that meet the threshold requirements are visualized on the image, marking the location and category of the illegal elements.

[0016] The benefit of the present invention is that it can achieve efficient and accurate review of multimodal data such as text, images, and videos by integrating computer vision and natural language processing technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 It is a schematic flow chart of the method of the present invention. DETAILED DESCRIPTION

[0019] The present invention will be further described below with reference to the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments are not intended to limit the present invention.

[0020] Example 1: The present invention provides an intelligent audit method based on target detection and a multimodal large model, comprising: Step 1: Collect data sets: Use web searches to collect text, audio, image, and video data containing illegal elements to form a data set for training the illegal review model.

[0021] Specifically, dataset collection involves selecting appropriate content from existing public data resources, such as the Hate Speech Dataset and Twitter Sentiment140 for text; COCO and ImageNet for images; YouTube-8M for video; and LibriSpeech for audio. Relevant data is extracted based on violation categories. Text data includes sensitive words, typos and grammatical errors, advertising content, and other illegal information. Sensitive words include terrorist-related terms and abusive language; typos and grammatical errors include a database of common typos and examples of grammatical errors; advertising content includes spam and marketing tactics; and other illegal information includes rumors and inflammatory speech. Audio data primarily includes speech content violations and audio feature violations. Speech content violations include speech containing sensitive words and hate speech; audio feature violations include terrifying sound effects. Image or video data primarily includes violent content, prohibited symbols, special clothing, vulgar content, advertisements and watermarks, and other illegal scenarios. For example, dangerous goods include explosives; prohibited symbols include flags of extremist organizations; and advertisements and watermarks include unauthorized commercial watermarks.

[0022] Step 2: Build a text review service: Build a text review service based on a text error correction model and database. The text error correction model is used to find errors in text content and give correct results. The database is used for sensitive word query matching, ultimately forming a text review service that includes text error correction, advertisement detection, and sensitive word query.

[0023] Step 2 may specifically include: building a text review service based on a text error correction model and database, which requires systematic integration of natural language processing technology and high-performance data storage to achieve full-process automated review from text error detection to sensitive information screening.

[0024] Text error detection and correction based on a text error correction model: supports grammatical and spelling error detection in Chinese text; can identify and correct problems such as inappropriate word usage and incorrect collocation; outputs the original error text, corrected text, and error location annotations; and provides error type classification such as spelling errors, grammatical errors, and inappropriate word usage.

[0025] Sensitive word detection is a high-performance sensitive word detection system built on a database: it uses the full-text indexing function of the database to achieve efficient sensitive word matching; supports multiple matching modes such as prefix, suffix, and containment; the sensitive word library supports classification management such as political sensitivity, vulgarity, and terror; supports sensitive word grading, and adopts different processing strategies according to risk levels; and provides batch import and export of sensitive word libraries and real-time update functions.

[0026] Ad detection can combine rule engines and machine learning algorithms to detect ad content: identify explicit ad features such as URLs, phone numbers, and email addresses based on regular expressions; train classification models to identify implicit ad content such as marketing rhetoric and inducement to share; support custom ad keywords and blocking rules; and be able to identify variant ads such as homophones, pinyin abbreviations, etc.

[0027] Through the above solution, a high-performance, high-accuracy, and easily scalable text review service is built to effectively identify errors, sensitive information, and advertising content in texts, providing safe and reliable text review guarantees.

[0028] Step 3: Build an audio review service: Convert audio to text based on the audio recognition model, then find errors in the text content and give the correct results, perform sensitive word query matching, and finally form an audio review service that includes text error correction, advertisement detection, and sensitive word query.

[0029] Step 3 may specifically include: building an audio review service based on the audio recognition model, which requires systematically integrating speech recognition, audio processing and text review technologies to achieve full process automation from audio input to review result output.

[0030] The audio recognition model is used for voice activity detection. The implementation steps are as follows: Audio format conversion: Convert the input audio to a mono WAV format with a 16kHz sampling rate and 16-bit depth; Audio framing: Segment the audio into frames of a fixed length, such as 10ms / frame; Model inference: Perform voice activity detection on each frame to determine whether it contains valid speech; Voice segment merging: Merge adjacent valid voice segments and filter out silent portions; Segment caching and batch processing: Cache voice segments sequentially and perform batch processing after they reach a certain length.

[0031] Speech recognition and text generation, using audio recognition models to achieve speech-to-text conversion: Feature extraction: extracting acoustic features such as MFCC, Fbank, etc. from audio clips; Speech recognition: using models to convert acoustic features into text sequences; Punctuation recovery: using models to add punctuation to recognized text; Text post-processing: digital normalization; entity recognition and standardization; language error correction.

