55 results about "Confidence region" patented technology

The invention discloses a natural language-to-SQL (Structured Query Language) method and system based on RAG (Random Access Gateway) enhancement, and the method comprises the steps: constructing a query intention graph through dependency syntax analysis, recognizing and complementing semantic missing components, and forming complete semantic representation; vector representation is carried out on the business term segments by adopting vectorization coding, accurate definitions of business terms are obtained from a factory structured knowledge base, and an enhanced context set is formed through expansion retrieval in a low-confidence region; an SQL template mapping network is established based on historical query records, a mapping relation matrix is generated through field candidate matching, mapping conflict positions are identified, and multiple SQL candidate sequences are generated; grammar verification is carried out on the candidate sequence to identify grammar errors, semantic consistency verification is carried out to calculate the intention alignment degree, and an optimal SQL statement is selected through a deviation correction factor; and performing formatting processing and statistical abstract on an execution result, and providing data query capability for scenes such as factory quality management and equipment maintenance.

The invention provides a universal visual perception method for an open world based on language guidance, and belongs to the crossing field of computer vision and natural language processing, and the method comprises the steps: generating an initial multi-modal fusion representation, and obtaining the initial multi-modal fusion representation through the collection of a video frame sequence and a language instruction through processing and fusion; then determining a target candidate region, and carrying out matching, screening and optimization based on semantic keywords and visual region features; then generating a target identifier, and allocating a unique identifier to the high-confidence region; a continuous tracking trajectory sequence is formed, and a bounding box is updated and the trajectory is smoothed in combination with an algorithm; when the target disappears, the state vector of the target is temporarily stored, and the tracking identifier is recovered when a similar region appears; optimizing the tracking sequence, and adjusting a bounding box to generate an optimized sequence; and finally, outputting target motion trail, position and state information. Through multi-modal fusion, an optimization algorithm and a recovery mechanism, the open world target identification and tracking effect is effectively improved, and the practical value is high.

The invention provides an SPECT pinhole collimator imaging truncation artifact filling method and related equipment, and relates to the technical field of CT imaging. The method comprises the following steps: reconstructing real projection data to obtain an original reconstructed image, and determining a high-confidence region in the original reconstructed image; mapping the real projection data to a projection image of an ideal detector radial distance to obtain mapping projection data; predicting projection data outside an effective area of a real acquisition detector and taking the projection data as simulation projection data; carrying out weighted reconstruction on the simulation projection data to obtain a complemented image containing a truncation region; and according to the high-confidence region, fusing the original reconstructed image and the complemented image to obtain a final image. The method aims at solving the problem of truncation artifacts caused by the fact that the probe cannot completely cover the whole human body projection image, negative effects caused by the truncation artifacts can be made up, and the effect of improving the imaging quality and the diagnosis accuracy is achieved.

The invention discloses an automatic detection system for packaging defects based on computer vision, and the system comprises an image preprocessing module which collects an RGB image, completes the normalization and geometric correction, and generates a standardized image frame; the energy diagram and gradient tensor field generation module is used for respectively calculating color variance, texture response and gray gradient; the initial superpixel division module is used for guiding geometric flow propagation based on the energy diagram and the tensor field to generate a first group of superpixels; the feature extraction and classification module is used for extracting regional features and outputting defect probability and category; the confidence evaluation module is used for constructing a confidence score graph and identifying a low-confidence region; the local re-segmentation module is used for re-dividing superpixels in a specified area; and the result output module is used for outputting the final defect position, category and confidence score. According to the invention, the accuracy and adaptability of defect detection are improved.

The invention discloses an engineering construction site real-time three-dimensional modeling method based on an unmanned aerial vehicle aerial image, and relates to the technical field of computer vision, and the method comprises the steps: collecting inclined image data and high-frequency pose data, obtaining sparse key point features, depth semantic features and high-frequency time sequence features, generating a three-dimensional space point cloud model, and recognizing a low-confidence region. And comparing the three-dimensional space point cloud model with the semantic tag with a dynamic semantic three-dimensional model of a previous period as a reference model to identify a semantic change area and semantic state conversion conforming to preset construction logic, constructing a dynamic semantic three-dimensional model of a current period, and generating an unmanned aerial vehicle flight parameter adjustment instruction. According to the method, the three-dimensional modeling quality is improved through cross-view semantic consistency verification, progress changes are interpreted based on construction logic, a building information model and time sequence prediction are combined, and high-confidence data support is provided for construction site management by solving a multi-objective optimization function and deciding unmanned aerial vehicle collection operation.

