Big model-based intrusion target detection method

Abstract

The application relates to the technical field of large models and target detection, and discloses an intrusion target detection method based on a large model, which comprises the following steps: acquiring a to-be-detected image, inputting the to-be-detected image into a first large model to obtain a preliminary analysis result for representing the content of the to-be-detected image; determining a target detection algorithm from a plurality of preset target detection algorithms according to the preliminary analysis result; running the target detection algorithm to process the to-be-detected image to obtain a preliminary detection result; and inputting the preliminary detection result into a second large model to verify the preliminary detection result, thereby generating a final detection result. The application can significantly improve the flexibility and automation level of the detection process, enable a single system to cope with more diversified detection tasks, and thus enhance the universality of the system. Moreover, the cognitive and reasoning capabilities of the artificial intelligence large model are utilized to identify and correct possible erroneous judgments, improve the accuracy and reliability of the final output result, and reduce the probability of false positives or false negatives.

Description

Technical Field

[0001] This invention relates to the field of large model and target detection technology, and in particular to an intrusion target detection method based on large model. Background Technology

[0002] Target detection technology plays a crucial role in numerous technological fields, including security monitoring, automated production line inspection, and intelligent traffic management. This technology aims to automatically identify, locate, and classify specific targets in images or video streams using computer vision algorithms.

[0003] In related technologies, several methods for target detection have been disclosed. For example, invention patent CN 118247713 A discloses an intrusion detection method for emergency management areas. However, these existing technologies still have shortcomings in terms of how to further improve the intelligence, flexibility, and reliability of the detection process.

[0004] On the one hand, existing technical solutions typically rely on pre-defined, fixed logic when selecting which detection algorithm to execute. For example, a system might be hard-coded to specifically perform intrusion detection or vehicle detection. This approach may lack flexibility when facing complex and ever-changing real-world scenarios. When the scene in the input image or the type of target to be detected changes (e.g., from recognizing people to recognizing a specific type of vehicle), the system may not be able to automatically and dynamically schedule the most suitable dedicated algorithm for processing, which to some extent limits the system's versatility and level of automation.

[0005] On the other hand, traditional object detection algorithms may suffer from reduced reliability in outputting poor-quality images (e.g., low-light, partially occluded, or blurry images), leading to false positives or false negatives. While some solutions attempt to improve accuracy by combining multiple traditional algorithms, such combinations are typically simple summaries of results, lacking a higher-dimensional verification mechanism capable of understanding image content and intelligently reviewing and confirming the algorithm's initial results. Therefore, providing an effective and intelligent verification mechanism for the algorithm's output to enhance the confidence of the final detection results is a technical problem that needs to be solved in this field. Summary of the Invention

[0006] To address the problems existing in the prior art, the present invention provides the following technical solution, specifically including:

[0007] Intrusion target detection methods based on large models include:

[0008] S1. Obtain the image to be detected;

[0009] S2. Input the image to be detected into the first large model to obtain preliminary analysis results for characterizing the content of the image to be detected;

[0010] S3. Based on the preliminary analysis results, determine one target detection algorithm from multiple preset target detection algorithms;

[0011] S4. Run the target detection algorithm to process the image to be detected in order to obtain preliminary detection results;

[0012] S5. Input the preliminary detection results into the second large model to verify the preliminary detection results, thereby generating the final detection results.

[0013] Preferably, step S2 includes:

[0014] S21. Input the image to be detected into multiple first-large models respectively to obtain multiple candidate analysis results;

[0015] S22. Perform a voting process on the multiple candidate analysis results to select the candidate analysis result with the most votes as the preliminary analysis result.

[0016] Preferably, step S5 includes:

[0017] S51. Input the preliminary detection results into multiple second-large models respectively to obtain multiple candidate verification results;

[0018] S52. Perform a voting process on the multiple candidate verification results to select the candidate verification result with the highest number of votes as the final detection result.

[0019] Preferably, after obtaining the image to be detected in step S1, the method further includes:

[0020] Preliminary feature recognition is performed on the image to be detected to obtain preliminary features related to the target in the image to be detected; the preliminary features and the image to be detected are used as input to the first large model.