[0032] By integrating speech recognition results into existing text review services, we can build a high-performance, high-accuracy, and easily scalable audio review service through the above solution. This can effectively identify illegal content in audio and provide safe and reliable audio review guarantees.

[0033] Step 4: Train the target detection audit model: For the collected images, use image annotation tools to classify and annotate the illegal elements in the images, train and fine-tune multiple types of audit models based on the YOLO target detection pre-trained model, and obtain the audit model of illegal elements for image and video auditing.

[0034] Step 4 may specifically include: Preprocessing of sensitive image data: After collecting sensitive images, a series of technical processing is required to meet the needs of model training. First, the image is normalized and the pixel values ​​of the image are uniformly scaled to the range of 0-1, which helps to improve the convergence speed and stability of model training. For images with large differences in resolution, a unified scaling strategy is adopted, such as maintaining the aspect ratio and scaling to a fixed size such as 460×460 pixels to ensure that the image size input to the model is consistent. In addition, in order to increase the diversity of the data and prevent the model from overfitting, data enhancement operations are also performed on the image, including random flipping, random cropping, rotation, brightness adjustment, noise addition, etc. These operations expand the data set so that the model can learn the characteristics of sensitive elements under different perspectives and conditions.

[0035] The annotation tool can use the YOLO-specific TXT format to store annotation information. When the annotator uses the annotation tool to draw a rectangular box to select a sensitive element, the tool will record the coordinates of the upper left corner and lower right corner of the rectangular box (x min ,y min ,x max ,y max ) and binds it to the corresponding sensitive element category label. When annotating special clothing, image feature extraction techniques, such as the HOG Histogram of Oriented Gradients or the SIFT algorithm, assist annotators in more accurately identifying clothing styles and features. For flags that violate regulations, color histograms and template matching techniques are used to quickly locate and identify flag patterns. Once annotation is complete, the generated annotation file corresponds one-to-one with the original image, forming a complete annotation dataset.

[0036] The YOLO (You Only Look Once) model series utilizes an end-to-end detection architecture, with its network structure primarily consisting of a backbone network (Backbone), a neck network (Neck), and a head network (Head). The backbone network is responsible for extracting basic image features and typically utilizes the CSPDarknet architecture. Through the Cross-Stage Local Network (CSPNet) design, it reduces computational effort while improving feature extraction capabilities. The neck network utilizes a Path Aggregation Network (PANet) to fuse feature maps of different scales, enhancing detection capabilities for objects of varying sizes. The head network outputs target category and location predictions. The YOLO model, based on an anchor mechanism, predicts objects in an image using pre-set anchor boxes of varying sizes. Each anchor box is assigned a confidence score and category probability. The model uses the non-maximum suppression (NMS) algorithm to remove duplicate detection boxes to obtain the final detection result.

[0037] When dividing the labeled sensitive image dataset into training, validation, and test sets, a ratio of 7:1:2 or 8:1:1 is generally used. Stratified sampling techniques are used during the division process to ensure that the proportion of sensitive elements in each category in each dataset is similar to that in the original dataset, avoiding model training bias due to category imbalance. Furthermore, to further evaluate the model's generalization ability, K-fold cross-validation can be used. The dataset is divided into K parts, with one part used as the validation set and the remaining K-1 parts used as the training set. The model is trained and validated multiple times, and the average performance indicator is used as the final evaluation result.

[0038] During model training, stochastic gradient descent (SGD) and its variants, such as Adam and Adagrad, are used as optimizers to update model parameters by minimizing the loss function. The YOLO model's loss function primarily consists of three parts: object confidence loss, category prediction loss, and bounding box regression loss. The total loss is calculated through a weighted summation. During training, the batch size is appropriately set based on the dataset size and computing resources. A larger batch size results in faster model training, but also requires more memory. A balance must be struck between training speed and memory usage.

[0039] During the fine-tuning phase, for specific categories of sensitive elements, some layers of the pre-trained model are unfrozen, allowing the parameters of these layers to be updated during the training process. For example, for the training of an audit model for sensitive elements such as firearms, the last few layers of the backbone network and the head network can be unfrozen, and the parameters of these layers can be optimized to focus more on learning the characteristics of firearms. At the same time, by adjusting the learning rate, a smaller learning rate is used in the early stages of fine-tuning to prevent excessive updates of model parameters that lead to performance degradation. As training progresses, the learning rate is gradually increased to speed up convergence. During the training process, performance indicators on the validation set, such as accuracy, recall rate, and F1 value, are monitored in real time. When the performance of the validation set no longer improves or shows a downward trend, training is stopped to avoid overfitting.

