431 results about "Information embedding" patented technology

The invention discloses a Transform-based traffic flow prediction method and device capable of sensing a local-global time-space relationship, and the method comprises the steps: firstly obtaining the historical traffic data of all to-be-predicted road sections, the historical data comprising the traffic flow feature data of the predicted road sections, the static adjacent matrix data between road network nodes, and the sampling timestamp data of the historical data; secondly, a spatio-temporal information embedding layer is constructed to provide multiple types of embedding input for a model trunk, the learning ability of the model is enhanced, and three different types of embedding are respectively historical data information embedding, time information embedding and space node self-adaptive embedding; then, constructing a local-global time dependence extraction module, respectively learning short-time and long-time time dependence relationships in the data by using a multi-scale TCN and a self-attention mechanism in a time dimension, and meanwhile, introducing a double-path self-adaptive information gating fusion technology to realize effective fusion of time features of different hierarchies; then constructing a local-global spatial dependency extraction module, respectively learning local and global spatial dependency relationships in the data by using a dynamic-static graph convolutional network and a self-attention mechanism in spatial dimension, and realizing effective fusion of spatial features of different levels based on a two-way adaptive information gating fusion technology; and finally, mapping the potential spatial-temporal feature representation into a prediction result through a full connection layer network.

The invention provides a potential diffusion model-based building planning image generation method and system. The method comprises the steps of data preprocessing, potential diffusion model structure improvement, input condition optimization, model training, image generation and the like. The relevance between text description and architectural styles is enhanced by preprocessing data sets of architectural design images and urban planning images and introducing channel crossing and space crossing attention mechanisms. Meanwhile, an architectural style embedding vector, an architectural description text embedding vector and a geometric information embedding vector are combined as additional input conditions to ensure that the generated image contains complete information. In a model training process, a multi-scale strategy and depth separable convolution are adopted, so that the recognition capability and training efficiency of the model to a large scene are improved. The system comprises a data preprocessing module, a model optimization module, a model training module, an image generation module and the like, and can generate a high-resolution building image conforming to design specifications. The method and system can be widely applied to the fields of building design and urban planning, and design efficiency and accuracy are improved.

The present application provides a watermark information embedding method, device, storage medium and program product, wherein the method includes obtaining the target watermark information to be embedded and the original carrier information in response to a watermark embedding instruction; mapping the target watermark information to a latent variable space to obtain a watermark latent vector, and mapping the original carrier information to the latent variable space to obtain a carrier latent vector; fusing the watermark latent vector and the carrier latent vector to generate a noise latent vector with a watermark; denoising the noise latent vector to generate a target carrier information with the target watermark information. The present application realizes that when the target carrier information is subjected to various complex editing processes, the watermark information can still be effectively extracted, thereby improving the robustness of the watermark and providing a reliable data basis for subsequent digital information ownership protection.

The invention relates to a multi-category welding spot defect classification method, system and equipment based on Transform and channel interaction and a medium. The method comprises the following steps: preprocessing an input image, and respectively carrying out division, embedding, position coding and category information embedding on an input support image and an input query image; an image block embedding sequence containing position and category information is transmitted to a Transform encoder to extract embedding features containing category related information and original image blocks; performing adaptive adjustment on the image block embedding features of each channel by using the category-related embedding vector and the global visual context vector; judging whether the support images and the query images belong to the same category or defect type or not by calculating the similarity of the support images and the query images; outputting defect categories of the images through a classification network; according to the method, the precision of welding spot defect classification is effectively improved, and the problems of category confusion and small sample learning confronted by a traditional method are solved especially in the scene of various defect types and small sample size.

The invention provides a Mama-based multi-modal dialogue emotion recognition method and system, and belongs to the field of dialogue emotion recognition. The problems that an existing multi-mode emotion recognition method is limited in memory ability, information is not fully utilized, and noise accumulation exists in long sequence fusion are solved. Comprising the following steps: processing dialogue data in different modes by adopting different preprocessing modes to obtain corresponding feature codes; performing convolution and information embedding on different feature codes; external attention is adopted to capture semantic information in each modal; interaction among different modes is realized by using a cross fusion mechanism; filtering noise in each mode by adopting Kalman filtering and establishing a relation between the modes; fusing the semantic information of different modals by using grouping pooling and cross attention, and then fusing the fused semantic information with a Kalman filtering result to obtain a fusion result; mapping the fusion result into a predefined emotion category; the method and the device are applied to multi-mode dialogue emotion recognition.