[0021] Preferably, step S2, before inputting the image to be detected into the first large model, further includes:

[0022] The image to be detected is subjected to content review to determine whether it contains preset illegal content. If so, the subsequent steps are terminated.

[0023] Preferably, the step of acquiring the image to be detected includes at least one of the following methods:

[0024] Extracting image frames from a real-time video stream;

[0025] Receive local images uploaded by users;

[0026] The image is retrieved from the network based on the image address provided by the user.

[0027] Preferably, the first large model and the second large model are the same large model.

[0028] Preferably, the plurality of preset target detection algorithms include at least one of the following: face recognition algorithm, vehicle recognition algorithm, license plate recognition algorithm, or intrusion detection algorithm.

[0029] Beneficial effects

[0030] 1. This invention, by performing a preliminary analysis of the image content using a primary model before executing a specific detection algorithm, enables the system to dynamically and intelligently select an appropriate algorithm from multiple pre-set dedicated algorithms based on the actual scene or target type present in the image. This "understand first, then execute" approach, compared to traditional schemes using fixed detection algorithms, significantly improves the flexibility and automation level of the detection process, allowing a single system to handle more diverse detection tasks, thereby enhancing the system's versatility.

[0031] 2. After the initial detection results are obtained by the dedicated target detection algorithm, this invention introduces a verification step using a second large-scale model. This verification step is equivalent to adding an intelligent review and confirmation mechanism to the output of the traditional algorithm. Utilizing the cognitive and reasoning capabilities of the large-scale artificial intelligence model, the initial detection results can be evaluated and confirmed a second time. This helps to identify and correct erroneous judgments that may arise from factors such as poor image quality in traditional algorithms, thereby improving the accuracy and reliability of the final output results and reducing the probability of false positives or false negatives. Attached Figure Description

[0032] Figure 1 This is a flowchart illustrating an intrusion target detection method based on a large model provided in a preferred embodiment of the present invention.

[0033] Figure 2 This is a flowchart illustrating an intrusion target detection method based on a large model, provided in another preferred embodiment of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0035] Figure 1 A flowchart illustrating a target detection method based on a large artificial intelligence model according to an embodiment of the present invention is shown. It should be understood that this flowchart is merely an example, and the implementation of the present invention is not limited to this. Figure 1 The order or steps shown. In some implementations, some steps may be omitted, combined, or performed in a different order.

[0036] S1. Obtain the image to be detected.

[0037] It should be understood that the method of acquiring the image to be detected can be varied. For example, in a scenario applied to real-time monitoring, the image to be detected can be an image frame obtained by performing frame extraction on a real-time video stream from a video acquisition device such as a camera. Optionally, frame extraction of the video stream can be performed using known video processing tools (such as the FFMPEG library).

[0038] For example, in an application with a graphical user interface (GUI), the image to be detected can be a local image file (such as on a PC or mobile client) uploaded by the user through the interface. In one implementation, to provide a good user experience, the interface can be built on an open-source engine such as Godot, supporting various interaction methods such as drag-and-drop upload or file selection.

[0039] For example, the image to be detected may also be an image obtained from the network based on a Uniform Resource Locator (URL) provided by the user.

[0040] After acquiring the image to be detected, it can be preprocessed to facilitate subsequent processing and network transmission. In one embodiment, preprocessing may include encoding the image data into Base64 format.

[0041] In a preferred embodiment, such as Figure 2 As shown, one or more optional pre-analysis steps can be included before submitting the image to the large model for analysis to improve the efficiency and accuracy of subsequent processing.

[0042] An optional step is to perform preliminary feature recognition on the image to be detected to obtain preliminary features related to the target in the image to be detected; the preliminary features and the image to be detected are used as input to the first large model.

[0043] The purpose of this step is to quickly extract key local features from an image using efficient traditional algorithms. For example, face detection algorithms from the OpenCV library (such as the Haar cascade classifier or a deep learning-based face detector) can be used to determine whether an image contains a face. If a face is detected, a preliminary feature label, such as "face detected," can be generated. Subsequently, this preliminary feature (e.g., the text keyword "face") will be used as input to the first main model along with the image itself. The advantage of this is that it provides the main model with explicit "hints" or "context," guiding it to focus more on image content related to this feature in subsequent analysis, thereby making a more accurate judgment.