[0040] In image review applications, pre-processed images are fed into a trained review model, which then outputs the category, location, and confidence score of any detected sensitive elements. Based on a preset confidence threshold, detection results with low confidence levels are filtered out, and detection results that meet the threshold are visualized on the image, annotating the location and category of sensitive elements.

[0041] Step 5: Build image and video review services: Use the inference server to deploy an internal visual service for the review model of illegal elements. The internal visual service detects illegal behaviors and builds image and video review services based on the visual service. Face detection involves database information comparison, and OCR detection involves text review. During video review, the video is first decomposed into audio stream and image stream. Audio review is performed on the audio stream, and frame extraction is performed on the image stream at a fixed frequency. Image review is performed on the extracted image frames. Finally, the audio review and image review results are combined as the result of the video review. Example

[0042] The present invention also provides an intelligent audit device based on target detection and multimodal large models, including an acquisition module, a text service building module, an audio service building module, a model training module and an image service building module. The acquisition module collects data sets: using online searches to collect text, audio, image, and video data containing illegal elements to form a data set for training the violation review model; The text service building module builds a text review service: This service is built based on a text error correction model and a database. The text error correction model is used to find errors in text content and provide correct results. The database is used for sensitive word query matching, ultimately forming a text review service that includes text error correction, advertisement detection, and sensitive word query. The audio service building module builds an audio review service: it converts audio into text based on the audio recognition model, then finds errors in the text content and gives the correct results, performs sensitive word query matching, and ultimately forms an audio review service that includes text error correction, advertisement detection, and sensitive word query; The model training module trains the object detection audit model: for the collected images, it uses image annotation tools to classify and annotate the illegal elements in the images. Based on the YOLO object detection pre-trained model, it trains and fine-tunes multiple types of audit models to obtain the audit model for illegal elements for image and video auditing. The image service building module builds image and video review services: the review model of illegal elements uses an inference server to deploy a set of internal visual services, the internal visual services detect illegal behaviors, and build image and video review services based on the visual services. Among them, face detection involves database information comparison, and OCR detection involves text review; during video review, the video is first disassembled into audio stream and image stream, audio review is performed on the audio stream, fixed-frequency frame extraction is performed on the image stream, and image review is performed on the extracted image frames. Finally, the audio review and image review results are combined as the result of the video review.

[0043] Since the information interaction, execution process and other contents between the modules in the above-mentioned device are based on the same concept as the embodiment of the method of the present invention, the specific contents can be found in the description of the embodiment of the method of the present invention and will not be repeated here.

[0044] Similarly, the device of the present invention achieves efficient and accurate review of multimodal data such as text, images, and videos by integrating computer vision and natural language processing technology.

[0045] It should be noted that not all steps and modules in the above-mentioned processes and device structures are required, and certain steps or modules can be omitted according to actual needs. The execution order of each step is not fixed and can be adjusted as needed. The system structure described in the above-mentioned embodiments can be a physical structure or a logical structure, that is, some modules may be implemented by the same physical entity, or some modules may be implemented by multiple physical entities, or may be implemented by certain components in multiple independent devices.

[0046] The above embodiments are merely preferred embodiments for the purpose of fully illustrating the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are within the scope of protection of the present invention. The scope of protection of the present invention shall be subject to the claims.

Claims

1. An intelligent audit method based on target detection and multimodal large model, characterized by include: Step 1: Collect data sets: Use online searches to collect text, audio, image, and video data containing illegal elements to form a data set for training the illegal review model; Step 2: Build a text review service: This service is based on a text error correction model and database. The text error correction model is used to find errors in text content and provide correct results. The database is used for sensitive word query matching. Ultimately, a text review service is formed that includes text error correction, ad detection, and sensitive word query. Step 3: Build an audio review service: This service converts audio to text based on an audio recognition model, then searches for errors in the text and presents the correct answer. It also performs sensitive word search and matching, ultimately forming an audio review service that includes text error correction, ad detection, and sensitive word search. Step 4: Train the object detection audit model: For the collected images, use image annotation tools to classify and annotate the illegal elements in the images. Based on the YOLO object detection pre-trained model, train and fine-tune multiple types of audit models to obtain the illegal element audit model for image and video auditing. Step 5: Build image and video review services: Use the inference server to deploy an internal visual service for the review model of illegal elements. The internal visual service detects illegal behaviors and builds image and video review services based on the visual service. Face detection involves database information comparison, and OCR detection involves text review. During video review, the video is first decomposed into audio stream and image stream. Audio review is performed on the audio stream, and frame extraction is performed on the image stream at a fixed frequency. Image review is performed on the extracted image frames. Finally, the audio review and image review results are combined as the result of the video review.