The invention discloses a class incremental learning method, system and device based on a pre-training model and a medium. The method comprises the following steps: step 1, in class incremental learning of a first stage, aiming at a first task D1, finely adjusting a pre-training model ViT through a Prompt embedding method; obtaining a pre-training model after fine tuning; 2, combining and freezing the embedding layer function in the fine-tuning pre-training model obtained in the step 1 and the embedding layer function of the pre-training model ViT before fine tuning to obtain a combined model; step 3, performing class increment training from the beginning by using the combined model obtained in the step 2, and finally obtaining a class increment model with strong generalization and robustness; and 4, predicting a test sample by using the class increment model obtained in the step 3, and calculating the final classification precision. According to the method, the generalization ability of the model is kept while the adaptability of the model is ensured.

The invention discloses a 3D human body posture estimation method based on a diffusion model. According to the method, firstly, a continuous human body dynamic video is split into RGB images, two-dimensional coordinate information, embedded time steps and positions of all joint points of a human body are extracted from the RGB images through a 2D posture detector, then predicted 3D postures are output through a diffusion model, and low-frequency key information is reserved and high-frequency noise is filtered in combination with discrete cosine transform (DCT), so that the 3D postures are obtained. And the calculation complexity is greatly reduced. And based on a multi-hypothesis aggregation strategy of confidence and consistency evaluation, weighted fusion is performed on the predicted 3D postures, so that the precision and robustness of posture estimation are improved. The performance of the model in a dynamic scene is further enhanced by a time step embedding mechanism and time consistency scoring, so that the model adapts to a complex motion environment and a shielding condition. The method is low in computing resource demand, is suitable for being deployed on resource-constrained equipment in the fields of human-computer interaction, virtual reality, motion analysis, medical rehabilitation and the like, and has a wide application prospect.

The invention discloses a multi-mode traffic demand prediction method based on a graph neural network, which is characterized in that technologies such as multi-source fusion, multi-graph processing and meta-information embedding are introduced, and a complex space-time dependency relationship of multi-mode data of an urban traffic system is captured; and carrying out traffic demand prediction on the fused multi-mode data by adopting a space-time diagram convolutional network, a diffusion convolutional recurrent neural network and a Graph-WaveNet network model. Compared with the prior art, the method has the advantages of adjusting distribution of traffic resources, optimizing scheduling, reducing congestion, improving traffic efficiency and the like, modeling is carried out by collecting spatio-temporal data of different traffic modes and combining multiple graph structures, high-precision prediction of multi-mode traffic demands is achieved, and the method is suitable for popularization and application. The method solves the problems of mutual influence between complex traffic modes in urban traffic, spatial heterogeneity and multi-mode data fusion, can be widely applied to the fields of intelligent traffic, urban planning, public traffic scheduling and the like, and has a good application prospect.

The invention provides an interest point recommendation method based on a graph-enhanced user context information network. The interest point recommendation method comprises the following steps: acquiring historical track data of a user and preprocessing the historical track data; utilizing an interest point supplementing module to fill the interest point between two adjacent interest points of which the road network distance and the sign-in time difference are greater than a threshold value in the historical trajectory data to obtain enhanced trajectory data; inputting the enhanced track data into a user context information embedding module to extract a context information embedding vector of the user; according to the enhanced trajectory data of all users in the network, constructing a global trajectory graph by taking the interest points as nodes and taking a sign-in sequence of the users to the adjacent interest points as edges; inputting the global trajectory graph into a general travel module, respectively extracting POI embedding and user embedding by utilizing a GCN network and an MCL network to obtain a POI embedding matrix and a user embedding matrix, and fusing the POI embedding matrix and the user embedding matrix to obtain a POI-user embedding vector of the user; and inputting the context information embedded vector of the user, the POI-user embedded vector of the user and the global trajectory graph into a prediction module for prediction to obtain an interest point recommendation result of the user. According to the method, the core problem in the prior art is effectively solved, the accuracy and scene adaptability of recommendation of the next interest point are remarkably improved, and the method has important application value and wide development prospects in the fields of intelligent transportation, personalized position service and the like.