[0044] An optional step is to perform content review on the image to be detected to determine whether it contains preset illegal content. If so, the subsequent steps are terminated.

[0045] This step aims to determine whether an image contains pre-defined prohibited content. Specifically, it can utilize image processing libraries such as OpenCV, combined with Optical Character Recognition (OCR) technology and a sensitive word database, to detect text information in the image. Alternatively, image classification models can be used for preliminary screening of prohibited content. If prohibited content is detected, the process can be immediately terminated, and an alert can be sent to the user or administrator, thereby preventing further processing and dissemination of illegal information and ensuring the secure operation of the system.

[0046] S4. Input the image to be detected into the first large model to obtain preliminary analysis results for characterizing the content of the image to be detected.

[0047] It should be noted that the first large model can be one or more large generative artificial intelligence models with image understanding capabilities. For example, it can be a multimodal large model or a large language model. Specific examples may include, but are not limited to, the GPT series models, the DeepSeek model, or the Wenxin Yiyan model.

[0048] The preliminary analysis results are often natural language descriptions or structured labels generated by the first large model, representing the overall content of the image. For example, for a picture of a construction site, the preliminary analysis result might be: "This is a picture of a construction site at dusk, with a pond in the foreground and several white cars and a crane in the background."

[0049] To further improve the accuracy and robustness of the preliminary analysis results, a preferred implementation method is to employ a multi-model voting mechanism. Specifically, this includes:

[0050] S21: Input the image to be detected into multiple first-level large models to obtain multiple candidate analysis results. Multiple large model services from different sources or versions can be called simultaneously, such as Gemini, GPT series, and Spark large models.

[0051] S22: A voting process is performed on the multiple candidate analysis results. All candidate analysis results returned by the first-largest model are collected, and a voting algorithm (e.g., based on text similarity clustering, keyword frequency statistics, etc.) is used to determine the most accepted result. For example, if two of the three models mention "car" and "construction site," this description will have a higher weight. Finally, the candidate analysis result with the highest number of votes is selected as the final preliminary analysis result. This ensemble learning approach can effectively offset the bias or random errors of a single model, significantly improving the reliability of the analysis results.

[0052] S5. Based on the preliminary analysis results, determine one target detection algorithm from multiple preset target detection algorithms.

[0053] The target detection algorithm described is traditional, efficient, and designed specifically for a particular task. In one specific embodiment, the plurality of preset target detection algorithms may include, but are not limited to, face recognition algorithms, vehicle recognition algorithms, license plate recognition algorithms, or general intrusion detection algorithms. Other target detection algorithms updated according to actual needs on-site and technological advancements are also included, and this invention does not further limit them.

[0054] Based on the preliminary analysis results obtained in step S4, decisions are made automatically. This can be achieved by parsing the preliminary analysis results (text) for keywords and matching the parsed keywords with a preset algorithm. For example, if the analysis results contain keywords such as "car" or "license plate," the system can determine to use the "license plate recognition algorithm"; if they contain keywords such as "person" or "intrusion," it can determine to use the "personnel intrusion detection algorithm." This method avoids the tediousness of manual configuration and achieves dynamic and intelligent algorithm selection.

[0055] S6. Run the target detection algorithm to process the image to be detected to obtain preliminary detection results. After determining the specific target detection algorithm, call the algorithm to process the original image to be detected. Since these are dedicated algorithms, they are fast and consume relatively few resources. The output of the algorithm is a structured "preliminary detection result". For example, a vehicle recognition algorithm may output a JSON object: {"object":"vehicle","count":10,"locations":[...]}, which accurately indicates the number and location of the detected vehicles.

[0056] S5. Input the preliminary detection result into the second large model to verify the preliminary detection result, thereby generating a final detection result.

[0057] The design of this step is to further ensure the reliability of the result. The purpose of this verification step is to use the context understanding and logical reasoning capabilities of the large model to conduct an intelligent review of the output result of the dedicated algorithm.