2. The intelligent audit method based on target detection and multimodal large models according to claim 1 is characterized in that in step 1, data is screened from existing public data resources, using Hate Speech Dataset and Twitter Sentiment140 data resources in the text field; COCO and ImageNet data resources in the image field; YouTube-8M data resources in the video field; and LibriSpeech data resources in the audio field. Relevant data is extracted based on the violation category requirements. Text data contains sensitive words, typos and grammatical errors, advertising content, and other violation information. Audio data contains voice content violations and audio feature violations. Image or video data contains prohibited logos, advertisements, watermarks, and other violation scenarios.

3. The intelligent audit method based on target detection and multimodal large model according to claim 1 is characterized by When building a text review service based on the text error correction model and database in step 2, text error detection and correction are implemented based on the text error correction model: grammatical errors and spelling errors in Chinese text are detected; inappropriate word usage and incorrect collocation are identified and corrected; Output the original error text, the corrected text, and the error location mark; Provide error type classification, including spelling errors, grammatical errors, and inappropriate word usage; Use the database's full-text indexing function to match sensitive words; use prefix, suffix, and inclusion matching patterns for matching; classify sensitive words and adopt different handling strategies based on risk levels; Provides batch import and export of sensitive word libraries and real-time update functions; Identify explicit advertising features of URLs, phone numbers, and email addresses based on regular expressions; Train classification models to identify marketing rhetoric and hidden advertising content that induces sharing; customize advertising keywords and blocking rules; and identify variant ads.

4. The intelligent audit method based on target detection and multimodal large model according to claim 1 is characterized by Step 3 specifically includes: using the audio recognition model for voice activity detection: converting the audio format to a mono WAV format with a 16kHz sampling rate and 16-bit depth; Audio framing: split the audio into frames of fixed length; Perform model inference: Perform voice activity detection on each frame to determine whether it contains valid speech; Merge voice segments: Merge adjacent valid voice segments and filter out the silent parts; Cache and batch process segments: cache speech segments in sequence and process them in batches when they reach a certain length; Perform speech recognition and text generation, and use audio recognition models to achieve speech-to-text conversion: Feature extraction: extract acoustic features from audio clips; perform speech recognition: use audio recognition models to convert acoustic features into text sequences; Punctuation recovery: use audio recognition models to add punctuation to recognized text; perform digital normalization post-processing on the text.

5. The intelligent audit method based on target detection and multimodal large model according to claim 1 is characterized by Step 4 specifically includes: Perform image data preprocessing: After collecting sensitive images, normalize the images and uniformly scale the pixel values ​​to the range of 0-1. For images with large resolution differences, adopt a unified scaling strategy; and perform data enhancement operations on the images. Use YOLO's TXT format to store annotation information. When using the annotation tool to draw a rectangular box to select the offending element, record the coordinates of the upper left and lower right corners of the rectangular box and bind the coordinates to the corresponding offending element category label. When dividing the labeled violation image dataset into training, validation, and test sets, a ratio of 7:1:2 or 8:1:1 is used; During the object detection audit model training phase, stochastic gradient descent (SGD) and its variants are used as optimizers to update model parameters by minimizing the loss function. Fine-tune the object detection audit model by unfreezing some layers of the pre-trained model for specific violation element categories, allowing the parameters of these layers to be updated during the training process; In image review applications, the image to be reviewed is pre-processed and input into a trained target detection review model. The target detection review model outputs the category, location, and confidence score of the detected sensitive elements. According to the preset confidence threshold, the detection results with low confidence are filtered out, and the detection results that meet the threshold requirements are visualized on the image, marking the location and category of the illegal elements.