The invention belongs to the technical field of graph classification, discloses a graph classification method based on subgraph integration and position sensing, and relates to a graph classification method SIPA combining substructure embedding and node position sensing. The SIPA firstly extracts sub-graphs through two different strategies so as to capture multi-sample sub-structures in the graphs, the structural features of the sub-structures are coded by adopting a graph convolutional network, and information of different sub-graphs is effectively fused through an attention mechanism. Then, anchor nodes are introduced to calculate relative position information of nodes, and the information is embedded into node representation to better capture global position features. And finally, a graph information bottleneck mechanism is used for optimizing node representation and removing redundant information irrelevant to a classification task. The method not only effectively learns the local structure information, but also enhances the perception capability of the relative position information of the nodes. Experimental results show that SIPA is superior to an existing baseline model in five data sets, and the superiority of SIPA in a graph classification task is verified.

The invention discloses an automobile part supply chain entity relationship extraction method based on a heterogeneous graph neural network. The method comprises the following steps: S1, obtaining to-be-processed automobile part supply chain information and a predefined relationship type; s2, word information embedding and high-dimensional word vector embedding are carried out on the automobile part supply chain information and the relation type respectively, and initial representations of word nodes and relation nodes are obtained; s3, inputting the initial representations of the word nodes and the relation nodes into the trained relation extraction model, and outputting a plurality of subject-relation-object triples; and S4, taking all the subject-relationship-object triads as entity relationship extraction results of the supply chain information of the automobile parts. According to the method, firstly, the relation between entities is modeled into nodes in a graph in a unified mode, then node representation is updated through iteration fusion of a multi-layer heterogeneous graph neural network, subject features are fused into the node representation through an attention mechanism, and finally a subject-relation-object triple is extracted at a time.

The invention discloses a policy service recommendation method and system based on a large language model, and relates to the field of policy information, and the method comprises the steps: analyzing a policy file, extracting policy terms with a numbering structure, building a unique identifier, and forming a term node set; identifying explicit or implicit reference behaviors in terms, constructing a reference path, and generating a reference directed graph; constructing a clause mapping tree based on the target clauses; mapping tree nodes are extracted to form structured prompt information, and the structured prompt information is embedded into large language model input; generating a policy report paragraph containing clause reference content; and outputting the structured policy text with clause reference chain closure and recommending the structured policy text to the user. According to the method, the clear reference context combined with the user information can be provided for the large language model, the large language model is guided to generate a text with a reasonable structure and complete terms for recommendation, and the problem of inaccurate generation when an existing model processes a complex reference structure is solved.

The invention discloses a vehicle re-identification method, device, equipment and product based on multi-modal pre-training. The method comprises the following steps: acquiring vehicle data of a target vehicle; wherein the vehicle data at least comprises one of vehicle image data or text description information of the target vehicle; performing feature extraction on the vehicle data by adopting a trained multi-modal vehicle re-identification model to obtain final features of the vehicle data, and outputting information of the target vehicle according to the final features; wherein the trained multi-modal vehicle re-identification model comprises a pre-trained visual converter model, a pre-trained text graph model and a visual converter main network. According to the method, multiple query modes and text information embedding can be supported, capturing of image text information and understanding of image semantics are improved, the robustness and discrimination of extracted features are enhanced through pre-training, and therefore the vehicle re-recognition performance is improved.

The invention discloses a tourist attraction recommendation system adopting a virtual reality technology, belongs to the field of image data processing, and is used for solving the problems of how to improve defects in image information processing and improve scene fidelity and how to increase a virtual reality scene interaction function and improve scene interactivity. The system comprises seven main modules including a data acquisition module, a virtual reality scene construction module, a user information acquisition module, a recommendation algorithm module, a virtual reality experience module, a feedback collection module and a recommendation adjustment module. According to the technical scheme, through integration of the above modules, in the image modeling unit, a finer three-dimensional modeling technology is combined with accurate geographic information fusion and rich text information embedding operation, so that real scenic spots in a virtual scene can be restored as much as possible in a one-grass-one-wood and one-tile manner, highly-vivid and immersive experience is provided for a user, and the user experience is improved. And the user can feel the real atmosphere and detail characteristics of the scenic spot like being personally on the scene.