The invention discloses an unformatted scale bill intelligent conversion method and system based on image recognition, relates to the technical field of character intelligent recognition, and solves the technical problems that the accuracy of subsequent OCR recognition is unstable due to the fact that the inclination angle detection precision is low and noise filtering and contrast enhancement effects are limited by scenes. According to the method, table line detection and text line analysis are adopted for the structured / no-table scale bill, the inclination angle detection precision is improved, layered denoising and adaptive contrast enhancement are achieved, the text recognition accuracy in a complex scene is improved, PaddleOCR is dynamically switched based on the Chinese proportion, a low-confidence region recheck mechanism is combined, the recognition error rate is reduced, and the recognition efficiency is improved. According to the technical scheme, a scale document field dictionary and a relation graph are constructed, intelligent term matching and error correction are achieved, the recognition problem of uncommon words and industry exclusive vocabularies is solved, a hierarchical rule system of basic verification, association verification and compliance verification is adopted, format, logic and industry / enterprise compliance full dimensions are covered, and the verification coverage rate is increased.

The invention discloses a cognitive function screening system and method based on an MMSE (Minimum Mean Square Error) prediction model, and relates to the technical field of data analysis, the method comprises the following steps: collecting physical examination index data, removing missing records, adopting a multiple interpolation method for interpolation, and obtaining multiple sets of complete data sets; a continuous prediction model is constructed, MMSE continuous prediction values are obtained, and a sensitivity analysis report is generated; constructing a first-stage classification model, adaptively dividing an optimal threshold combination, and dividing a sample into a high-confidence region, a to-be-discriminated region and a low-confidence region; if the sample size of the to-be-discriminated region is higher than a preset training threshold value, constructing an enhanced feature set, and constructing a second-stage classification model; when the prediction probability reaches the optimal re-discrimination threshold value, the classification result in the first stage is corrected, and otherwise, the classification result is maintained; if not, maintaining the classification result; and integrating the classification results to obtain a final classification result of all the samples.

The invention relates to the technical field of obstacle avoidance path planning, and discloses an intelligent sports vehicle dynamic obstacle avoidance path continuous planning method comprising the following steps: S1, obtaining a racing track pre-stored map, and constructing a racing track reference coordinate system; relates to the technical field of obstacle avoidance path planning, and the method comprises the steps: predicting the probability distribution of the future position of an obstacle through employing a recursive Bayesian filtering algorithm based on the historical observation data of the obstacle, carrying out the modeling through multivariate normal distribution, converting the modeling into a geometric confidence region, and forming a confidence occupancy set covering the uncertainty of the obstacle; a space-time safety channel is generated through set operation by combining the intelligent vehicle appearance envelope sequence and the racing track boundary; the method does not need to depend on a single prediction track, and by depicting the uncertainty of obstacle movement, even if the actual position of the obstacle deviates from prediction, the intelligent vehicle can still run in the safety channel, so that the problems of planning failure and conflict between the vehicle and the obstacle caused by deterministic prediction of the actual position of the obstacle in the prior art are effectively relieved.

The invention discloses a medical image recognition method and system based on artificial intelligence, and the method comprises the steps: obtaining CT image data, pathological section data and PET metabolic activity data, introducing a GNN graph neural network to model the semantic association between multi-modal features, and outputting a fusion feature vector; inputting the fusion feature vector into a DSRN double-flow space-time recursive network, extracting 3D lesion morphological features of a single image by using a spatial flow, analyzing lesion growth kinetic parameters of a historical image sequence through a time flow, and generating a lesion malignancy probability index based on a gating fusion unit; based on the focus malignancy probability index, Monte Carlo Dropout sampling is utilized to generate a confidence interval, a clinically interpretable credibility score is output, and when the confidence is smaller than a threshold value, a low-confidence area is displayed for a doctor to check. The deep fusion of multi-source information is realized, the one-sidedness of a single mode is avoided, and the recognition accuracy and efficiency are improved.

