Object recognition method and device, equipment and storage medium

Abstract

The invention discloses an object recognition method and device, equipment and a storage medium, and relates to the technical field of machine learning, and the method comprises the steps: carrying out the feature similarity analysis of a target object feature of a to-be-recognized object and a class index feature group corresponding to a target object class, and obtaining target feature similarity data; the category index feature group is a clustering center feature corresponding to the positive sample object feature; the positive sample object is a sample object which is screened based on a feature density clustering result of a plurality of original sample objects and belongs to a target object category; based on the target feature similarity data and a preset feature similarity threshold, determining object category indication information of the to-be-recognized object; the preset feature similarity threshold value is obtained by performing threshold value adaptive adjustment based on a preset identification index threshold value and the sample feature similarity data; the sample feature similar data represents the feature similarity condition between the positive sample object feature of the positive sample object and the category index feature group. By using the scheme of the invention, accurate identification of a new category can be rapidly realized.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an object recognition method, apparatus, device and storage medium. Background Technology

[0002] With the development of artificial intelligence technology, machine learning models are widely used, especially in the field of object recognition. Deep learning models are typically used to classify objects, and before a model can be deployed, it requires sample collection, data labeling, model training, and model iteration. Because the model needs to directly output classification results (usually the prediction confidence scores for each class), the samples must accurately reflect the true distribution of the data. This typically requires collecting a large amount of sample data to ensure effectiveness. However, for a specific class of objects to be identified, other classes are considered negative samples. In practice, using deep learning models to identify new classes of images not present in the sample set can easily lead to misclassification. Therefore, it is necessary to frequently collect bad cases for model iteration. Model development and each iteration require significant time investment. In scenarios with urgent business needs, object recognition algorithms need to be developed and deployed quickly, but existing methods, with their lengthy sample collection, data labeling, model training, and model iteration processes, are insufficient to meet the requirements for rapid identification of new classes. Summary of the Invention

[0003] This application provides an object recognition method, apparatus, device, and storage medium. It utilizes a category index feature group corresponding to the target object category and a preset feature similarity threshold to determine the target object category. It relies on a small sample size and does not require training of the feature extraction model. While saving development time for object recognition algorithms, it can quickly and accurately identify new categories. The technical solution of this application is as follows:

[0004] On the one hand, an object recognition method is provided, the method comprising:

[0005] The target object features and category index feature groups corresponding to the target object category are obtained. The category index features in the category index feature group are cluster center features obtained by clustering the positive sample object features corresponding to the positive sample objects. The positive sample objects are sample objects belonging to the target object category selected from multiple original sample objects based on the feature density clustering results. The feature density clustering results are feature clustering results obtained by performing density clustering on the sample object features corresponding to each of the multiple original sample objects.

[0006] The target object features are compared with the category index features in the category index feature group to obtain target feature similarity data.

[0007] Based on the comparison result between the target feature similarity data and the preset feature similarity threshold corresponding to the target object category, the object category indication information of the object to be identified is determined; the preset feature similarity threshold is obtained by adaptively adjusting the threshold based on the preset identification index threshold corresponding to the target object category and the sample feature similarity data corresponding to the positive sample object; the sample feature similarity data is used to characterize the feature similarity between the positive sample object features corresponding to the positive sample object and the category index features in the category index feature group.

[0008] On the other hand, an object recognition device is provided, the device comprising:

[0009] The feature acquisition module is used to acquire the target object features corresponding to the object to be identified and the category index feature group corresponding to the target object category. The category index feature group consists of cluster center features obtained by clustering the positive sample object features corresponding to the positive sample objects. The positive sample objects are sample objects belonging to the target object category selected from multiple original sample objects based on the feature density clustering results. The feature density clustering results are feature clustering results obtained by performing density clustering on the sample object features corresponding to each of the multiple original sample objects.

[0010] The feature similarity analysis module is used to perform feature similarity analysis between the target object features and the category index features in the category index feature group to obtain target feature similarity data.

[0011] The object category indication information determination module is used to determine the object category indication information of the object to be identified based on the comparison result between the target feature similarity data and the preset feature similarity threshold corresponding to the target object category; the preset feature similarity threshold is obtained by adaptively adjusting the threshold based on the preset identification index threshold corresponding to the target object category and the sample feature similarity data corresponding to the positive sample object; the sample feature similarity data is used to characterize the feature similarity between the positive sample object features corresponding to the positive sample object and the category index features in the category index feature group.

[0012] Furthermore, the category index feature group includes: an index feature group corresponding to each feature extraction model in the multiple feature extraction models, and the feature acquisition module includes:

[0013] An object feature extraction unit is used to input the object to be identified into the multiple feature extraction models respectively to extract object features, and obtain the object features output by each of the multiple feature extraction models.

[0014] The target object feature unit is used to take the object features output by each of the multiple feature extraction models as the target object features;

[0015] The feature similarity analysis module includes:

[0016] The feature similarity data unit is used to perform feature similarity analysis between the object features corresponding to each feature extraction model and the index feature group of the corresponding feature extraction model, so as to obtain the feature similarity data corresponding to each feature extraction model.

[0017] The target feature similarity data determination unit is used to determine the target feature similarity data based on the feature similarity data corresponding to each of the multiple feature extraction models.

[0018] Furthermore, the feature density clustering result includes: multiple sample feature sets corresponding to various feature extraction models, and the positive sample objects are selected through the following device:

[0019] The sample object feature acquisition module is used to acquire the sample object features corresponding to each of the multiple original sample objects. The sample object features corresponding to each original sample object include: the sample object features of the corresponding original sample object output by each of the multiple feature extraction models.

[0020] The density clustering module is used to perform density clustering processing on multiple features to be clustered corresponding to the target model to obtain multiple sample feature sets corresponding to the target model. The target model is any feature extraction model among the multiple feature extraction models, and the multiple features to be clustered are the sample object features corresponding to the multiple original sample objects output by the target model.

[0021] A positive and negative sample feature set determination module is used to determine the positive sample feature set and the negative sample feature set among the plurality of sample feature sets;

[0022] The sample category indication information determination module is used to determine the sample category indication information corresponding to the target model based on the positive sample feature set and the negative sample feature set. The sample category indication information is used to indicate whether each original sample object belongs to the target object category.

[0023] The category voting module is used to perform category voting on the multiple original sample objects based on the sample category indication information corresponding to each of the multiple feature extraction models, and to determine the positive sample object.

[0024] Furthermore, the density clustering module includes:

[0025] The feature distance analysis unit is used to perform feature distance analysis on every two features to be clustered among the plurality of features to be clustered, and to obtain feature distance data between every two features to be clustered;

[0026] An initial feature set unit is used to perform density clustering on the multiple features to be clustered based on the feature distance data and a preset feature neighborhood range, to obtain multiple initial feature sets;

[0027] The neighborhood range enlargement unit is used to increase the neighborhood range of the preset feature when the number of the plurality of initial feature sets is greater than the number of preset sets, so as to obtain the enlarged neighborhood range;

[0028] The repeated execution unit is used to repeatedly execute the step of performing density clustering on the multiple features to be clustered based on the feature distance data and the preset feature neighborhood range to obtain multiple initial feature sets, until the number of the multiple initial feature sets is less than or equal to the preset set number.

[0029] A sample feature set determination unit is used to use the current multiple initial feature sets as the multiple sample feature sets.

[0030] Furthermore, the category index feature group includes: an index feature group corresponding to each feature extraction model among multiple feature extraction models, and the category index feature group is constructed by the following device:

[0031] The positive sample object feature acquisition module is used to acquire the positive sample object features corresponding to the positive sample object. The positive sample object features include: sample object features corresponding to the positive sample object output by various feature extraction models.

[0032] The feature clustering module is used to perform feature clustering processing on the sample object features corresponding to the positive sample object output by each feature extraction model, so as to obtain at least one set of object features corresponding to each feature extraction model.

[0033] The cluster center feature determination module is used to use the cluster center features corresponding to each of the at least one set of object features as at least one category index feature corresponding to each feature extraction model.

[0034] The index feature group construction module is used to construct an index feature group corresponding to each feature extraction model based on at least one category index feature corresponding to each feature extraction model.

[0035] Furthermore, the device also includes:

[0036] A new sample object feature acquisition module is used to acquire the new sample object features corresponding to the new positive sample object. The new sample object features include: the sample object features corresponding to the new positive sample object output by each of the multiple feature extraction models.

[0037] A new feature clustering module is added, which is used to perform feature clustering processing on the sample object features corresponding to the newly added positive sample object output by each feature extraction model, so as to obtain at least one set of newly added object features corresponding to each feature extraction model.

[0038] A new index feature module is added, which is used to take the cluster center features corresponding to each of the at least one new object feature set as at least one new index feature corresponding to each feature extraction model.

