Target recognition method and system

Abstract

The present invention provides a target recognition method and system, belonging to the field of image processing technology. The method includes: parsing a user's input command, calling at least one first model based on the parsing result to perform target recognition on the corresponding type of target in the image to be recognized, and determining the recognition target with the highest confidence output by each first model; generating prompt point information based on the position information of the recognition target with the highest confidence, and matching a preset second model based on the numerical range of the highest confidence; performing target recognition on the image to be recognized based on the prompt point information and the second model to obtain expanded recognition targets; calculating the similarity between the image of any expanded recognition target with a confidence less than a first threshold and a sample image of the same type in a task type library, and retaining any expanded recognition target if the similarity is greater than a second threshold. The present invention improves the accuracy and robustness of target recognition.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a target recognition method and system. Background Art

[0002] With the continuous development of remote sensing technology, remote sensing imagery is increasingly being used in environmental monitoring, river and lake shoreline management, disaster warning, and other fields. Remote sensing intelligent recognition, as a key technology in remote sensing data processing, has become an important means of achieving efficient and accurate monitoring by extracting diverse information from massive amounts of remote sensing imagery. Traditional remote sensing recognition methods rely on large numbers of training samples, extracting information from remote sensing imagery through deep learning and pattern recognition techniques. While these methods have improved recognition efficiency and accuracy to a certain extent, they still face challenges in practical application, primarily manifesting in low recognition accuracy and weak robustness.

[0003] Currently, remote sensing imagery is complex and diverse, and factors such as image resolution, lighting variations, and weather conditions significantly impact recognition performance. Traditional methods often rely on a single model for classification and recognition, lacking comprehensive adaptability to complex and diverse scenarios, resulting in reduced recognition accuracy in certain complex scenarios. To address these issues, improving the accuracy and robustness of target recognition has become a core issue in the field of remote sensing recognition. Summary of the Invention

[0004] The present invention provides a target recognition method and system to solve at least one defect of target recognition in the prior art.

[0005] In a first aspect, the present invention provides a target recognition method, comprising: parsing a user's input instruction, calling at least one first model to perform target recognition on a target of a corresponding type in an image to be recognized based on the parsing result, and determining the recognition target with the highest confidence output by each first model; generating prompt point information based on the position information of the recognition target with the highest confidence, and matching a preset second model based on the numerical range of the highest confidence; performing target recognition on the image to be recognized based on the prompt point information and the second model to obtain an extended recognition target with the same target type as the recognition target with the highest confidence; calculating the similarity between an image of any extended recognition target with a confidence less than a first threshold and a sample image of the same type in a task type library, and retaining any extended recognition target when the similarity is greater than a second threshold; wherein the sample image in the task type library includes the image corresponding to the recognition target with the highest confidence output by the first model.

[0006] According to a target recognition method provided by the present invention, a user's input instruction is parsed, and at least one first model is called according to the parsing result, including: performing text parsing on the input instruction based on a large language model to obtain at least one type keyword; based on the type keyword, using semantic similarity, matching the type name of the first model in the recognition type library; the recognition type library includes the type names of multiple first models; and calling the first model based on the type name of the matched first model.

[0007] According to a target recognition method provided by the present invention, based on type keywords, semantic similarity is used to match the type name of the first model in the recognition type library, including: using the CLIP multimodal pre-training model to encode the type keywords and the type names in the recognition type library respectively and convert them into vector form; in the same vector space, the similarity between the vector of the type keyword and the vector of the type name of each first model in the recognition type library is calculated; based on the similarity, the type name of the first model that best matches is determined.

[0008] According to a target recognition method provided by the present invention, cue point information is generated based on the position information of the recognized target with the highest confidence, including: extracting an outer contour based on the position information of the recognized target with the highest confidence; fitting the outer contour using at least one rectangle, and setting multiple generalized cue points within each rectangle; screening all generalized cue points, retaining the generalized cue points located within the outer contour, to generate cue point information.

[0009] According to a target recognition method provided by the present invention, multiple generalization cue points are set inside each rectangle, including: setting a generalization cue point at the center point of the rectangle; and setting at least one generalization cue point on the central axis of the long side of the rectangle and on each side of the center point.

[0010] According to a target recognition method provided by the present invention, the similarity between the image of any extended recognition target with a confidence level less than a first threshold and the sample image of the same type in a task type library is calculated, including: determining a feature extraction model based on the target type of any extended recognition target; using the feature extraction model to respectively extract the first image feature of the image of any extended recognition target and the second image feature of the sample image of the same type; and calculating the similarity between the first image feature and the second image feature.