The invention discloses an automobile intelligent key positioning method, device and system and a storage medium, a built-in sensor of an intelligent key receives a field intensity signal emitted by a body control module BCM through a low-frequency antenna, a set membership filtering algorithm is adopted to carry out optimization processing on sensor data, and then an ellipsoid confidence region containing a real field intensity value is constructed; therefore, the robustness of key position judgment is obviously improved. By the adoption of the technical scheme, the problem that in the prior art, key positioning based on Bluetooth signal strength or a traditional filtering algorithm is prone to being affected by signal jitter and instantaneous interference, and consequently judgment inside or outside a vehicle is inaccurate is solved.

A machine learning (ML) model may be trained and / or implemented to assist in quality control of variant calls in cohort level sequencing data. A computing device may receive a variant call file comprising cohort level sequencing data during training. Training data may be identified based on a predefined set of features related to genotyping rate or quality distribution. The training data may include positive labeled training data from at least one high-confidence region of the genome for variant calling and / or negative labeled training data from the variants identified from at least one low-confidence region of the genome for variant calling. The training data may be used to train the ML model to predict a machine learning site quality (MLSQ) score. The trained ML model may predict the MLSQ score for each variant. The MLSQ score may be used to filter the variants identified in the cohort level sequencing data.

The invention discloses a method and a system for generating randomized vortex induced force under the action of uniform flow, relates to the technical field of ship and ocean engineering, is used for predicting the drift trajectory of a target object under the interference of uniform flow, and comprises the following steps: acquiring vortex induced force original data of the target object in uniform flow with different flow speeds; the method comprises the steps of constructing a deterministic vortex induced force calculation formula based on vortex induced force original data, obtaining calculation data, decomposing a residual error of the calculation data and the original data, constructing a random process model driven by a physical mechanism, coupling the random process model with the deterministic formula, and constructing a randomized vortex induced force time history calculation formula. And inputting the randomized vortex induction force time history into an underwater motion equation, predicting an underwater drift trajectory of the target object through Monte Carlo simulation, obtaining a trajectory end point, generating a trajectory end point probability distribution diagram, and constructing a confidence region. According to the method, the prediction precision of the underwater drift trajectory of the target object is improved, the range of a confidence region is narrowed, and the marine search efficiency is improved.

The invention provides a breast ultrasound image segmentation method and system based on decoupling learning, and relates to the technical field of breast ultrasound image processing and analysis, and the method comprises the steps: obtaining a to-be-segmented breast ultrasound image; inputting the breast ultrasound image into the decoupling learning segmentation model to generate a prediction probability graph, and mapping the prediction probability graph into a final focus segmentation result; according to decoupling learning, a two-stage training method is adopted, an encoder and a decoder of a segmentation model are trained respectively, in the first-stage training process, fuzzy activation is achieved based on region recombination, a high-confidence region is dynamically replaced with a fuzzy block, gradient flowing is enhanced through a context with stable disturbance, and the robustness of the segmentation model is improved. Structural conflicts and semantic variability are generated in a high confidence region, forcing an encoder to learn semantic invariance under context changes. According to the method, the challenges that a current breast ultrasound image segmentation method is insufficient in fuzzy region processing optimization, fragmentized in feature representation and reduced in boundary quality are solved, and the segmentation precision and robustness are improved.

This invention provides a method and system for large-scale model preference optimization based on reward margin constraints. It constructs a third-order margin partitioning criterion and dynamically estimates the quantile threshold of the reward margin using exponential moving averages, adaptively dividing preference pairs into uncertainty, buffer, and confidence regions. Subsequently, a differentiated mapping strategy is applied to different regions. Finally, the differentiated mapping is integrated into a truncated sigmoid function, and a TruncPO loss function is constructed to optimize model parameters for human preference alignment tasks in large language models. This invention can improve the original win rate of different models by 4% in benchmark tests such as AlpacaEval2 and Arena-Hard, while reducing KL divergence fluctuation by more than one-third, effectively balancing the order consistency and probabilistic calibration of preference optimization. It is suitable for deployment in large language model application systems requiring accurate preference alignment, such as intelligent dialogue, automatic problem solving, and intelligent education.

This invention discloses a data annotation method and system for autonomous driving, relating to the field of data annotation. First, multi-source sensor data is acquired and initially annotated using an automated model. Then, a joint optimization algorithm decomposes and reconstructs features, improves boundary annotation accuracy, and identifies low-confidence regions. Based on a deep active learning strategy, prediction entropy, Bayesian divergence, and task-level uncertainty are fused to screen high-value samples. Ground truth labels are obtained through manual verification, while low-confidence regions are optimized to generate supplementary labels. These two types of labels are used as incremental training data, and the model mapping matrix is ​​updated through topological residual projection. Finally, the model is deployed for road testing, and problematic data is collected, triggering a new annotation optimization process to form a closed-loop iteration. This invention improves annotation accuracy and efficiency, achieves efficient incremental model updates, and constructs a continuously evolving annotation closed loop, providing support for the iteration of autonomous driving models.