[0039] The index feature group addition module is used to add the at least one new index feature to the index feature group corresponding to each feature extraction model.

[0040] Furthermore, the preset feature similarity threshold is adjusted by the following device:

[0041] A threshold determination module is used to determine an initial similarity threshold and the preset identification index threshold;

[0042] The sample feature similarity data module is used to perform feature similarity analysis between the positive sample object features corresponding to the positive sample object and the category index features in the category index feature group to obtain the sample feature similarity data.

[0043] The prediction category indication information determination module is used to determine the prediction category indication information of the positive sample object based on the comparison result between the sample feature similarity data and the initial similarity threshold;

[0044] The recognition performance analysis module is used to perform recognition performance analysis on the target object category based on the predicted category indication information to obtain recognition performance indicators;

[0045] The threshold adaptive adjustment module is used to adaptively adjust the initial similarity threshold based on the comparison result between the recognition performance index and the preset recognition index threshold, so as to obtain the preset feature similarity threshold.

[0046] On the other hand, an object recognition device is provided, the device including a processor and a memory, the memory storing at least one instruction or at least one program, the at least one instruction or the at least one program being loaded and executed by the processor to implement the object recognition method as described above.

[0047] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction or at least one program is stored therein, the at least one instruction or the at least one program being loaded and executed by a processor to implement the object recognition method as described above.

[0048] On the other hand, a computer program product or computer program is provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the object recognition method as described above.

[0049] The object identification method, apparatus, device, and storage medium provided in this application have the following technical advantages:

[0050] In object recognition applications, this application employs a heuristic strategy. It performs density clustering on the features of multiple original sample objects to obtain feature density clustering results. Then, based on these results, positive sample objects belonging to the target object category are selected from the original sample objects, effectively improving the efficiency and accuracy of positive sample selection. Next, based on the cluster center features obtained from clustering the positive sample object features, a category index feature group corresponding to the target object category is constructed. Feature similarity analysis is performed between the positive sample object features and the category index features in the category index feature group to obtain sample feature similarity data. Based on a preset recognition index threshold corresponding to the target object category and the sample feature similarity data, a preset feature similarity threshold for category determination is obtained, enhancing the robustness and adaptability of the object recognition system. During system application, the category index feature group corresponding to the target object category and the preset feature similarity threshold are used to determine the target object category. This requires a small sample size and eliminates the need to train the feature extraction model, saving development time for object recognition algorithms and enabling rapid and accurate identification of new categories. Attached Figure Description

[0051] To more clearly illustrate the technical solutions and advantages in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a schematic diagram of an application environment provided in an embodiment of this application;

[0053] Figure 2 This is a flowchart illustrating an object recognition method provided in an embodiment of this application;

[0054] Figure 3 This is a flowchart illustrating a positive sample object screening scheme provided in an embodiment of this application;

[0055] Figure 4 This is a flowchart illustrating a positive sample object screening process provided in an embodiment of this application;

[0056] Figure 5 This is a flowchart illustrating a category index feature group construction scheme provided in an embodiment of this application;

[0057] Figure 6 This is a flowchart illustrating a preset feature similarity threshold adjustment scheme provided in an embodiment of this application;

[0058] Figure 7 This is a flowchart illustrating another object recognition method provided in an embodiment of this application;

[0059] Figure 8 This is a schematic diagram of the framework of an object recognition system provided in an embodiment of this application;

[0060] Figure 9 This is a block diagram of an object recognition device provided in an embodiment of this application;

[0061] Figure 10 This is a schematic diagram of the structure of an object recognition device provided in an embodiment of this application. Detailed Implementation

[0062] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0063] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, products or devices.

[0064] It is understood that in the specific embodiments of this application, data such as user information are involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0065] To facilitate understanding of the embodiments of this application, several concepts will be briefly introduced below:

[0066] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0067] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, pre-trained model technology, operating / interactive systems, and mechatronics. Among these, pre-trained models, also known as large-scale models or foundational models, can be widely applied to downstream tasks across various AI fields after fine-tuning. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0068] Computer vision (CV) is a science that studies how to enable machines to "see." More specifically, it refers to machine vision, which uses cameras and computers to replace human eyes in recognizing and measuring targets, and further processes images to create images more suitable for human observation or transmission to instruments. As a scientific discipline, computer vision studies related theories and technologies, attempting to build artificial intelligence systems capable of extracting information from images or multidimensional data. Large model technology has brought significant changes to the development of computer vision technology. Pre-trained models in the vision field, such as Swin-transformer, ViT, V-MOE, and MAE, can be quickly and widely applied to downstream tasks after fine-tuning. Computer vision technology typically includes image processing, image recognition, image semantic understanding, image retrieval, OCR, video processing, video semantic understanding, video content / behavior recognition, 3D object reconstruction, 3D technology, virtual reality, augmented reality, simultaneous localization and mapping (SLAM), and common biometric recognition technologies such as facial recognition and fingerprint recognition.

[0069] Key technologies in speech technology include Automatic Speech Recognition (ASR), Text-to-Speech (TTS), and Voiceprint Recognition. Enabling computers to hear, see, speak, and feel is the future direction of human-computer interaction, with speech being one of the most promising methods. Large-scale modeling has revolutionized speech technology; pre-trained models such as WavLM and UniSpeech, which utilize the Transformer architecture, possess strong generalization and versatility, enabling them to excel in various speech processing tasks.

[0070] Natural Language Processing (NLP) is an important field within computer science and artificial intelligence. It studies the theories and methods for enabling effective communication between humans and computers using natural language. NLP deals with natural language, the language people use in daily life, and is closely related to linguistics; it also involves computer science and mathematics. Pre-trained models, a crucial technique for model training in artificial intelligence, evolved from large language models in NLP. After fine-tuning, large language models can be widely applied to downstream tasks. NLP techniques typically include text processing, semantic understanding, machine translation, question answering, and knowledge graphs.

[0071] Pre-trained models, also known as foundational models or large models, refer to deep neural networks (DNNs) with a large number of parameters. These DNNs are trained on massive amounts of unlabeled data. Leveraging the function approximation capabilities of large-parameter DNNs, Proximity-Based Transformers (PTMs) extract common features from the data. Through fine-tuning, efficient parameter fine-tuning (PEFT), and prompt-tuning techniques, they are suitable for downstream tasks. Therefore, pre-trained models can achieve ideal results in small-shot or zero-shot scenarios. PTMs can be categorized according to the data modality they process, such as language models (ELMO, BERT, GPT), visual models (Swin-transformer, ViT, V-MOE), speech models (VALL-E), and multimodal models (ViBERT, CLIP, Flamingo, Gato). Multimodal models refer to models that establish feature representations for two or more data modalities. Pre-trained models are important tools for outputting AI-generated content (AIGC) and can also serve as a general interface connecting multiple task-specific models.

[0072] Machine Learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and pre-trained learning. Pre-trained models represent the latest development in deep learning, integrating all of these techniques.

[0073] With the research and advancement of artificial intelligence (AI) technology, AI is being studied and applied in various fields, such as smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, autonomous driving, drones, digital twins, virtual humans, robots, AI-generated content (AIGC), conversational interaction, smart healthcare, smart customer service, and game AI. It is believed that with the development of technology, AI will be applied in more fields and play an increasingly important role.

[0074] In addition, the technical terms used in this application include:

[0075] Heuristic strategies: Heuristic strategies are a problem-solving approach based on experience and intuition. They simplify the problem-solving process by introducing heuristic rules or approximation algorithms, thereby quickly finding an acceptable solution within a limited time. This method can effectively reduce computational load and improve search efficiency when dealing with complex problems.

[0076] The solutions provided in this application involve technologies such as natural language processing and machine learning in artificial intelligence, which are specifically illustrated through the following embodiments:

[0077] The object recognition method provided in this application embodiment can be applied to, for example, Figure 1 In the application environment shown, this environment may include a client 10 and a server 20, which can be indirectly connected via wireless communication. A relevant object (such as a user) can send an object recognition request carrying the object to be identified to the server 20 through the client 10. The server 20 responds to the object recognition request by obtaining the target object features corresponding to the object to be identified and the category index feature group corresponding to the target object category. The category index features in the category index feature group are cluster center features obtained after clustering the positive sample object features corresponding to positive sample objects. Positive sample objects are sample objects belonging to the target object category selected from multiple original sample objects based on feature density clustering results. Feature density clustering results are feature clustering results obtained after density clustering of the sample object features corresponding to each of the multiple original sample objects. Then, feature similarity analysis is performed between the target object features and the category index features in the category index feature group to obtain target feature similarity data. Based on the comparison between the target feature similarity data and the preset feature similarity threshold corresponding to the target object category, the object category indication information of the object to be identified is determined. The preset feature similarity threshold is obtained by adaptively adjusting the preset identification index threshold corresponding to the target object category and the sample feature similarity data corresponding to the positive sample object. The sample feature similarity data is used to characterize the feature similarity between the positive sample object features corresponding to the positive sample object and the category index features in the category index feature group. The object category indication information of the object to be identified is then fed back to the client 10. It should be noted that... Figure 1 This is just one example.