[0011] According to a target recognition method provided by the present invention, the image to be recognized is a remote sensing image.

[0012] In a second aspect, the present invention further provides a target recognition system, comprising:

[0013] A type agent is used to parse the user's input instruction, call at least one first model according to the parsing result to perform target recognition on the corresponding type of target in the image to be recognized, and determine the recognition target with the highest confidence output by each first model;

[0014] A generalized intelligent agent is used to generate cue point information based on the position information of the recognition target with the highest confidence, and match a preset second model according to the numerical range of the highest confidence;

[0015] further configured to perform target recognition on the image to be recognized based on the cue point information and the second model, so as to obtain an expanded recognition target having the same target type as the recognition target with the highest confidence;

[0016] an inductive agent, configured to calculate a similarity between an image of any expanded recognition target whose confidence is less than a first threshold and a sample image of the same type in the task type library, and retain any expanded recognition target if the similarity is greater than a second threshold;

[0017] Among them, the sample images in the task type library include images corresponding to the recognition targets with the highest confidence output by the first model.

[0018] In a third aspect, the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of any of the target recognition methods described above are implemented.

[0019] In a fourth aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the target recognition methods described above.

[0020] The target recognition method and system provided by the present invention have the following beneficial effects compared with the prior art:

[0021] (1) The target recognition method and system provided by the present invention improve the accuracy and robustness of target recognition by introducing the first model, the second model, etc. to collaboratively recognize the target.

[0022] (2) The target recognition method and system provided by the present invention call a specific recognition model according to user instructions, quickly identify the target in the image and output the result with the highest confidence.

[0023] (3) The target recognition method and system provided by the present invention analyzes the highest confidence information, calls different second models, and extracts all objects of the same type in the image, thereby expanding the recognition range.

[0024] (4) When faced with a low-confidence target, the present invention matches the low-confidence target with samples in the task type library based on the feature extraction model, eliminates results with strong interference, and ensures the accuracy and completeness of the final recognition result. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to 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.

[0026] Figure 1 It is a schematic diagram of the framework of the target recognition system provided by the present invention;

[0027] Figure 2 is a schematic diagram of a generalized prompt point provided by the present invention;

[0028] Figure 3 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0029] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0030] It should be noted that, in the description of the embodiments of the present invention, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "include a ..." do not exclude the presence of other identical elements in the process, method, article or device comprising the elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances.

[0031] The terms "first," "second," and the like in this application are used to distinguish similar objects, and are not used to describe a specific order or precedence. It should be understood that such terms are interchangeable where appropriate, so that embodiments of this application can be implemented in an order other than that illustrated or described herein. Furthermore, the terms "first," "second," and the like generally distinguish objects of a class and do not limit the number of objects; for example, the first object can be one or more. Furthermore, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates an "or" relationship between the connected objects.

[0032] The following combination Figure 1-Figure 3 The target recognition method and system provided by the embodiments of the present invention are described.

[0033] The target recognition method provided by the present invention includes but is not limited to the following steps:

[0034] Step 101: Parse the user's input instruction, call at least one first model (recognition model) based on the parsing result to perform target recognition on the corresponding type of target in the image to be recognized, and determine the recognition target with the highest confidence output by each first model.

[0035] Optionally, the image to be identified in the present invention is a remote sensing image, and the target to be identified is a ground object of different types.

[0036] The present invention can design a type of intelligent agent to implement the function of step 101.

[0037] The user's input instruction is parsed, and at least one first model is called according to the parsing result. The specific implementation method includes:

[0038] (1) Perform text parsing on the input instruction based on the large language model to obtain at least one type keyword.

[0039] For example, if the input command is: Please identify the net-enclosure farming from the picture; the type keyword parsed from the text is: net-enclosure farming.

[0040] (2) Based on the type keywords, using semantic similarity, the type name of the first model is matched in the recognition type library; the recognition type library includes multiple type names of the first model.

[0041] For example, if the recognition type library includes cage aquaculture, and the semantic similarity between net enclosure aquaculture and cage aquaculture is the highest, then the type name of the first model can be determined to be cage aquaculture.