The present invention relates to a knowledge reasoning and fault diagnosis method based on knowledge graph, including obtaining fault record information and data knowledge, removing redundant and repeated information, performing data cleaning, obtaining relationship paths and entity relationship triples, and constructing case resource knowledge graph; embedding relationship path information into a convolutional encoder, obtaining a path feature vector, then inputting it into a BiLSTM network, obtaining a relationship path representation vector, and then inputting it into an ATT model, completing the relationship path through multi-step reasoning; extracting fault attribute feature information; calculating the weight value of each attribute, and calculating the overall similarity between the current case and the historical case through a neural network, completing the construction of a CBRDA-KG decision model; obtaining actual fault case information, and loading it into the CBRDA-KG decision model to obtain a fault diagnosis result. Compared with the prior art, the present invention effectively reduces the detection cost and detection time, and the accuracy and integrity are also greatly enhanced, and the diagnostic efficiency and the accuracy of the results can also be improved.

The invention belongs to the technical field of image analysis, and particularly relates to an electronic component defect detection method and system based on machine vision, and the method comprises the following steps: S1, obtaining visible light, X-ray and three-dimensional structured light images of an electronic component to be detected, and generating respective multi-scale feature pyramids through independent feature extraction networks; s2, generating a multi-modal fusion feature pyramid; s3, applying a candidate region generation network to propose a candidate defect region; s4, processing the spatial topological relation graph through a graph neural network, and embedding spatial context information into candidate defect region features; and S5, if the minimum distance between the candidate region features and all the pre-stored prototypes is greater than a set discrimination threshold, determining that the candidate region is a potential defect of an unknown type. According to the method, the detection range is widened, the robustness of feature expression is improved, the limitation of insufficient information of a single data source is overcome, and the accuracy, comprehensiveness and perspectiveness of detection are improved.

The invention provides an active-disturbance-rejection control method for an electromagnetic bearing two-level digital switching power amplifier embedded with model information. The method comprises the steps that an electromagnetic bearing two-level digital switching power amplifier mathematical model suitable for controller design is established based on the Fourier series theory and coil current characteristics; a direct current component in the mathematical model is defined as a modeling disturbance term, and feedforward compensation is directly carried out in the control law; parameter perturbation and an unmodeled part in the mathematical model are defined as unknown perturbation items, and real-time estimation is carried out through an extended state observer; a current item and a control input item in the mathematical model are used as known model information to be embedded into the design of the extended state observer; designing a feedback control law to carry out real-time compensation on the current item, the modeling disturbance item and the unknown disturbance item; a zero-order retainer and a current observer are adopted to carry out discretization processing on the controller, and the discretization processing is realized on a digital control platform based on a DSP (Digital Signal Processor). The method is high in current response speed and high in control precision.

The invention discloses a multi-protocol compatible Internet of Things platform processing method and system, and the method comprises the steps: carrying out the embedded coding of the message information and TCP attribute information of a target message object through employing an information analysis and coding technology based on artificial intelligence; then attention multi-dimensional fusion is carried out on the embedded target message information embedding features and the target message TCP attribute embedding features, and meanwhile, processing is carried out on message information and TCP attribute information of the messages to obtain multi-dimensional information interaction fusion features of all reference messages; therefore, the TCP type label of the target message object is intelligently judged according to the quick matching query representation between the multi-dimensional information interaction fusion feature of the target message and the multi-dimensional information interaction fusion feature of each reference message. Therefore, a new or custom extended protocol can be automatically learned and adapted, the dynamic change of a network environment and the evolution of a communication standard can be coped with without manually updating a rule, and the accuracy and the reliability of data processing are further improved.

The invention discloses a reversible information hiding method and system for enhancing image smoothness, and relates to the technical field of information security, and the system comprises a preprocessing module, an information embedding module and an optimization recovery module. According to the method, the Sobel operator is utilized to calculate the image gradient map, the region is divided based on the threshold T, the smooth region and the texture region can be accurately distinguished, an accurate basis is provided for differential processing of different regions, region characteristics are fitted, large-size blocks and sub-blocks of the smooth region are divided to facilitate subsequent refinement processing based on the gradient mean value, and the processing efficiency is improved. The small-size blocks of the texture area are beneficial to operation aiming at texture features, the rationality and effectiveness of overall processing are improved, meanwhile, the pixel mean value in the large-size blocks of the smooth area is calculated, the pixel and mean value difference value is recorded, reversible low-pass filtering is carried out, and the reversibility of the image in the whole information hiding and recovering process is ensured.