[0078] The client can be a physical device such as a smartphone, computer (e.g., desktop computer, tablet computer, laptop computer), digital assistant, smart voice interaction device (e.g., smart speaker), smart wearable device, in-vehicle terminal, etc., or it can be software running on the physical device, such as a computer program. The operating system corresponding to the first client can be Android, iOS (a mobile operating system developed by Apple), Linux, Microsoft Windows, etc.

[0079] The server side can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. The server may include network communication units, processors, and memory, etc. The server side can provide backend services to the corresponding clients.

[0080] The aforementioned client 10 and server 20 can be used to build a system for object identification, which can be a distributed system.

[0081] It should be noted that the object recognition method provided in this application can be applied to both the client and the server, and is not limited to the embodiments described above.

[0082] The following describes a specific embodiment of an object recognition method provided in this application. Figure 2 This is a flowchart illustrating an object recognition method provided in an embodiment of this application. This application provides the operational steps of the method described in the embodiment or flowchart, but based on conventional or non-inventive labor, more or fewer operational steps may be included. The order of steps listed in the embodiment is merely one possible execution order among many steps and does not represent the only execution order. In actual systems or products, the method can be executed sequentially or in parallel (e.g., in a parallel processor or multi-threaded processing environment) as shown in the embodiment or drawings. Specifically, as... Figure 2 As shown, the method may include:

[0083] S201, obtain the target object features corresponding to the object to be identified and the category index feature group corresponding to the target object category. The category index features in the category index feature group are the cluster center features obtained after clustering the positive sample object features corresponding to the positive sample object. The positive sample object is a sample object belonging to the target object category selected from multiple original sample objects based on the feature density clustering result. The feature density clustering result is the feature clustering result obtained after density clustering the sample object features corresponding to each of the multiple original sample objects.

[0084] In the embodiments of this specification, the target object category can be any one of a variety of preset object categories. Specifically, the various preset object categories can be pre-set in combination with the object classification requirements in actual applications. In a specific embodiment, a category index feature group corresponding to the target object category can be pre-constructed, and the similarity between the target object features corresponding to the object to be identified and the category index feature group can be used to determine whether the object to be identified belongs to the target object category.

[0085] In the embodiments of this specification, the category index feature group corresponding to the target object category can be a set of category robust features of the target object category, and the category index feature in the category index feature group can be the cluster center feature obtained by clustering the positive sample object features corresponding to the positive sample objects of the target object category.

[0086] In the embodiments of this specification, the positive sample objects of the target object category can be sample objects belonging to the target object category selected from multiple original sample objects based on the feature density clustering results. The feature density clustering results can be feature clustering results obtained by performing density clustering on the sample object features corresponding to each of the multiple original sample objects.

[0087] In a specific embodiment, the sample object features corresponding to each original sample object can be used to characterize the object features of the corresponding original sample object. Specifically, the sample object features can be represented as a sample object feature vector.

[0088] In one specific embodiment, the sample object features corresponding to each original sample object can be the feature data obtained after inputting the corresponding original sample object into the object feature extraction model for object feature extraction. In one specific embodiment, since the feature representation capability of a single model is limited, multiple models can be selected for combination. Therefore, the object feature extraction model can include at least one feature extraction model, and the sample object features corresponding to each original sample object can include the sample object features of the corresponding original sample object output by each of the at least one feature extraction model. Correspondingly, the category index feature group corresponding to the target object category can include the index feature group corresponding to each feature extraction model in the at least one feature extraction model.

[0089] In practical applications, different pre-trained models have different model structures, training data, and training methods, which will lead to differences in feature representation. These differences are implicit, but there are some phenomena that can be observed. For example, the target object category can be divided into multiple subcategories (taking security video data as an example, the target object category can be abnormal behavior data, and multiple subcategories can include, but are not limited to: illegal intrusion, stranger loitering, fire alarm). Model A performs best in some subcategories, while Model B performs best in other subcategories. Therefore, by measuring the feature representation ability of different models in each subcategory, appropriate combinations of multiple models can be selected.

[0090] Specifically, the object feature extraction model can be a pre-trained feature extraction model, and in the embodiments of this specification, it is not necessary to train the object feature extraction model.

[0091] In an optional embodiment, the category index feature group corresponding to the target object category can be constructed according to the object modality to which the target object category belongs. Specifically, multiple sample objects of the object modality can be collected as multiple original sample objects, and the category index feature group can be constructed based on these multiple original sample objects. Optionally, the object modality here can include, but is not limited to, images, text, audio, video, etc. For example, taking an image as the object modality to which the target object category belongs, the target object category is used to classify images. Multiple sample images can be collected as multiple original sample objects, and positive sample objects are selected based on these multiple sample images to construct the category index feature group corresponding to the target object category.

[0092] In an optional embodiment, an object feature extraction model corresponding to the object modality of the original sample object can be used to extract object features from the original sample object. Specifically, the object feature extraction model corresponding to the target object modality may include, but is not limited to: a feature extraction model pre-trained only for the target object modality and a sub-model used for feature extraction of the target object modality from a feature extraction model pre-trained for multimodal matching. In an optional embodiment, when the target object modality is an image, the object feature extraction model can be at least one image feature extraction model. Schematic, at least one image feature extraction model may include: an image encoder in ResNet101 (a pre-trained deep residual network), ViT (Visual Transformer), ALBEF (Align the image and text representations BEfore Fusing, a pre-trained model that aligns image and text representations before fusion), an image encoder in CLIP (Constrastive Language-Image Pre-training, a text-image pre-trained model using contrastive learning), and BLIP (Bootstrapping Language-Image Pre-training for UnifiedVision-Language Understanding and... Image encoders in Generation (a pre-trained model for unifying visual-language understanding and generation tasks); when the target object modality is text, the object feature extraction model can be at least one text feature extraction model, illustratively speaking, at least one text feature extraction model can include: text encoders in Transformer, BERT, ALBEF, CLIP, and BLIP; when the target object modality is audio, the object feature extraction model can be at least one audio feature extraction model, illustratively speaking, at least one audio feature extraction model can include: audio extractors in VGGish, OpenL3, and MULAN (a joint embedding of music audio and natural language); when the target object modality is video, the object feature extraction model can be at least one video feature extraction model, illustratively speaking, at least one video feature extraction model can include: Video ViT (A Video Vision Transformer), VideoMAE (a self-supervised pre-trained framework for video), etc.

[0093] In a specific embodiment, such as Figure 3As shown, the positive sample objects of the above target object category are filtered in the following way:

[0094] S301, obtain the sample object features corresponding to each of the multiple original sample objects. The sample object features corresponding to each original sample object include: the sample object features of the corresponding original sample object output by each of the multiple feature extraction models.

[0095] In a specific embodiment, obtaining the sample object features corresponding to each of the multiple original sample objects may include: inputting each original sample object into multiple feature extraction models for object feature extraction, and obtaining the sample object features of the corresponding original sample object output by each of the multiple feature extraction models.

[0096] S302, perform density clustering on multiple features to be clustered corresponding to the target model to obtain multiple sample feature sets corresponding to the target model. The target model is any feature extraction model among multiple feature extraction models, and the multiple features to be clustered are the sample object features corresponding to the multiple original sample objects output by the target model.

[0097] Specifically, clustering is the process of dividing a collection of physical or abstract objects into multiple classes composed of similar objects. A cluster generated by clustering is a set of data objects that are similar to objects within the same cluster and dissimilar to objects in other clusters. Density clustering, on the other hand, is based on the concept of density; after density clustering, the density of objects assigned to the same cluster can reach a preset density threshold.

[0098] In a specific embodiment, the above-described density clustering processing of multiple features to be clustered corresponding to the target model to obtain a set of multiple sample features corresponding to the target model may include:

[0099] S3021, Perform feature distance analysis on every two features to be clustered among multiple features to obtain feature distance data between every two features to be clustered.

[0100] Specifically, the feature distance data between any two features to be clustered can characterize the degree of proximity between them. In one specific embodiment, the features to be clustered can be represented as feature vectors, and correspondingly, the vector distance between any two feature vectors to be clustered can be determined as the feature distance data. In one specific embodiment, the vector distance may include, but is not limited to, cosine distance, Euclidean distance, and Manhattan distance.

[0101] S3022, based on feature distance data and preset feature neighborhood range, performs density clustering on multiple features to be clustered to obtain multiple initial feature sets.

[0102] Specifically, the preset feature neighborhood range can be a neighborhood region with a specified feature distance threshold as its radius. In optional embodiments, the specified feature distance threshold can be preset in combination with the feature clustering accuracy requirements in actual applications.