[0042] Optionally, the method for matching the type name of the first model in the recognition type library by using semantic similarity includes:

[0043] 1) Using the CLIP multimodal pre-trained model, we encode type keywords and type names in the recognition type library and convert them into vector form.

[0044] The CLIP model is used to encode type keywords and type names from the recognition type library. CLIP (Contrastive Language–Image Pre-training) is a multimodal model that can understand both images and text. In this step, both type keywords and type names from the recognition type library are converted into vectors.

[0045] 2) In the same vector space, calculate the similarity between the vector of the type keyword and the vector of the type name of each first model in the identification type library.

[0046] Within the same vector space, calculate the similarity between the type keyword vector and the type name vector for each first model in the recognition type library. For example, this can be achieved by calculating the cosine similarity between the vectors. Cosine similarity effectively measures the directional proximity of two vectors and is less affected by their magnitude. This similarity score reflects the semantic similarity between the type keyword and the type name in the recognition type library.

[0047] 3) According to the similarity, determine the type name of the first model that best matches.

[0048] Based on the similarity scores obtained in the previous step, the type name of the first model that best matches is determined. The type name with the highest similarity score is typically chosen as the best match. This approach considers not only literal matches but also underlying semantic information. Therefore, even if the type keyword doesn't exactly match certain names in the recognition type library, as long as they are semantically similar, a correct match can be achieved.

[0049] (3) Based on the type name of the matched first model, call the first model.

[0050] It should be noted that the first model in the present invention is a target recognition model. The first model has different types and is used to recognize different targets. For example, the first model of the type named cage aquaculture is particularly suitable for target recognition in cage aquaculture.

[0051] It can be understood that the recognition target of each first model, including but not limited to the category, confidence and location information (such as bounding box information or mask information, etc.) of the recognition target, can determine the final output of each first model with the highest confidence.

[0052] Step 102: Generate cue point information based on the position information of the recognized target with the highest confidence, and match a preset second model (generalized model) based on the numerical range of the highest confidence.

[0053] Based on the cue point information and the second model, target recognition is performed on the image to be recognized to obtain an expanded recognition target having the same target type as the recognition target with the highest confidence.

[0054] The present invention can design a generalized intelligent agent to implement the function of step 102.

[0055] First, the location information of the object identified with the highest confidence is obtained from the first model. If the location information is provided in the form of a mask, the mask indicates the specific location and outline of the object in the image.

[0056] Based on this mask information, key points can be extracted as cue points. These cue points may include the center point of the object, the corner points of the bounding box, or a series of points selected along the object's outline. The purpose of cue points is to provide positioning references for subsequent models, helping them to more accurately identify similar types of objects.

[0057] The cue point information is generated according to the position information of the recognition target with the highest confidence, including:

[0058] (1) Extract the outer contour based on the position information of the target with the highest confidence;

[0059] The position information can be given in the form of a mask, which marks the area of the target in the image to be identified. From this mask, the outer contour of the target can be extracted.

[0060] (2) Use at least one rectangle to fit the outer contour and set multiple generalization prompt points inside each rectangle.

[0061] Once the outer contour is obtained, at least one rectangle may be used to fit the outer contour. One or more rectangles may be selected based on actual needs to cover the target area as accurately as possible. It should be noted that before fitting the outer contour, the outer contour may be smoothed.

[0062] Among them, multiple generalization prompt points are set inside each rectangle, which can be:

[0063] Set a generalization cue point at the center of the rectangle;

[0064] At least one generalized prompt point is set on the central axis of the long side of the rectangle and on each side of the center point.

[0065] (3) All generalized cue points are screened and the generalized cue points located within the outer contour are retained to generate cue point information.

[0066] After screening, the remaining cue points are the final cue point information, which will be used to guide further image recognition tasks.

[0067] The present invention can pre-divide the confidence level into several predefined intervals, each corresponding to a second model with different accuracy or characteristics. The most appropriate second model is selected based on the confidence level within that interval. For example, a higher confidence level may indicate that the target is relatively clear, so a model focused on analyzing detailed features may be selected; whereas a lower confidence level may require a model that can handle a wider range of uncertainty.

[0068] After selecting an appropriate second model, the cue point information generated in the first step is used to guide the model's object recognition process in the image to be recognized, further improving the recognition range.

[0069] Step 103: Calculate the similarity between any image of the expanded recognition target whose confidence is less than a first threshold and a sample image of the same type in the task type library. If the similarity is greater than a second threshold, retain the expanded recognition target; otherwise, discard the expanded recognition target. The sample images in the task type library include the image corresponding to the recognition target with the highest confidence output by the first model.