The embodiment of the invention provides an address data matching method and related equipment, and belongs to the technical field of geographic information services. The method comprises the following steps: constructing an address annotation corpus according to input address information data and a preset address database; the method comprises the following steps: generating a geographic information embedding vector according to a preset geographic information knowledge graph, performing address element analysis in combination with an address annotation corpus to obtain an address element sequence so as to construct a dictionary tree, and performing similarity screening through a spatial hierarchical matching algorithm to obtain a similar address set; generating an address embedding vector matrix through a preset word embedding vector model, and performing feature extraction through a preset semantic feature extraction model to obtain semantic-level similar features; according to input address information data, multi-dimensional character similarity matching is carried out to obtain character-level similar features, then weighted fusion is carried out in combination with semantic-level similar features, and target matching address data is determined according to a weighted fusion result. According to the embodiment of the invention, the address data matching accuracy and efficiency can be improved.

The invention discloses a generative image steganography method and device based on Stable Diffusion and discrete wavelet transform, and the method comprises the following steps: S1, sampling a potential representation of a to-be-generated image through employing a Stable Diffusion diffusion model, embedding secret information into a potential space of the diffusion model, processing the potential representation through employing discrete wavelet transform to carry out frequency domain information modulation, and carrying out the frequency domain information modulation; obtaining the potential representation after the secret information is embedded; s2, inputting the potential representation embedded with the secret information into a diffusion model, and generating a visual natural steganographic image through a diffusion inversion process; s3, inputting the received steganographic image into a diffusion model, and decrypting the received steganographic image without original model weight modification and empty prompt conditions to obtain embedded secret information; according to the method, the frequency domain embedding technology is combined with the potential diffusion model, so that high-fidelity, high-capacity and high-robustness image steganography is realized.

At least one embodiment relates to a method comprising obtaining an occupancy information indicating if blocks of a 2D regular grid defined over a projection plane are associated with depth information of an orthogonal projection of at least one point of a point cloud onto said projection plane; embedding said occupancy information in an image storing said depth information; and encoding said image.

The invention provides a Chinese short text-oriented place name recognition method, which relates to the field of natural language processing, and comprises the following steps of: designing a plurality of place name data storage tables based on an OSM map, a GeoNames database and an Gade map, and performing Chinese place name extraction on each place name data storage table to obtain a multi-source global Chinese place name dictionary; obtaining and preprocessing text data of foreign language social media, and matching place names of the preprocessed text data based on the multi-source global Chinese place name dictionary to obtain a place name candidate word set; the method comprises the following steps of: constructing an integrated neural network model by taking multi-source text information embedding of place name candidate word embedding, text information fusion coding of Cross-Transform Encoder and fusion information dynamic screening and decoding as composition units; and inputting the preprocessed text data and the place name candidate word set into the integrated neural network model, and identifying the real place name of the text data to obtain an identification result.

The invention provides a hyperspectral target detection method based on generative self-supervised learning and wavelet transform. The hyperspectral target detection method comprises the following implementation steps: acquiring a pre-training / testing sample set and a fine tuning sample set; constructing a spatial-spectral reconstruction model based on wavelet transformation and performing generative self-supervised pre-training on the spatial-spectral reconstruction model; constructing a priori constrained hyperspectral target detection network model and carrying out fine tuning training on the model; and obtaining a hyperspectral target prediction result and a detection result. In the pre-training process, the multi-scale spatial and spectral features can be effectively extracted and fused by combining a dual-branch information embedding module of wavelet transform, so that an encoder can learn feature representation from multi-scale spatial-spectral information; and the hyperspectral target detection network model is finely adjusted by using a priori constrained cross entropy loss function, and nonlinear transformation is performed on a prediction result, so that background information can be suppressed by fully utilizing a priori target spectrum, and the hyperspectral target detection precision is improved.

The invention discloses a heterogeneous network embedding method based on dynamic heterogeneous network decomposition and a hierarchical space-time attention mechanism. The method comprises the following steps: dynamic network decomposition: decomposing a heterogeneous network sequence into independent subgraph sets according to edge types; multi-order node attention aggregation: for each sub-graph, fusing information of nodes and different edge types; semantic level attention is embedded, wherein global semantic embedding is generated in combination with global relation type weight and meta-path long-short-term memory network coding; time dynamic modeling: respectively calculating a time attention weight and a Monte Carlo sampling approximate Horkes intensity, and then aggregating and splicing results of the time attention weight and the Monte Carlo sampling approximate Horkes intensity to obtain a final embedding result. According to the method, through cross-granularity semantic modeling and efficient time sequence dependence learning, an innovative solution is provided for node representation of the dynamic heterogeneous network.