[0103] In an optional embodiment, the DBSCAN algorithm can be used for density clustering. DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is a density-based clustering algorithm that defines a cluster as the largest set of density-connected points. It can divide regions with sufficiently high density into clusters and can discover clusters of arbitrary shapes in noisy spatial databases. Specifically, the above-mentioned density clustering processing of multiple features to be clustered based on feature distance data and a preset feature neighborhood range to obtain multiple initial feature sets may include:

[0104] S1, determine the initial clustering features from multiple clustering features.

[0105] S2, based on the feature distance data between every two features to be clustered, determine the feature density within the preset feature neighborhood range corresponding to the initial features to be clustered.

[0106] S3, when the feature density corresponding to the initial feature to be clustered reaches the preset density threshold, determine the initial feature set corresponding to the initial feature to be clustered.

[0107] Specifically, the feature density corresponding to the initial feature to be clustered can be the number of features whose feature distance data from the initial feature to be clustered is less than a specified feature distance threshold. When the number of features whose feature distance data from the initial feature to be clustered is less than the specified feature distance threshold within the neighborhood region centered on the feature space coordinates of the initial feature to be clustered and with the specified feature distance threshold as the radius is greater than the minimum sample number threshold, the feature density corresponding to the initial feature to be clustered can be considered to have reached the preset density threshold. The minimum sample number threshold can be preset based on the feature clustering accuracy requirements in actual applications.

[0108] S4, jump to the step of determining the initial clustering features from multiple clustering features, until all initial feature sets corresponding to multiple clustering features are determined.

[0109] In practical applications, we can first perform feature distance analysis on every two features among multiple features to be clustered, obtaining the feature distance data between each pair of features. Then, we can randomly select an unvisited feature from the multiple features to be clustered as the initial feature. Using this initial feature as the center and a specified feature distance threshold as the radius, we construct a preset feature neighborhood range corresponding to this initial feature. If the number of features to be clustered in this preset feature neighborhood range reaches the minimum sample size threshold, we can determine that the feature density corresponding to the initial feature to be clustered has reached the preset density threshold. At this point, the features to be clustered in the preset feature neighborhood range and the initial feature to be clustered can form a cluster, that is, an initial feature set, and the initial feature to be clustered is marked as visited. Then, we recursively process all unvisited features to be clustered within this cluster in the same way, thereby expanding the cluster. If the cluster is sufficiently expanded, that is, all features to be clustered within the cluster are marked as visited, the same algorithm can be used to process other unvisited features to be clustered. If the number of features to be clustered within the preset feature neighborhood corresponding to the initial feature to be clustered does not reach the minimum sample size threshold, it can be determined that the feature density corresponding to the initial feature to be clustered has reached the preset density threshold. In this case, the initial feature to be clustered can be temporarily marked as a noise feature. After density clustering is performed on multiple features to be clustered, the multiple features to be clustered can be divided into multiple initial feature sets, and the features to be clustered that are not in the initial feature sets are noise features.

[0110] In an optional embodiment, the method for density clustering of multiple features to be clustered is not limited to DBSCAN; OPTICS clustering methods, etc., can also be used. Among them, OPTICS (Ordering points to identify the clustering structure) is a density-based clustering algorithm that aims to cluster data in space according to density distribution and can obtain clusters of different densities.

[0111] S3023, when the number of multiple initial feature sets is greater than the number of preset sets, increase the preset feature neighborhood range to obtain the increased neighborhood range.

[0112] Specifically, the neighborhood range can be increased by increasing the specified feature distance threshold, and the neighborhood range can be decreased by decreasing the specified feature distance threshold.

[0113] S3024, Based on the increased neighborhood range, repeatedly execute the step of performing density clustering on multiple features to be clustered based on feature distance data and preset feature neighborhood range to obtain multiple initial feature sets, until the number of multiple initial feature sets is less than or equal to the preset set number.

[0114] Specifically, when the original sample objects corresponding to the features to be clustered in the initial feature set have the same category, the accuracy of density clustering can be improved by increasing the neighborhood range to make more samples in the same initial feature set and less in the number of initial feature sets.

[0115] In an optional embodiment, if there are many inconsistent categories of the original sample objects corresponding to the features to be clustered within the initial feature set, it indicates that the standard for classifying them as the same type is too low. Therefore, the preset feature neighborhood range can be reduced to obtain a reduced neighborhood range. Based on the reduced neighborhood range, the step of performing density clustering on multiple features to be clustered based on feature distance data and the preset feature neighborhood range is repeated to obtain multiple initial feature sets until the proportion of the original sample objects corresponding to the features to be clustered within each initial feature set is less than or equal to a preset proportion threshold.

[0116] S3025, take the current multiple initial feature sets as multiple sample feature sets.

[0117] As can be seen from the above embodiments, feature distance analysis is performed on every two features to be clustered to obtain feature distance data between each pair of features to be clustered. Then, based on the feature distance data and the preset feature neighborhood range, density clustering is performed on the multiple features to be clustered to obtain multiple initial feature sets. Based on the number of initial feature sets and the similarity within each initial feature set, the preset feature neighborhood range is adjusted to obtain multiple initial feature sets with accurate clustering, thereby improving the accuracy of subsequent positive sample screening.

[0118] S303, determine the positive sample feature set and the negative sample feature set from multiple sample feature sets.

[0119] Specifically, positive and negative sample feature sets can be manually labeled from multiple sample feature sets. The original sample objects corresponding to the features to be clustered in the positive sample feature set belong to the target object category, while the original sample objects corresponding to the features to be clustered in the negative sample feature set do not belong to the target object category.

[0120] In an optional embodiment, after density clustering of multiple features to be clustered is completed, the features to be clustered that are marked as noise features can be added to the negative sample feature set.

[0121] S304. Based on the positive sample feature set and the negative sample feature set, determine the sample category indication information corresponding to the target model. The sample category indication information is used to indicate whether each original sample object belongs to the target object category.

[0122] Specifically, the sample category indication information corresponding to the target model may include the category indication information of the original sample objects of multiple features to be clustered. This category indication information can be divided into positive sample indication information (i.e., the original sample objects corresponding to the features to be clustered belong to the target object category) and negative sample indication information (i.e., the original sample objects corresponding to the features to be clustered do not belong to the target object category). In a specific embodiment, the category indication information can be represented as category labels.

[0123] In an optional embodiment, when a certain sample feature set is labeled as a positive sample feature set, it can be assumed that the category indication information of the original sample objects corresponding to all features to be clustered in the sample feature set is positive sample indication information; when a certain sample feature set is labeled as a negative sample feature set, it can be assumed that the category indication information of the original sample objects corresponding to all features to be clustered in the sample feature set is negative sample indication information.

[0124] S305, based on the sample category indication information corresponding to each of the multiple feature extraction models, performs category voting on multiple original sample objects to determine the positive sample objects.

[0125] Specifically, based on the category indication information of each original sample object from various feature extraction models, if the number of positive sample indication information for a certain original sample object is greater than the number of negative sample indication information, the original sample object is determined to be a positive sample object.

[0126] In an optional embodiment, the clustering order of multiple feature extraction models can be determined. After using the feature density clustering result corresponding to the first feature extraction model to screen multiple original sample objects for positive sample objects and obtain multiple initial positive sample objects, based on the clustering order, the feature density clustering and positive sample object screening process is repeated using other feature extraction models based on the positive sample screening results of the previous round. Since the feature expression ability of a certain feature extraction model is limited, repeating this process on the positive sample screening results of the previous feature extraction model can continuously delete noisy negative samples and finally obtain a clean set of positive sample objects.

[0127] See Figure 4 , Figure 4 This is a flowchart illustrating a positive sample object screening process provided in an embodiment of this application. Specifically, positive sample objects are screened out from multiple original sample objects through sample feature extraction, feature density clustering, and positive sample set selection.

[0128] As can be seen from the above embodiments, based on the feature density clustering results corresponding to each of the multiple feature extraction models, the sample category indication information corresponding to each of the multiple feature extraction models is determined. Then, the category voting is performed on multiple original sample objects. The positive sample objects belonging to the target object category are selected from multiple original sample objects. This can improve the problem of the limited feature expression ability of a single model. Removing negative samples multiple times on different models can effectively improve the accuracy of positive sample selection.

[0129] In a specific embodiment, the above-mentioned category index feature group may include: the index feature group corresponding to each feature extraction model in multiple feature extraction models, such as... Figure 5 As shown, the above category index feature groups are constructed in the following way:

[0130] S501, Obtain the positive sample object features corresponding to the positive sample object. The positive sample object features include the sample object features corresponding to the positive sample object output by each of the multiple feature extraction models.

[0131] Specifically, based on step S301, the positive sample object features corresponding to the positive sample object can be obtained.