[0070] Likewise, the present invention can design a generalized intelligent agent to implement the function of step 103 .

[0071] The present invention first selects the expanded recognition targets with a confidence level less than a first threshold from the expanded recognition targets output in step 102;

[0072] Then, for any expanded recognition target whose confidence is less than the first threshold, the similarity with the sample image of the same type in the task type library is calculated;

[0073] Finally, based on the size of the similarity, the expanded recognition targets with a confidence level less than the first threshold are screened.

[0074] Calculating the similarity between an image of any expanded recognition target with a confidence level less than a first threshold and a sample image of the same type in the task type library includes:

[0075] (1) determining a feature extraction model according to the target type of any of the expanded recognition targets;

[0076] According to the target type of the expanded recognition target, a suitable feature extraction model is selected. This feature extraction model should be able to effectively extract the features of the specific target type from the image.

[0077] (2) using a feature extraction model to extract first image features of any of the expanded recognition target images and second image features of the same type of sample images;

[0078] (3) Calculate the similarity between the first image feature and the second image feature.

[0079] The present invention can use the expanded recognition target that is finally retained and the expanded recognition target with a confidence level greater than or equal to a first threshold as the final output result.

[0080] Figure 1 This is a schematic diagram of the framework of the target recognition system provided by the present invention. Figure 1 The target recognition system is introduced, and the system includes:

[0081] A type agent is used to parse the user's input instruction, call at least one first model according to the parsing result to perform target recognition on the corresponding type of target in the image to be recognized, and determine the recognition target with the highest confidence output by each first model;

[0082] A generalized intelligent agent is used to generate cue point information based on the position information of the recognition target with the highest confidence, and match a preset second model according to the numerical range of the highest confidence;

[0083] further configured to perform target recognition on the image to be recognized based on the cue point information and the second model, so as to obtain an expanded recognition target having the same target type as the recognition target with the highest confidence;

[0084] an inductive agent, configured to calculate a similarity between an image of any expanded recognition target whose confidence is less than a first threshold and a sample image of the same type in the task type library, and retain any expanded recognition target if the similarity is greater than a second threshold;

[0085] Among them, the sample images in the task type library include images corresponding to the recognition targets with the highest confidence output by the first model.

[0086] The following combination Figure 1 The target recognition system provided by the present invention is further described as follows:

[0087] Part 1 (Type Agent): The Type Agent's goal is to "accurately identify" specific ground object types from remote sensing images. It consists of three main parts: a semantic understanding module, a recognition model invocation module, and a result screening module.

[0088] The semantic understanding module is to accurately understand the user's intention from the user's instructions. The user instructions are as follows:

[0089] Please identify the pier in the picture.

[0090] "Please find all the buildings and bridges"

[0091] “Extraction storage tanks, farmland, cage aquaculture, water body targets”.

[0092] This module is primarily based on the understanding capabilities of a large language model. Conventional thinking involves directly inputting user commands into the large language model, allowing it to automatically select the model to call. To improve the understanding capabilities of existing large language models in the context of intelligent recognition, this paper proposes an optimization method combining "prompt word template guidance and large model generation," specifically as follows:

[0093] (1) Define an initialization prompt word template, such as "Please choose which model to call according to the user's instructions. Enter the instruction: {text}."

[0094] (2) “{text}” is the user’s input command, which obtains and analyzes the output results of the large language model.

[0095] (3) Correction of identification type deviation, that is, whether there is any deviation in the expression between the type keywords output by the large language model and the type names in the identification type library. For example, enclosure farming has an expression deviation from cage farming in the identification type library, but the semantics are similar. It is necessary to use the type name in the identification type library to replace it.

[0096] (4) The recognition model calling module refers to calling the corresponding recognition model (first model) for recognition based on the output of the large language model. This process requires calling a single type or multiple types of recognition models.

[0097] (5) The result screening module is to screen the results output by the large language model. For each input image, for each recognition model output result, the target with the highest confidence value is taken as the screening result of that type. The sum of all types of screening results is the output result of the first part.

[0098] Assume the input image is , the recognition model that needs to be called is , the screening results can be expressed as follows:

[0099]

[0100]

[0101] in, Representation category c The target object with the highest confidence value is identified. Represents the output of the first part, Indicates the output The confidence value of the target predicted by the model, Representation category recognition model.