The invention discloses an intelligent decision support method and system fusing multi-source financial data and a security fence, and relates to the field of financial data processing, and the method comprises the steps: analyzing a customer portrait generation request, determining a portrait task scene, and converting the portrait task scene into a task context vector; according to the target customer identifier, recalling the original data, and decomposing the original data into a plurality of feature atoms; labeling metadata for each feature atom; constructing a financial feature topology network based on the plurality of feature atoms; calculating a dynamic confidence score of each feature atom; according to the dynamic confidence score, distributing the plurality of feature atoms to different confidence regions of a structured cue word template to generate a structured cue word; inputting the structured cue word into the large language model to generate a customer portrait report; and based on the customer portrait report, customer insight analysis and personalized verbal skills are generated and displayed. According to the method, effective credibility evaluation can be carried out on the multi-source financial data, so that the fact consistency is ensured.

The invention provides an asset tracking system (100) for tracking a target tag (10) in a space (500), wherein the target tag (10) is configured to emit a target beacon signal, wherein the asset tracking system (100) comprises a plurality of listener nodes (110) arranged in the space (500) and configured to detect the target beacon signal, wherein the asset tracking system (100) comprises a control system (300), wherein the control system (300) has access to (i) listener position data and (ii) map data, wherein in an operational mode: the control system (300) determines a presence of an object (200), wherein an object tag (210) is associated with the object (200), and wherein the object tag (210) is configured to emit an object beacon signal, wherein the plurality of listener nodes (110) is configured to detect the object beacon signal and provide a related object signal to the control system (300); the control system (300) determines, for each listener node (110), a set of confidence regions (230) based on the related object signal, the listener position data, and the map data; the plurality of listener nodes (110) detects the target beacon signal and provides a related target signal to the control system (300); and, the control system (300) determines a target tag position of the target tag (10) based on the related target signal, the listener position data, the map data, and the set of confidence regions (230) of the plurality of listener nodes (110).

The invention provides an image definition evaluation method, device and equipment, and the method comprises the steps: dividing an image into different confidence regions according to the local features of different positions in the image, and enabling the local features to be used for evaluating the confidence of the definition of the corresponding position to the real resolution, the local features comprise local frequency domain features and local time domain features; respectively segmenting the different confidence regions obtained by division into a preset number of confidence sub-regions; respectively carrying out definition detection on each confidence coefficient sub-region to obtain the definition of each confidence coefficient sub-region; and calculating the definition of the image by using a fusion algorithm based on the definition of each confidence coefficient sub-region. By means of the image definition evaluation method and device, in the image definition evaluation process, the algorithm complexity can be reduced, and meanwhile high accuracy is achieved.

The invention relates to the technical field of map generation, in particular to a topographic map intelligent labeling and generating method based on a knowledge graph, and the method comprises the steps: obtaining data in real time; constructing a knowledge graph; generating a to-be-labeled element; carrying out preliminary annotation integration; carrying out low-confidence correction; dynamically updating the annotation; generating a primary topographic map; adjusting the historical duration; and generating a final plate. According to the method, a knowledge graph structure is introduced, spatial geometry, texture and spectral features of image data are fused with semantic features in text terrain description information, a feature vector library and hierarchical attribute relationship is constructed, and a labeling confidence calculation and dynamic updating mechanism is introduced in a labeling stage; and the spatial distribution and feature difference degree of the low-confidence region is corrected, so that the problems of inaccurate terrain change capture and limited labeling precision caused by single data source processing, static knowledge representation and insufficient utilization of multi-temporal data are effectively solved.