6. An intelligent audit device based on target detection and multimodal large model, characterized by Including collection module, text service building module, audio service building module, model training module and image service building module, The acquisition module collects data sets: using online searches to collect text, audio, image, and video data containing illegal elements to form a data set for training the violation review model; The text service building module builds a text review service: This service is built based on a text error correction model and a database. The text error correction model is used to find errors in text content and provide correct results. The database is used for sensitive word query matching, ultimately forming a text review service that includes text error correction, advertisement detection, and sensitive word query. The model training module trains the object detection audit model: for the collected images, it uses image annotation tools to classify and annotate the illegal elements in the images. Based on the YOLO object detection pre-trained model, it trains and fine-tunes multiple types of audit models to obtain the audit model for illegal elements for image and video auditing. The image service building module builds image and video review services: the review model of illegal elements uses an inference server to deploy a set of internal visual services, the internal visual services detect illegal behaviors, and build image and video review services based on the visual services. Among them, face detection involves database information comparison, and OCR detection involves text review; during video review, the video is first disassembled into audio stream and image stream, audio review is performed on the audio stream, fixed-frequency frame extraction is performed on the image stream, and image review is performed on the extracted image frames. Finally, the audio review and image review results are combined as the result of the video review.

7. The intelligent audit device based on target detection and multimodal large model according to claim 6 is characterized by The acquisition module selects data from existing public data resources. In the text field, it uses Hate Speech Dataset and Twitter Sentiment140 data resources; in the image field, it uses COCO and ImageNet data resources; in the video field, it uses YouTube-8M data resources; in the audio field, it uses LibriSpeech data resources. Relevant data is extracted based on the violation category requirements. Text data contains sensitive words, typos and grammatical errors, advertising content, and other violation information. Audio data contains voice content violations and audio feature violations. Image or video data contains prohibited logos, advertisements, watermarks, and other violation scenarios.

8. The intelligent audit device based on target detection and multimodal large model according to claim 6 is characterized in that the text When the service building module builds a text review service based on the text error correction model and database, it implements text error detection and correction based on the text error correction model: detecting grammatical errors and spelling errors in Chinese texts; identifying and correcting inappropriate word usage and collocation errors; Output the original error text, the corrected text, and the error location mark; Provide error type classification, including spelling errors, grammatical errors, and inappropriate word usage; Use the database's full-text indexing function to match sensitive words; use prefix, suffix, and inclusion matching patterns for matching; classify sensitive words and adopt different handling strategies based on risk levels; Provides batch import and export of sensitive word libraries and real-time update functions; Identify explicit advertising features of URLs, phone numbers, and email addresses based on regular expressions; Train classification models to identify marketing rhetoric and hidden advertising content that induces sharing; customize advertising keywords and blocking rules; and identify variant ads.

9. The intelligent audit device based on target detection and multimodal large model according to claim 6 is characterized by The audio service building module uses an audio recognition model for voice activity detection: it converts the audio format into a mono WAV format with a 16kHz sampling rate and 16-bit depth; Audio framing: split the audio into frames of fixed length; Perform model inference: Perform voice activity detection on each frame to determine whether it contains valid speech; Merge voice segments: Merge adjacent valid voice segments and filter out the silent parts; Segment caching and batch processing: Speech segments are cached sequentially and batch processed when they reach a certain length; Perform speech recognition and text generation, and use audio recognition models to achieve speech-to-text conversion: Feature extraction: extract acoustic features from audio clips; perform speech recognition: use audio recognition models to convert acoustic features into text sequences; Punctuation recovery: use audio recognition models to add punctuation to recognized text; perform digital normalization post-processing on the text.

10. The intelligent audit device based on target detection and multimodal large model according to claim 6 is characterized by: The model training module performs image data preprocessing: after collecting sensitive images, the images are normalized and their pixel values ​​are uniformly scaled to the range of 0-1; for images with large resolution differences, a unified scaling strategy is adopted; and data enhancement operations are performed on the images. Use YOLO's TXT format to store annotation information. When using the annotation tool to draw a rectangular box to select the offending element, record the coordinates of the upper left and lower right corners of the rectangular box and bind the coordinates to the corresponding offending element category label. When dividing the labeled violation image dataset into training, validation, and test sets, a ratio of 7:1:2 or 8:1:1 is used; During the object detection audit model training phase, stochastic gradient descent (SGD) and its variants are used as optimizers to update model parameters by minimizing the loss function. Fine-tune the object detection audit model by unfreezing some layers of the pre-trained model for specific violation element categories, allowing the parameters of these layers to be updated during the training process; In image review applications, the image to be reviewed is pre-processed and input into a trained target detection review model. The target detection review model outputs the category, location, and confidence score of the detected sensitive elements. According to the preset confidence threshold, the detection results with low confidence are filtered out, and the detection results that meet the threshold requirements are visualized on the image, marking the location and category of the illegal elements.

Images