[0102] Part 2 (Generalized Agent): The goal of the generalized agent is to "find everything" and identify all target objects from the image. The main workflow is as follows:

[0103] (1) According to the confidence value, the generalization model that the generalization agent needs to call is divided into multiple types to process the results of different recognition levels of the type agent.

[0104] (2) First, a generalization cue point is generated based on the target position of the recognition target with the highest confidence in the first part (e.g. Figure 2 As shown, Figure 2 (Figure 1 is a schematic diagram of the generalized cue points provided by the present invention), and then the generalized cue points are input into the generalized model (second model) to assist in generating more ground objects of the same type. The present invention proposes a new strategy for generating generalized cue points, as follows:

[0105] 1) Extract the outer contour of the target object based on its contour mask;

[0106] 2) Smooth the outer contour and fit the outer contour of each part with a rectangle. Generate three points inside each rectangle: the center of gravity and the center points of the two short sides. The short side center points need to be a certain distance away from the short sides. The specific value is determined during method debugging.

[0107] 3) Filter generalized cue points. The filtering rule is whether the cue point is included in the target object contour mask. If not, it is eliminated.

[0108] (3) Summarize the results of each type of ground object to form an expanded result. Each ground object result contains a category, a mask, and a confidence value.

[0109] Part 3 (Inductive Agent): The goal of the inductive agent is to "fill in the gaps" and filter out the low-confidence results of the generalized agent to further improve the recognition effect of this method. The workflow of the inductive agent is as follows:

[0110] (1) Collect typical sample data of different land feature types from public remote sensing sample sets, including corresponding images and labels, to form a general land feature type library.

[0111] (2) The general feature type library and the output results of the type agent are aggregated to form a task feature type library, thereby improving the diversity of different types of features and avoiding the lack of representativeness of the task type library due to insufficient feature morphology and background, which affects the overall performance of this method.

[0112] (3) Filter out the results whose generalized agent confidence is less than the first threshold, use the picture corresponding to the result as the input picture, and calculate the similarity between the feature picture corresponding to the input picture mask and the sample picture in the task feature type library by calling the feature extraction model of the corresponding type. If the similarity is higher than the threshold and the matching feature type is consistent with the feature type of the input picture, then retain the result, otherwise discard the result. The formula is described as follows:

[0113]

[0114]

[0115] in, Indicates that the generalized agent type is c 、Serial number is m The output result is, Indicates that the confidence value of the filtered thres The output of the generalized agent of 1; express The serial number is j Type is p The ground feature image is intercepted by the mask corresponding to the element of Indicates type p The feature extraction model, Represents the similarity calculation function, Indicates the serial number is k Type is p The image of the feature type library sample, represents the calculated similarity score, express The similarity calculated between the sample images in the task object type library is greater than the threshold thres 2 elements.

[0116] (4) The results in (3) (i.e., the retained results) are combined with the high-confidence results of the generalized agent (results with confidence values higher than the preset threshold) as the final output results.

[0117] To better highlight the technical advantages of this invention, experiments compared this method with the latest remote sensing feature recognition methods, TransUnet and CM-UNet, as well as a single agent. The single agent primarily employed the type agent described in this invention, without selecting the highest confidence value. Table 1 shows the comparison results between this method and the single-agent method.

[0118] Table 1 is a schematic diagram of the comparison results between the present invention and the single agent method.

[0119]

[0120] It can be seen from the table that this method achieves the best results in all indicators.

[0121] The present invention has the following beneficial effects compared to the prior art:

[0122] (1) The present invention provides a target recognition method that can be applied to remote sensing intelligent recognition of ground objects. It introduces multi-agent collaborative work and effectively overcomes the inaccurate recognition problem of traditional single agent or single model.

[0123] (2) This paper proposes a method for constructing a type of intelligent agent that can understand user instructions and accurately extract high-precision ground objects.

[0124] (3) This paper proposes a generalized intelligent agent construction method that can make up for the problem of missed detection in traditional models.

[0125] (4) The present invention proposes a method for constructing an inductive intelligent agent, which can effectively filter out misidentification results, thereby improving the accuracy of the present technical method.

[0126] (5) The type agent, generalized agent, and inductive agent in the present invention have strong flexibility and can be combined with existing technical methods as modules.

[0127] It should be noted that the target recognition system provided by the embodiment of the present invention can execute the target recognition method described in any of the above embodiments during specific operation, which will not be described in detail in this embodiment.