The invention discloses a robust unsupervised domain adaptive streetscape semantic segmentation method based on Vision Mama. The invention provides an end-to-end training framework for a cross-domain semantic segmentation task. The end-to-end training framework is a coding and decoding model-DAEDM constructed based on a Vision Mama framework. According to the method, a long-distance context dependency relationship in an image is captured through state space modeling, and a domain invariant context enhancement module is introduced to apply statistical disturbance to features of a low-confidence region in a target domain, so that the robustness of cross-domain features is improved. In order to enhance the segmentation capability of a small target and a boundary region, the designed decoder combines a Vision Mama structure and a channel attention mechanism to realize effective fusion and semantic detail recovery of multi-scale features. Compared with a traditional unsupervised domain adaptation method based on a convolutional neural network or visual Traformer, the method has the advantages that the segmentation performance of small targets and rare categories in a target domain is remarkably improved in a plurality of typical migration tasks, and excellent cross-domain generalization ability and practical application value are shown.

A method for generating an augmented reality image from first and second images, wherein at least a portion of at least one of the first and the second image is captured from a real scene, identifies a confidence region in which a confident determination as to which of the first and second image to render in that region of the augmented reality image can be made, and identifies an uncertainty region in which it is uncertain as to which of the first and second image to render in that region of the augmented reality image. At least one blending factor value in the uncertainty region is determined based upon a similarity between a first colour value in the uncertainty region and a second colour value in the confidence region, and an augmented reality image is generated by combining, in the uncertainty region, the first and second images using the at least one blending factor value.

The invention relates to the technical field of geographic information measurement, in particular to a high-precision geographic space information acquisition surveying and mapping method and system, and the method specifically comprises the steps: constructing a fusion point cloud data set through RGB images, thermal infrared images and combined point cloud data at all moments, and analyzing the terrain complexity; shielding omission features of the point cloud data are analyzed through the data set, and blind area effective scores are constructed; based on the difference between the vertical displacement of the echo intensity at each moment and a preset reference truth value, constructing the reliability score of the data at each moment, and combining the effective score of the blind area and the terrain complexity to construct the shielding compensation confidence at each moment; triggering a path planning algorithm based on the shielding compensation confidence, and guiding the unmanned aerial vehicle to avoid obstacles to complementarily collect a low-confidence region; the integrity and the micro-feature capture rate of the data in the shielding area are remarkably improved, the blind area of high-precision surveying and mapping is reduced, the integrity of the scene data is improved, and the engineering availability of the complex scene data is ensured.

The invention provides an electric vehicle and power grid combined dispatching method, device and equipment and a storage medium, and relates to the technical field of artificial intelligence. The method comprises the steps that an electric vehicle and power grid combined dispatching problem is coded into a bigraph structure, the electric vehicle and power grid combined dispatching problem is modeled into an MISOCP, the MISOCP is limited by target constraints, and the target constraints comprise a power distribution network constraint and a vehicle-network cooperation constraint; according to the bipartite graph structure, scheduling prediction is carried out, and a scheduling prediction probability is obtained; constructing a confidence domain according to the scheduling prediction probability; and solving the sub-problem determined by the confidence domain in the confidence domain by taking minimization of power generation cost and minimization of electric vehicle energy interaction cost as targets, and obtaining scheduling parameters in the scheduling period. According to the scheme, the scheduling parameters are solved in the confidence domain, so that the calculation complexity of electric vehicle and power grid joint scheduling can be reduced, and the calculation overhead is reduced.

The invention discloses a BD-Net-based wind driven generator blade defect two-stage detection and positioning method, and the method comprises the steps: firstly carrying out the preliminary detection of a blade high-resolution image obtained by an unmanned plane through GSS-Det, and outputting a defect candidate region and the confidence coefficient of the defect candidate region; a dynamic decision-making mechanism is carried out according to the confidence, and a high-confidence region directly enters a fine screening stage; performing high-resolution re-cutting on the medium-confidence region and performing secondary detection; the low confidence coefficient region is filtered out; carrying out fine screening by adopting a double-branch attention mechanism network D2A-Net, carrying out fine-grained classification and accurate segmentation on leaf defects, and outputting defect types, severity levels and corresponding segmentation masks; and establishing a camera internal reference calibration model, calculating a mapping relation between pixel coordinates and blade physical coordinates, mapping defect pixel coordinates to the blade physical coordinates, and outputting an accurate physical position of the defect. According to the invention, the accuracy, efficiency and positioning precision of blade defect detection can be obviously improved.