The invention discloses a mapping knowledge domain embedding method fusing meta-information and logic rules, and relates to the technical field of computers. The method comprises the following steps: inputting a to-be-predicted knowledge graph into a knowledge graph embedding module; in the multi-granularity meta-information embedding module, for each entity to be predicted in the knowledge graph to be predicted, obtaining ontology information embedding and multi-hop neighbor relation embedding of the entity to be predicted; carrying out average pooling processing on each of the multi-hop neighbor relation embedding, and processing the neighbor relation embedding after the average pooling processing through a multi-layer perceptron and learnable parameters to obtain a hierarchical attention weight of each neighbor relation embedding; each multi-hop neighbor relation embedding comprises a plurality of neighbors; and performing weighted summation on the ontology information embedding and the multi-hop neighbor relation embedding based on the hierarchical attention weight to obtain meta-information embedding of a to-be-predicted entity in the to-be-predicted knowledge graph. According to the method, the determination precision of the meta-information embedding of the entity can be improved.

The invention discloses a rule engine quality inspection method and device, a storage medium and a computer program product, and relates to the technical field of software testing, the method comprises the following steps: receiving an expected target description input by a user, and determining an executable script code corresponding to a rule engine to be subjected to quality inspection based on the expected target description; extracting parameter information corresponding to the rule definition text of the rule engine to be subjected to quality inspection; embedding the expected target description, the executable script code and the parameter information into a preset cue word template; and calling a quality inspection model to perform quality inspection on the cue word template to obtain a rule engine quality inspection result. The rule engine is converted into the script code based on the mainstream computer programming language from the proprietary grammar, so that the quality inspection efficiency and the quality inspection accuracy of the quality inspection model on the rule engine are improved.

The present application belongs to the technical field of image analysis, and particularly relates to a kind of electronic component defect detection method and system based on machine vision, comprising the following steps: S1, the visible light, X-ray and three-dimensional structure light image of the electronic component to be measured are acquired, and respective multi-scale feature pyramids are generated through independent feature extraction network respectively;S2, multi-modal fusion feature pyramid is generated;S3, candidate region generation network is applied to propose candidate defect region;S4, spatial topological relation graph is processed through graph neural network, and spatial context information is embedded into candidate defect region feature;S5, if the minimum distance of candidate region feature and all pre-stored prototypes is greater than the set discrimination threshold, then it is determined that the candidate region is unknown type potential defect.The present application widens the detection range and improves the robustness of feature expression, overcomes the limitation of insufficient information of single data source, and improves the accuracy, comprehensiveness and foresight of detection.

The invention discloses a composite circuit substrate material parameter pixel gray scale coding method and device and a storage medium, and relates to the technical field of composite substrate material physical property modeling simulation. According to the method, the composite circuit substrate unit image is pixelated, the pixel gray value is used as a matrix material parameter coding space, and then matrix material parameter information is embedded into the substrate unit image, so that fusion coding of wiring pattern features and material mechanical parameters is realized when two or more materials exist in the same layer of substrate; and the accuracy of predicting the equivalent mechanical parameters of the substrate unit through the wiring image information of the substrate unit and the deep convolutional neural network is improved.

The invention discloses a YO-SAM2-based industrial pellet raw material instance segmentation method, which belongs to the field of image detection, and carries out target detection (an efficient up-sampling module EUCB, a multi-scale heterogeneous convolution module MSHCC and fast normalized weighted fusion FFusion) by means of a YOLOv10 framework and fusion of a brand new EMBSFPN (an efficient up-sampling module EUCB, a multi-scale heterogeneous convolution module MSHCC and a fast normalized weighted fusion FFusion); and then the output bounding box is used as prompt information to be embedded into a prompt encoder of the customized SAM2 to output a high-quality segmentation result. Experiments show that the model constructed by the method well overcomes the defects of traditional instance segmentation in an industrial pellet segmentation task, an excellent mAPmask score of 92.8% is shown on a self-constructed PelletORE pellet data set, the best accuracy is realized, and the method has high practical application and popularization values.