[0132] S502, perform feature clustering processing on the sample object features corresponding to the positive sample objects output by each feature extraction model to obtain at least one set of object features corresponding to each feature extraction model.

[0133] In an optional embodiment, a partitioning clustering algorithm can be used to perform feature clustering processing on the sample object features corresponding to the positive sample objects output by each feature extraction model, thereby obtaining at least one set of object features corresponding to each feature extraction model. Illustratively, the partitioning clustering algorithm here may include, but is not limited to, K-means clustering, Agglomerative Clustering, etc., and this application does not impose any particular limitation on it.

[0134] S503, take the cluster center feature corresponding to each of the at least one set of object features as at least one category index feature corresponding to each feature extraction model.

[0135] Specifically, after obtaining at least one set of object features corresponding to each feature extraction model, the cluster center features corresponding to each set of object features can be determined based on the sample object features within each set of object features.

[0136] In an optional embodiment, the sample object features within each object feature set can be averaged to obtain the cluster center features corresponding to each object feature set.

[0137] S504, construct an index feature group corresponding to each feature extraction model based on at least one category index feature corresponding to each feature extraction model.

[0138] Specifically, the index feature group corresponding to each feature extraction model can contain at least one category index feature of the corresponding feature extraction model.

[0139] As can be seen from the above embodiments, feature clustering is performed on the sample object features corresponding to the positive sample objects output by various feature extraction models to obtain at least one cluster center feature corresponding to each of the various feature extraction models, so as to construct the category index feature group corresponding to the target object category, and robust features for category feature matching of the target object category can be extracted.

[0140] In one specific embodiment, the above method may further include:

[0141] S505, obtain the features of the newly added positive sample objects corresponding to the newly added positive sample objects. The features of the newly added sample objects include the sample object features output by each of the multiple feature extraction models corresponding to the newly added positive sample objects.

[0142] In practical applications, datasets may change over time. To enable the system to adapt to such changes, incremental learning methods can be used to collect new samples periodically or as needed. These new samples may come from new application scenarios or from new data in existing application scenarios. The purpose of collecting new samples is to enable the system to identify newly emerging categories or to better identify existing categories.

[0143] S506, Perform feature clustering processing on the sample object features corresponding to the newly added positive sample objects output by each feature extraction model to obtain at least one set of newly added object features corresponding to each feature extraction model.

[0144] Specifically, the details of step S506 are similar to those of step S502, and will not be repeated here.

[0145] S507, take the cluster center feature corresponding to each of the at least one newly added object feature set as at least one newly added index feature for each feature extraction model.

[0146] Specifically, the details of step S507 are similar to those of step S503, and will not be repeated here.

[0147] S508, add at least one new index feature to the index feature group corresponding to each feature extraction model.

[0148] In an optional embodiment, the features of the newly added positive sample objects corresponding to the newly added positive sample objects in S505 and the features of the positive sample objects in step S501 can be merged into a feature set corresponding to each feature extraction model. Based on the merged feature set corresponding to each feature extraction model, the processing of steps S502-S504 can be repeated to update the index feature group corresponding to each feature extraction model, thereby updating the recognition capability of the object recognition system.

[0149] As can be seen from the above embodiments, by incremental learning, adding features of new samples to the original dataset can update the recognition capability of the object recognition system, enabling the system to recognize newly emerging categories or better recognize existing categories.

[0150] In the embodiments of this specification, the object to be identified can be any object that needs to be categorized, and the target object features can characterize the object features of the object to be identified. Specifically, the target object features can be the feature data obtained after inputting the object to be identified into an object feature extraction model for object feature extraction. In a specific embodiment, the target object features can be represented as a target object feature vector.

[0151] S202, perform feature similarity analysis between the target object features and the category index features in the category index feature group to obtain target feature similarity data.

[0152] In the embodiments of this specification, target feature similarity data can be used to characterize the feature similarity between target object features and category index features in the category index feature group. In a specific embodiment, target object features can be represented as target object feature vectors, and category index features in the category index feature group can be represented as category index feature vectors. Accordingly, the vector distance between the target object feature vector and the category index feature vector can be used to measure the target feature similarity data. Optionally, the vector distance here may include, but is not limited to, cosine distance, Euclidean distance, and Manhattan distance.

[0153] S203, based on the comparison results between the target feature similarity data and the preset feature similarity threshold corresponding to the target object category, determine the object category indication information of the object to be identified; the preset feature similarity threshold is obtained by adaptively adjusting the threshold based on the preset identification index threshold corresponding to the target object category and the sample feature similarity data corresponding to the positive sample object; the sample feature similarity data is used to characterize the feature similarity between the positive sample object features corresponding to the positive sample object and the category index features in the category index feature group.

[0154] In the embodiments of this specification, object category indication information can be used to indicate whether the object to be identified belongs to the target object category. In a specific embodiment, the object category indication information can be represented as an object category label.

[0155] In a specific embodiment, the object category indication information can be divided into: positive object category indication information and negative object category indication information. Positive object category indication information indicates that the object to be identified belongs to the target object category, while negative object category indication information indicates that the object to be identified does not belong to the target object category. Specifically, if the value of the target feature similarity data is greater than a preset feature similarity threshold, the object category indication information is considered positive; if the value of the target feature similarity data is not greater than the preset feature similarity threshold, the object category indication information is considered negative.

[0156] In the embodiments of this specification, the preset recognition index threshold can be the index threshold of the recognition performance index. The recognition performance index is used to evaluate the recognition performance of the object recognition system for the target object category. Specifically, the recognition performance index may include, but is not limited to, the false positive rate index and the false negative rate index. Correspondingly, the preset recognition index threshold may include, but is not limited to, the false positive rate threshold and the false negative rate threshold.

[0157] In a specific embodiment, the positive sample object features can be represented as positive sample object feature vectors, and the category index features in the category index feature group can be represented as category index feature vectors. Correspondingly, the vector distance between the positive sample object features and the category index feature vectors can be used to measure the similarity of sample features. Optionally, the vector distance here can include, but is not limited to, cosine distance, Euclidean distance, and Manhattan distance.

[0158] As can be seen from the above embodiments, in object recognition applications, a heuristic strategy is adopted. By performing density clustering on the sample object features corresponding to multiple original sample objects, feature density clustering results are obtained. Then, positive sample objects belonging to the target object category are selected from multiple original sample objects based on the feature density clustering results. This can effectively improve the efficiency and accuracy of positive sample selection. Then, based on the cluster center features obtained after clustering the positive sample object features corresponding to the positive sample objects, a category index feature group corresponding to the target object category is constructed. Feature similarity analysis is performed on the positive sample object features corresponding to the positive sample objects and the category index features in the category index feature group to obtain sample feature similarity data. Based on the preset recognition index threshold corresponding to the target object category and the sample feature similarity data, the threshold is adaptively adjusted to obtain the preset feature similarity threshold used for category determination. This can improve the robustness and adaptability of the object recognition system. In the process of system application, the category index feature group corresponding to the target object category and the preset feature similarity threshold are used to determine the target object category of the object to be identified. The number of samples required is small, and there is no need to train the feature extraction model. While saving the development time of the object recognition algorithm, the accurate identification of new categories can be achieved quickly.

[0159] In a specific embodiment, such as Figure 6 As shown, the preset feature similarity threshold is adjusted in the following way:

[0160] S601, determine the initial similarity threshold and the preset identification index threshold.

[0161] Specifically, the initial similarity threshold corresponding to the target object category can be used to determine whether an object belongs to the target object category, and the preset recognition index threshold can be used to evaluate the recognition performance of the object recognition system for the target object category. The initial similarity threshold and the preset recognition index threshold can be preset in combination with the object recognition accuracy requirements in actual applications.

[0162] S602, perform feature similarity analysis between the positive sample object features corresponding to the positive sample object and the category index features in the category index feature group to obtain sample feature similarity data.

[0163] In a specific embodiment, the positive sample object features may include: sample object features corresponding to the positive sample object output by each of the multiple feature extraction models; the category index feature group may include: at least one category index feature corresponding to each feature extraction model among the multiple feature extraction models; the above-mentioned performing feature similarity analysis between the positive sample object features corresponding to the positive sample object and the category index features in the category index feature group to obtain sample feature similarity data may include:

[0164] S6021, Perform feature similarity analysis between at least one category index feature corresponding to the model to be analyzed and the sample object feature corresponding to the positive sample object output by the model to be analyzed, to obtain the initial similarity data of each category index feature corresponding to the model to be analyzed; the model to be analyzed is any one of the above feature extraction models.

[0165] Specifically, the initial similarity data for each category index feature can be used to characterize the feature similarity between the corresponding category index feature and the sample object feature. In a specific embodiment, the initial similarity data can be represented as the vector distance between feature vectors.