[0128] Figure 3 Schematic diagram of the structure of the electronic device provided by the present invention, such as Figure 3 As shown, the electronic device may include: a processor 310, a communication interface 320, a memory 330, and a communication bus 340. The processor 310, the communication interface 320, and the memory 330 communicate with each other via the communication bus 340. The processor 310 may call logic instructions in the memory 330 to execute the target recognition method.

[0129] Furthermore, the logic instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product, stored in a storage medium, includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0130] On the other hand, the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the target recognition method provided by the above embodiments.

[0131] On the other hand, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented by a processor to perform the target recognition method provided in the above embodiments when the computer program is executed.

[0132] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0133] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A target recognition method, characterized in that: include: Parsing the user's input command, calling at least one first model to perform target recognition on the corresponding type of target in the image to be recognized based on the parsing result, and determining the recognized target with the highest confidence level output by each first model; Generate prompt point information based on the position information of the recognition target with the highest confidence, and match the preset second model based on the numerical range of the highest confidence; Based on the cue point information and the second model, performing target recognition on the image to be recognized to obtain an expanded recognition target of the same target type as the recognition target with the highest confidence; Calculate the similarity between the image of any expanded recognition target whose confidence is less than the first threshold and the sample image of the same type in the task type library, and retain any expanded recognition target if the similarity is greater than the second threshold; Among them, the sample images in the task type library include images corresponding to the recognition targets with the highest confidence output by the first model.

2. The target recognition method according to claim 1, characterized in that: Parsing the user's input instruction and calling at least one first model according to the parsing result, including: Parse the input command text based on the large language model to obtain at least one type keyword; Based on the type keyword, using semantic similarity, matching the type name of the first model in the identification type library; the identification type library includes multiple type names of the first model; Based on the matched type name of the first model, the first model is called.

3. The target recognition method according to claim 2, characterized in that: Based on the type keywords, using semantic similarity, the type name of the first model is matched in the recognition type library, including: Using the CLIP multimodal pre-training model, the type keywords and type names in the recognition type library are encoded and converted into vector form; In the same vector space, calculating the similarity between the vector of the type keyword and the vector of the type name of each first model in the identification type library; According to the similarity, the type name of the first model that best matches is determined.

4. The target recognition method according to claim 1, characterized in that: Generate cue point information based on the location information of the recognized target with the highest confidence, including: Extract the outer contour based on the position information of the recognized target with the highest confidence; Fitting the outer contour using at least one rectangle, and setting multiple generalization prompt points inside each rectangle; All generalized cue points are screened and the generalized cue points located within the outer contour are retained to generate cue point information.

5. The target recognition method according to claim 4, characterized in that: Set multiple generalization cue points inside each rectangle, including: Set a generalization cue point at the center of the rectangle; At least one generalized prompt point is set on the central axis of the long side of the rectangle and on each side of the center point.

6. The target recognition method according to claim 1, characterized in that: Calculating the similarity between an image of any expanded recognition target with a confidence less than a first threshold and a sample image of the same type in the task type library, including: Determining a feature extraction model according to a target type of any of the expanded recognition targets; Using a feature extraction model, respectively extracting first image features of the image of any of the expanded recognition targets and second image features of sample images of the same type; Calculate the similarity between the first image feature and the second image feature.

7. The target recognition method according to claim 1, characterized in that: The image to be identified is a remote sensing image.

8. A target recognition system, characterized in that: include: A type agent is used to parse the user's input command, call at least one first model based on the parsing result to perform target recognition on the corresponding type of target in the image to be recognized, and determine the recognition target with the highest confidence level output by each first model; A generalized intelligent agent is used to generate cue point information based on the position information of the recognition target with the highest confidence, and match a preset second model according to the numerical range of the highest confidence; further configured to perform target recognition on the image to be recognized based on the cue point information and the second model, so as to obtain an expanded recognition target having the same target type as the recognition target with the highest confidence; an inductive agent, configured to calculate a similarity between an image of any expanded recognition target whose confidence is less than a first threshold and a sample image of the same type in the task type library, and retain any expanded recognition target if the similarity is greater than a second threshold; Among them, the sample images in the task type library include images corresponding to the recognition targets with the highest confidence output by the first model.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the target recognition method according to any one of claims 1 to 7 are implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the target recognition method according to any one of claims 1 to 7 are implemented.

Images