[0058] The role of the second large model is to "polish", "interpret", and "confirm" the preliminary detection result. It uses its powerful logical reasoning and language generation capabilities to transform structured data into a smooth, rich, and natural language description that is easy for humans to understand. For example, after receiving {"object":"vehicle","count":10}, it can combine the context of the image to generate the following "final detection result": "After verification, a total of 10 vehicles were detected in the target area. Considering the night scene, there may be a risk of unauthorized vehicle intrusion. It is recommended to pay attention." Another example is to submit the image and the preliminarily identified license plate number "Su E8B888" to the large model. The large model may combine the visual features of the license plate in the image and determine that the character "B" is more likely to be "8", thereby correcting the result.

[0059] It should be understood that, similar to the first large model, the second large model can also be one or more models with cognitive and reasoning capabilities. In one implementation, the first large model and the second large model can be two different and independently deployed models. For example, the first model can be a model with a wider cognitive range, while the second model can be a model that has been specially fine-tuned for the verification task. It is worth noting that in a simple and efficient implementation, the first large model and the second large model can be the same large model. That is, the system calls the same large model interface to complete different tasks at different stages of the process: the first call is for scene understanding and algorithm distribution, and the second call is for result verification and report generation. This can simplify the system architecture and the management of third-party services.

[0060] Similarly, in the verification step, a multi-model voting mechanism can also be preferably adopted to improve the quality and credibility of the final detection result, ensuring that the finally output analysis result is not only accurate, but also more objective and comprehensive in expression. Specifically, this step can be refined as follows:

[0061] S51. Input the preliminary detection result into multiple second large models respectively to obtain multiple candidate verification results.

[0062] S52. Conduct a voting process on the multiple candidate verification results to select the candidate verification result with the highest number of votes as the final detection result.

[0063] Optionally, an interface for manual intervention can be provided throughout the process. For example, after the final test results are generated, users can be allowed to confirm or modify the results to handle extremely complex scenarios.

[0064] Optionally, an interface may also be included for displaying and statistically analyzing the generated final detection results on a user-friendly interface. The interface can provide real-time statistical analysis of the total number of tests, average accuracy, and detailed distribution of various algorithms (such as human intrusion and license plate recognition), display the original image and a natural language analysis report (model expert) generated by the second-largest model, providing users with intuitive and comprehensive decision support.

[0065] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for intrusion target detection based on a large model, characterized by, The method comprises the following steps: S1, acquiring an image to be detected; S2, inputting the image to be detected into a first large model to acquire a preliminary analysis result for representing content of the image to be detected; S3, determining a target detection algorithm from a plurality of preset target detection algorithms according to the preliminary analysis result; S4, running the target detection algorithm to process the image to be detected to acquire a preliminary detection result; S5, inputting the preliminary detection result into a second large model to verify the preliminary detection result to generate a final detection result; Step S2 comprises: S21, inputting the image to be detected into a plurality of first large models respectively to acquire a plurality of candidate analysis results; S22, performing voting processing on the plurality of candidate analysis results to select a candidate analysis result with the highest votes as the preliminary analysis result; Step S5 comprises: S51, inputting the preliminary detection result into a plurality of second large models respectively to acquire a plurality of candidate verification results; S52, performing voting processing on the plurality of candidate verification results to select a candidate verification result with the highest votes as the final detection result; After step S1 of acquiring the image to be detected, the method further comprises: performing preliminary feature recognition on the image to be detected to acquire preliminary features related to a target in the image to be detected; and inputting the preliminary features and the image to be detected together as input of the first large model.

2. The large model-based intrusion target detection method of claim 1, wherein, Before step S2 of inputting the image to be detected into the first large model, the method further comprises: performing content review on the image to be detected to determine whether the image to be detected contains preset illegal content, and if so, terminating subsequent steps. 3.The large model-based intrusion target detection method of claim 1, wherein, The step of acquiring the image to be detected comprises at least one of the following manners: extracting an image frame from a real-time video stream; receiving a local image uploaded by a user; acquiring an image from a network according to an image address provided by a user.

4. The large model-based intrusion target detection method of claim 1, wherein, The first large model and the second large model are the same large model. 5.The large model-based intrusion target detection method of claim 1, wherein, The plurality of preset target detection algorithms comprise at least one of the following: a face recognition algorithm, a vehicle recognition algorithm, a license plate recognition algorithm, or an intrusion detection algorithm.

Images