[0166] S6022, take the maximum value of the initial similar data of at least one category index feature corresponding to the model to be analyzed as the maximum similar data corresponding to the model to be analyzed.

[0167] S6023 uses the maximum value among the maximum similar data corresponding to each of the multiple feature extraction models as the sample feature similarity data.

[0168] S603, based on the comparison results between sample feature similarity data and the initial similarity threshold, determine the predicted category indication information of positive sample objects.

[0169] In a specific embodiment, the predicted category indication information can be divided into: predicted positive class indication information and predicted negative class indication information. The predicted positive class indication information is used to indicate that a positive sample object belongs to the target object category, and the predicted negative class indication information is used to indicate that a positive sample object does not belong to the target object category. Specifically, if the value of the sample feature similarity data is greater than the initial similarity threshold, the predicted category indication information is considered to be predicted positive class indication information; if the value of the sample feature similarity data is not greater than the initial similarity threshold, the predicted category indication information is considered to be predicted negative class indication information.

[0170] S604, based on the predicted category indication information, performs recognition performance analysis on the target object category to obtain recognition performance indicators.

[0171] In a specific embodiment, the identification performance metrics may include: a false positive rate metric and a false negative rate metric. The false positive rate metric can be obtained by performing category false positive analysis based on the predicted category indication information of positive sample objects, and the false negative rate metric can be obtained by performing category false negative analysis based on the predicted category indication information of positive sample objects.

[0172] S605, based on the comparison results between the recognition performance index and the preset recognition index threshold, adaptively adjusts the initial similarity threshold to obtain the preset feature similarity threshold.

[0173] In optional embodiments, the recognition performance indicators may include: a false positive rate indicator and a false negative rate indicator. Correspondingly, the preset recognition indicator thresholds may include: a false positive rate threshold and a false negative rate threshold. The above-mentioned adaptive adjustment of the initial similarity threshold based on the comparison results between the recognition performance indicators and the preset recognition indicator thresholds to obtain the preset feature similarity threshold may include: increasing the initial similarity threshold when the false positive rate indicator is greater than the false positive rate threshold to obtain the preset feature similarity threshold; or, decreasing the initial similarity threshold when the false negative rate indicator is greater than the false negative rate threshold to obtain the preset feature similarity threshold.

[0174] In an optional embodiment, positive sample objects can be randomly divided into training objects and test objects. A category index feature group is constructed based on the sample object features of the training objects. Based on the sample feature similarity data of the test objects and the preset recognition index threshold, the initial similarity threshold is adaptively adjusted to obtain the preset feature similarity threshold.

[0175] In an optional embodiment, after the category index feature group is updated, the recognition performance of the updated category index feature group can be verified. If the corresponding recognition performance index cannot meet the preset recognition index threshold required for actual application, some parameters can be adaptively adjusted, such as the clustering parameters in step S302 (specifying the feature distance threshold and the minimum sample number threshold, the preset feature similarity threshold in step S605, etc.).

[0176] As can be seen from the above embodiments, based on the comparison results between the sample feature similarity data of positive sample objects and the initial similarity threshold, the predicted category indication information of positive sample objects is determined, thereby obtaining the recognition performance index of the target object category. Based on the comparison results between the recognition performance index and the preset recognition index threshold, the initial similarity threshold is adaptively adjusted to obtain the preset feature similarity threshold. Compared with manually specifying a fixed similarity threshold, this adaptive method can be adjusted according to the actual application scenario and the characteristics of the dataset, effectively improving the robustness and adaptability of the object recognition system.

[0177] In a specific embodiment, such as Figure 7 As shown, the above category index feature group may include: the index feature group corresponding to each feature extraction model in multiple feature extraction models, and the above acquisition of the target object features corresponding to the object to be identified may include:

[0178] S2011, the object to be identified is input into multiple feature extraction models to extract object features, and the object features output by each of the multiple feature extraction models are obtained.

[0179] Specifically, the multiple feature extraction models input to the object to be identified should be consistent with the multiple feature extraction models corresponding to the category index feature groups.

[0180] S2012 uses the object features output by various feature extraction models as the target object features.

[0181] Accordingly, the above-mentioned feature similarity analysis between the target object features and the category index features in the category index feature group, to obtain target feature similarity data, may include:

[0182] S2021, Perform feature similarity analysis between the object features corresponding to each feature extraction model and the index feature group of the corresponding feature extraction model to obtain feature similarity data corresponding to each feature extraction model.

[0183] In a specific embodiment, the above-mentioned method of performing feature similarity analysis between the object features corresponding to each feature extraction model and the index feature group of the corresponding feature extraction model to obtain feature similarity data corresponding to each feature extraction model may include: performing feature similarity analysis between at least one category index feature in the index feature group of each feature extraction model and the object features output by the corresponding feature extraction model to obtain predicted similarity data for each category index feature of each feature extraction model; and determining the feature similarity data corresponding to each feature extraction model based on the predicted similarity data of at least one category index feature corresponding to each feature extraction model.

[0184] Optionally, the maximum value in the predicted similarity data of at least one category index feature corresponding to each feature extraction model can be used as the feature similarity data corresponding to each feature extraction model; or the average value in the predicted similarity data of at least one category index feature corresponding to each feature extraction model can be used as the feature similarity data corresponding to each feature extraction model.

[0185] Specifically, the predicted similarity data for each category index feature can be used to characterize the feature similarity between the corresponding category index feature and the object feature. In a specific embodiment, the predicted similarity data can be represented as the vector distance between feature vectors.

[0186] S2022, based on the feature similarity data corresponding to each of the multiple feature extraction models, determine the target feature similarity data.

[0187] In one specific embodiment, the maximum value among the feature similarity data corresponding to each of the multiple feature extraction models is taken as the target feature similarity data.

[0188] Accordingly, the object category indication information for determining the object to be identified, based on the comparison results between the target feature similarity data and the preset feature similarity threshold corresponding to the target object category, may include:

[0189] S2031, if the comparison result indicates that the value of the target feature similarity data is greater than the preset feature similarity threshold, the object category indication information is considered to be the object positive class indication information;

[0190] Alternatively, in S2032, if the comparison result indicates that the value of the target feature similarity data is not greater than the preset feature similarity threshold, the object category indication information is considered to be the object negative class indication information.

[0191] As can be seen from the above embodiments, by using multiple feature extraction models to extract object features of the object to be identified, and performing feature similarity analysis between the object features corresponding to each feature extraction model and the index feature group of the corresponding feature extraction model, feature similarity data corresponding to each feature extraction model is obtained. Then, based on the feature similarity data corresponding to each of the multiple feature extraction models, target feature similarity data is determined. Finally, based on the comparison results between the target feature similarity data and the preset feature similarity threshold corresponding to the target object category, the object category indication information of the object to be identified is determined. This can improve the problem of the limited feature expression ability of a single model. By using the features extracted by multiple models to perform object recognition together, the accuracy of object recognition results is effectively improved.

[0192] illustrative, for reference Figure 8 , Figure 8 This is a schematic diagram of the framework of an object recognition system provided in an embodiment of this application. Specifically, the object recognition system may include: a feature extractor selection module, a raw sample collection module, a semi-automated dataset module, a robust feature extraction module, a feature similarity threshold adjustment module, and an object recognition module, wherein: the feature extractor selection module can be used to select at least one feature extraction model suitable for the target object category; the raw sample collection module can be used to collect multiple raw sample objects related to the target object category; the semi-automated dataset module can be used to implement the positive sample object screening process in steps S301-S305; the robust feature extraction module can be used to implement the category index feature group construction process in steps S501-S508; the feature similarity threshold adjustment module can be used to implement the feature similarity threshold adaptive adjustment process in steps S601-S605; and the object recognition module can be used to implement the object category recognition process in steps S2011, S2012, S2021, S2022, S2031, and S2032.

[0193] To illustrate, taking the object to be identified as an image as an example, the object recognition system can be an image recognition system. In social media content moderation, users upload and share a large number of pictures and videos on social media platforms every day. Among them, some content may contain illegal or harmful information, such as bullying, fake news, etc. In order to maintain the safety and order of the community, the platform needs to review and filter this content. However, due to the massive amount and diversity of content, new malicious types of images will continue to appear. Using the image recognition system (image recognition method) proposed in this application, these new malicious types of images can be quickly identified, realizing automated and real-time content moderation, which greatly improves the efficiency and accuracy of moderation. At the same time, it can also nip harmful information in the bud in time, preventing it from fermenting and spreading. Taking video as an example of the object to be identified, the object recognition system can be a video recognition system. In security monitoring, security cameras generate a large amount of security video data every day, which may include some abnormal behaviors, such as intrusion, fighting, fire, etc. In order to ensure public safety, it is necessary to monitor and warn of these behaviors in real time. The video recognition system (video recognition method) proposed in this application can quickly learn and identify the characteristics of these events, realize automated and real-time event monitoring and early warning, and help security personnel respond and handle quickly.

[0194] As can be seen from the technical solutions provided in the embodiments of this application above, by adopting a heuristic strategy, density clustering is performed on the sample object features corresponding to multiple original sample objects to obtain feature density clustering results. Then, positive sample objects belonging to the target object category are selected from multiple original sample objects based on the feature density clustering results, which can effectively improve the efficiency and accuracy of positive sample selection. Then, based on the cluster center features obtained after clustering the positive sample object features corresponding to the positive sample objects, a category index feature group corresponding to the target object category is constructed. Feature similarity analysis is performed on the positive sample object features corresponding to the positive sample objects and the category index features in the category index feature group to obtain sample feature similarity data. Based on the preset recognition index threshold corresponding to the target object category and the sample feature similarity data, the threshold is adaptively adjusted to obtain the preset feature similarity threshold used for category determination. This can improve the robustness and adaptability of the object recognition system. In the process of system application, the category index feature group corresponding to the target object category and the preset feature similarity threshold are used to determine the target object category of the object to be identified. The number of samples required is small, and there is no need to train the feature extraction model. While saving the development time of the object recognition algorithm, the accurate identification of new categories can be achieved quickly.

[0195] This application also provides an object recognition device, such as... Figure 9 As shown, the object recognition device may include:

[0196] The feature acquisition module 910 is used to acquire the target object features corresponding to the object to be identified and the category index feature group corresponding to the target object category. The category index features in the category index feature group are the cluster center features obtained after clustering the positive sample object features corresponding to the positive sample objects. The positive sample objects are sample objects belonging to the target object category selected from multiple original sample objects based on the feature density clustering results. The feature density clustering results are the feature clustering results obtained after performing density clustering on the sample object features corresponding to each of the multiple original sample objects.

[0197] The feature similarity analysis module 920 is used to perform feature similarity analysis between the target object features and the category index features in the category index feature group to obtain target feature similarity data.

[0198] The object category indication information determination module 930 is used to determine the object category indication information of the object to be identified based on the comparison result between the target feature similarity data and the preset feature similarity threshold corresponding to the target object category; the preset feature similarity threshold is obtained by adaptively adjusting the threshold based on the preset identification index threshold corresponding to the target object category and the sample feature similarity data corresponding to the positive sample object; the sample feature similarity data is used to characterize the feature similarity between the positive sample object features corresponding to the positive sample object and the category index features in the category index feature group.

[0199] In a specific embodiment, the aforementioned category index feature group may include: an index feature group corresponding to each feature extraction model in multiple feature extraction models, and the aforementioned feature acquisition module 910 may include:

[0200] The object feature extraction unit is used to input the object to be identified into multiple feature extraction models to extract object features and obtain the object features output by each of the multiple feature extraction models.

[0201] The target object feature unit is used to take the object features output by various feature extraction models as the target object features.

[0202] The aforementioned feature similarity analysis module 920 may include:

[0203] The feature similarity data unit is used to perform feature similarity analysis between the object features corresponding to each feature extraction model and the index feature group of the corresponding feature extraction model, so as to obtain the feature similarity data corresponding to each feature extraction model.

[0204] The target feature similarity data determination unit is used to determine the target feature similarity data based on the feature similarity data corresponding to various feature extraction models.

[0205] In a specific embodiment, the above feature density clustering result may include: multiple sample feature sets corresponding to various feature extraction models, wherein the above positive sample objects are selected by the following device:

[0206] The sample object feature acquisition module is used to acquire the sample object features corresponding to multiple original sample objects. The sample object features corresponding to each original sample object include: the sample object features of the corresponding original sample object output by multiple feature extraction models.

[0207] The density clustering module is used to perform density clustering on multiple features to be clustered corresponding to the target model, so as to obtain multiple sample feature sets corresponding to the target model. The target model is any feature extraction model among multiple feature extraction models, and the multiple features to be clustered are the sample object features corresponding to the multiple original sample objects output by the target model.

[0208] The positive and negative sample feature set determination module is used to determine the positive and negative sample feature sets from multiple sample feature sets.

[0209] The sample category indication information determination module is used to determine the sample category indication information corresponding to the target model based on the positive sample feature set and the negative sample feature set. The sample category indication information is used to indicate whether each original sample object belongs to the target object category.

[0210] The category voting module is used to vote on the categories of multiple original sample objects based on the sample category indication information corresponding to each of the various feature extraction models, and to determine the positive sample objects.

[0211] In one specific embodiment, the density clustering module described above may include:

[0212] The feature distance analysis unit is used to perform feature distance analysis on every two features to be clustered among multiple features to be clustered, and to obtain the feature distance data between every two features to be clustered.

[0213] The initial feature set unit is used to perform density clustering on multiple features to be clustered based on feature distance data and a preset feature neighborhood range, to obtain multiple initial feature sets.

[0214] The neighborhood range enlargement unit is used to increase the neighborhood range of a preset feature when the number of multiple initial feature sets is greater than the number of preset sets, so as to obtain the enlarged neighborhood range.

[0215] The repeated execution unit is used to repeatedly execute the steps of performing density clustering on multiple features to be clustered based on feature distance data and preset feature neighborhood range, to obtain multiple initial feature sets, until the number of multiple initial feature sets is less than or equal to the preset number of sets.

[0216] The sample feature set determination unit is used to take the current multiple initial feature sets as multiple sample feature sets.

[0217] In one specific embodiment, the above-mentioned category index feature group may include: an index feature group corresponding to each feature extraction model in multiple feature extraction models, and the above-mentioned category index feature group is constructed by the following means:

[0218] The positive sample object feature acquisition module is used to acquire the positive sample object features corresponding to the positive sample object. The positive sample object features include: the sample object features corresponding to the positive sample object output by each of the multiple feature extraction models.

[0219] The feature clustering module is used to perform feature clustering processing on the sample object features corresponding to the positive sample objects output by each feature extraction model, so as to obtain at least one set of object features corresponding to each feature extraction model.

[0220] The cluster center feature determination module is used to take the cluster center features corresponding to each of the at least one set of object features as at least one category index feature corresponding to each feature extraction model;

[0221] The index feature group construction module is used to construct the index feature group corresponding to each feature extraction model based on at least one category index feature corresponding to each feature extraction model.

[0222] In one specific embodiment, the above-described apparatus may further include:

[0223] A new sample object feature acquisition module is added to acquire the new sample object features corresponding to the new positive sample object. The new sample object features include: the sample object features corresponding to the new positive sample object output by various feature extraction models.

[0224] A new feature clustering module is added to perform feature clustering on the sample object features corresponding to the newly added positive sample objects output by each feature extraction model, so as to obtain at least one set of new object features corresponding to each feature extraction model.

[0225] A new index feature module is added, which is used to take the cluster center feature corresponding to each of the at least one new object feature set as at least one new index feature for each feature extraction model.

[0226] The index feature group addition module is used to add at least one new index feature to the index feature group corresponding to each feature extraction model.

[0227] In one specific embodiment, the aforementioned preset feature similarity threshold is adjusted by the following device:

[0228] The threshold determination module is used to determine the initial similarity threshold and the preset identification index threshold;

[0229] The sample feature similarity data module is used to perform feature similarity analysis between the positive sample object features corresponding to the positive sample object and the category index features in the category index feature group to obtain sample feature similarity data.

[0230] The predicted category indication information determination module is used to determine the predicted category indication information of positive sample objects based on the comparison results between sample feature similarity data and the initial similarity threshold;

[0231] The recognition performance analysis module is used to perform recognition performance analysis on the target object category based on the predicted category indication information, and obtain recognition performance indicators.

[0232] The threshold adaptive adjustment module is used to adaptively adjust the initial similarity threshold based on the comparison results between the recognition performance index and the preset recognition index threshold, so as to obtain the preset feature similarity threshold.

[0233] It should be noted that the apparatus and method embodiments described above are based on the same inventive concept.

[0234] This application provides an object recognition device, which includes a processor and a memory. The memory stores at least one instruction or at least one program segment, which is loaded and executed by the processor to implement the object recognition method provided in the above method embodiments.

[0235] Furthermore, Figure 10 A schematic diagram of the hardware structure of an object recognition device for implementing the object recognition method provided in the embodiments of this application is shown. The object recognition device may participate in or include the object recognition apparatus provided in the embodiments of this application. Figure 10 As shown, the object recognition device 100 may include one or more ( Figure 10 The processor 1002 (which may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA), 1004 for storing data, and 1006 for communication functions may also be included. In addition, it may include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports in the I / O interface), a network interface, a power supply, and / or a camera. Those skilled in the art will understand that... Figure 10 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, the object recognition device 100 may also include a... Figure 10The more or fewer components shown, or having the same Figure 10 The different configurations shown.

[0236] It should be noted that the aforementioned one or more processors 1002 and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be wholly or partially integrated into any other element within the object recognition device 100 (or mobile device). As involved in the embodiments of this application, the data processing circuits serve as a processor control mechanism (e.g., selection of a variable resistor termination path connected to an interface).

[0237] The memory 1004 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the object recognition method described in the embodiments of this application. The processor 1002 executes various functional applications and data processing by running the software programs and modules stored in the memory 1004, thereby realizing the object recognition method described above. The memory 1004 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 1004 may further include memory remotely located relative to the processor 1002, and these remote memories can be connected to the object recognition device 100 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0238] The transmission device 1006 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the object identification device 100. In one example, the transmission device 1006 includes a network interface controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In one embodiment, the transmission device 1006 may be a radio frequency (RF) module for wireless communication with the Internet.

[0239] The display may be, for example, a touchscreen liquid crystal display (LCD) that allows a user to interact with the user interface of the object recognition device 100 (or a mobile device).

[0240] Embodiments of this application also provide a computer-readable storage medium, which can be disposed in an object recognition device to store at least one instruction or at least one program segment for implementing at least one instruction or at least one program segment, which is loaded and executed by the processor to implement the object recognition method provided in the above method embodiments.

[0241] Optionally, in this embodiment, the storage medium may be located in at least one of the multiple network servers in a computer network. Optionally, in this embodiment, the storage medium may include, but is not limited to, various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0242] Embodiments of this application also provide a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the object recognition method as provided in the method embodiments.

[0243] In this application embodiment, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0244] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims can be performed in a different order than that shown in the embodiments and still achieve the desired results. Additionally, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are also possible or may be advantageous.

[0245] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and apparatus embodiments are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0246] In this application embodiment, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0247] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0248] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An object recognition method, characterized in that, The method includes: The target object features and category index feature groups corresponding to the target object category are obtained. The category index features in the category index feature group are cluster center features obtained by clustering the positive sample object features corresponding to the positive sample objects. The positive sample objects are sample objects belonging to the target object category selected from multiple original sample objects based on the feature density clustering results. The feature density clustering results are feature clustering results obtained by performing density clustering on the sample object features corresponding to each of the multiple original sample objects. The target object features are compared with the category index features in the category index feature group to obtain target feature similarity data. Based on the comparison result between the target feature similarity data and the preset feature similarity threshold corresponding to the target object category, the object category indication information of the object to be identified is determined; the preset feature similarity threshold is obtained by adaptively adjusting the threshold based on the preset identification index threshold corresponding to the target object category and the sample feature similarity data corresponding to the positive sample object; the sample feature similarity data is used to characterize the feature similarity between the positive sample object features corresponding to the positive sample object and the category index features in the category index feature group.

2. The method according to claim 1, characterized in that, The category index feature group includes: the index feature group corresponding to each feature extraction model in multiple feature extraction models, and the step of obtaining the target object features corresponding to the object to be identified includes: The object to be identified is input into the various feature extraction models respectively to extract object features, and the object features output by each of the various feature extraction models are obtained. The object features output by each of the multiple feature extraction models are used as the target object features; The step of performing feature similarity analysis between the target object features and the category index features in the category index feature group to obtain target feature similarity data includes: The object features corresponding to each feature extraction model are compared with the index feature group of the corresponding feature extraction model to obtain the feature similarity data corresponding to each feature extraction model. Based on the feature similarity data corresponding to each of the various feature extraction models, the target feature similarity data is determined.

3. The method according to claim 1, characterized in that, The feature density clustering results include: multiple sample feature sets corresponding to various feature extraction models, and the positive sample objects are selected in the following way: Obtain the sample object features corresponding to each of the multiple original sample objects. The sample object features corresponding to each original sample object include: the sample object features of the corresponding original sample object output by each of the multiple feature extraction models. Density clustering is performed on multiple features to be clustered corresponding to the target model to obtain multiple sample feature sets corresponding to the target model. The target model is any feature extraction model among the multiple feature extraction models, and the multiple features to be clustered are the sample object features corresponding to the multiple original sample objects output by the target model. Determine the positive sample feature set and the negative sample feature set from the plurality of sample feature sets; Based on the positive sample feature set and the negative sample feature set, the sample category indication information corresponding to the target model is determined. The sample category indication information is used to indicate whether each original sample object belongs to the target object category. Based on the sample category indication information corresponding to each of the multiple feature extraction models, category voting is performed on the multiple original sample objects to determine the positive sample object.

4. The method according to claim 3, characterized in that, The step of performing density clustering on multiple features to be clustered corresponding to the target model to obtain a set of multiple sample features corresponding to the target model includes: Perform feature distance analysis on every two features to be clustered among the plurality of features to be clustered to obtain feature distance data between every two features to be clustered; Based on the feature distance data and the preset feature neighborhood range, density clustering is performed on the multiple features to be clustered to obtain multiple initial feature sets; If the number of the multiple initial feature sets is greater than the number of preset sets, the neighborhood range of the preset features is increased to obtain the increased neighborhood range; Based on the increased neighborhood range, the step of performing density clustering on the multiple features to be clustered based on the feature distance data and the preset feature neighborhood range to obtain multiple initial feature sets is repeated until the number of the multiple initial feature sets is less than or equal to the preset set number. The current set of multiple initial features is used as the set of multiple sample features.

5. The method according to claim 1, characterized in that, The preset feature similarity threshold is adjusted in the following way: Determine the initial similarity threshold and the preset identification index threshold; The positive sample object features corresponding to the positive sample object are compared with the category index features in the category index feature group to obtain the sample feature similarity data. Based on the comparison results between the sample feature similarity data and the initial similarity threshold, the predicted category indication information of the positive sample object is determined; Based on the predicted category indication information, the recognition performance of the target object category is analyzed to obtain recognition performance indicators; Based on the comparison result between the recognition performance index and the preset recognition index threshold, the initial similarity threshold is adaptively adjusted to obtain the preset feature similarity threshold.

6. The method according to claim 1, characterized in that, The category index feature group includes: an index feature group corresponding to each feature extraction model in the multiple feature extraction models, and the category index feature group is constructed in the following way: Obtain the positive sample object features corresponding to the positive sample object, wherein the positive sample object features include: sample object features corresponding to the positive sample object output by each of the multiple feature extraction models; The features of the sample objects corresponding to the positive sample objects output by each feature extraction model are subjected to feature clustering processing to obtain at least one set of object features corresponding to each feature extraction model; The cluster center features corresponding to each of the at least one set of object features are used as at least one category index feature corresponding to each feature extraction model; Based on at least one category index feature corresponding to each feature extraction model, construct an index feature group corresponding to each feature extraction model.

7. The method according to claim 6, characterized in that, The method further includes: Obtain the new sample object features corresponding to the newly added positive sample object, wherein the new sample object features include: the sample object features corresponding to the newly added positive sample object output by each of the multiple feature extraction models; The features of the sample objects corresponding to the newly added positive sample objects output by each feature extraction model are subjected to feature clustering processing to obtain at least one set of newly added object features corresponding to each feature extraction model. The cluster center features corresponding to each of the at least one newly added object feature sets are used as at least one newly added index feature for each feature extraction model. The at least one newly added index feature is added to the index feature group corresponding to each feature extraction model.

8. An object recognition device, characterized in that, The device includes: The feature acquisition module is used to acquire the target object features corresponding to the object to be identified and the category index feature group corresponding to the target object category. The category index feature group consists of cluster center features obtained by clustering the positive sample object features corresponding to the positive sample objects. The positive sample objects are sample objects belonging to the target object category selected from multiple original sample objects based on the feature density clustering results. The feature density clustering results are feature clustering results obtained by performing density clustering on the sample object features corresponding to each of the multiple original sample objects. The feature similarity analysis module is used to perform feature similarity analysis between the target object features and the category index features in the category index feature group to obtain target feature similarity data. The object category indication information determination module is used to determine the object category indication information of the object to be identified based on the comparison result between the target feature similarity data and the preset feature similarity threshold corresponding to the target object category; the preset feature similarity threshold is obtained by adaptively adjusting the threshold based on the preset identification index threshold corresponding to the target object category and the sample feature similarity data corresponding to the positive sample object; the sample feature similarity data is used to characterize the feature similarity between the positive sample object features corresponding to the positive sample object and the category index features in the category index feature group.

9. An object recognition device, characterized in that, The device includes a processor and a memory, the memory storing at least one instruction or at least one program, the at least one instruction or the at least one program being loaded and executed by the processor to implement the object recognition method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores at least one instruction or at least one program segment, which is loaded and executed by a processor to implement the object recognition method as described in any one of claims 1 to 7.

11. A computer program product, characterized in that, The computer program product includes at least one instruction or at least one program segment, which is loaded and executed by a processor to implement the object recognition method as described in any one of claims 1 to 7.

Images