The invention relates to the technical field of data calibration, and discloses a geophysical survey data feedback calibration method, which is characterized in that an initial model is established through geological data integration, and a low-resolution region is calculated and identified by using a sensitivity matrix. A multidirectional measuring line array is deployed in an acquisition stage, and the position of a seismic source is dynamically adjusted by monitoring an offset distance distribution heat map in real time. And when frequency spectrum missing is detected, low-frequency seismic source supplementary excitation is automatically triggered. And for the acquisition data missing region, a low-rank tensor completion algorithm is adopted to recover a complete data volume. In the inversion process, long-offset data are preferentially utilized to construct a macroscopic velocity model, and then short-offset reflected wave data are gradually introduced to refine a local structure. And the model uncertainty analysis module automatically identifies a low-confidence region, generates supplementary acquisition coordinates and drives field equipment to implement directional encryption, newly added data is reinjected into an inversion process after being quickly processed, and a continuously optimized closed-loop system is formed.

The invention discloses a camouflage target detection system based on an attribute classification guidance Transform model. The camouflage target detection system comprises a visual feature extraction module, an attribute classification guidance branch, a target segmentation branch and a multi-task joint loss function module. An attribute classification task is introduced into a Transform network, and modeling and guiding of semantic information of camouflage targets such as multiple targets, shielding and complex shapes are achieved; meanwhile, a detail texture extractor and a cross-layer attention mechanism are designed to enhance fine-grained structure perception, and a low-confidence region is optimized and adjusted through a feature self-adaptive refining strategy; and optimizing an attribute classification result and a camouflage target segmentation result through a multi-task joint loss function. According to the method, the problems of fuzzy segmentation boundaries and detail missing are effectively improved, and the discrimination and integrity of target features are enhanced; the semantic perception capability of the system is obviously improved, and the camouflage target can be identified more accurately.

The invention discloses an asphalt mixture CT (Computed Tomography) image small sample multi-scale classification method based on a deep reasoning decision, belongs to the technical field of traffic infrastructure digitization, and aims to solve the problems that a traditional threshold segmentation method is insufficient in multiphase material processing capability and a pure deep learning model is insufficient in mesoscopic interface segmentation precision. The method comprises the following steps: S1, constructing an asphalt mixture CT (Computed Tomography) image mixed data set; s2, a multi-scale expansion convolution model CNN-S is built, and aggregate and asphalt mortar edge multi-scale classification is achieved; the CNN-S captures macro-micro characteristics of aggregates, gaps and mortar through multi-expansion-rate parallel convolution; a causal feature correction module is introduced, a high confidence region and a low confidence region are divided, a backbone network processes a macrostructure, and an edge branch decoder identifies a mesoscopic interface; s3, dynamic reasoning optimization: deploying a DeepSeek-R1 reasoning model as an auxiliary dynamic decision maker, and recording a misjudgment mode; and based on training indexes and expert knowledge, dynamically correcting CNN-S model weight and a loss function to realize self-optimization training.

This invention discloses an intelligent conversion method and system for unformatted weight slips based on image recognition. This invention relates to the field of intelligent text recognition technology, solving the technical problems of low accuracy in tilt angle detection and scene-dependent noise filtering and contrast enhancement, leading to unstable OCR recognition accuracy. This invention improves tilt angle detection accuracy by employing table line detection and text line analysis for structured / unformatted weight slips respectively. Layered denoising and adaptive contrast enhancement improve text recognition accuracy in complex scenes. PaddleOCR is dynamically switched based on the proportion of Chinese characters, combined with a low-confidence region re-examination mechanism to reduce the recognition error rate. A weight slip domain dictionary and relationship graph are constructed to achieve intelligent terminology matching and error correction, solving the problem of recognizing rare words and industry-specific vocabulary. A layered rule system of basic verification, association verification, and compliance verification is adopted, covering all dimensions of format, logic, and industry / enterprise compliance, improving verification coverage.

The application belongs to the field of computer data processing, and more particularly relates to a weakly supervised semantic segmentation method based on shape block semantic correlation degree. The method comprises the following steps: S1, inputting an original image into a classification network to obtain a class activation map; S2, obtaining graph structure data with shape blocks as nodes by dividing the original image through a shape division module; S3, performing shape block pooling on the class activation map by using the shape block division result in S2 to obtain a pooled class activation map; S4, training a semantic correlation degree network by using the confidence region in the pooled class activation map; S5, performing semantic classification on the graph nodes by using the adjacency matrix output by the semantic correlation degree network, and aggregating the nodes into pseudo labels; and S6, training a semantic segmentation network by using the pseudo labels, wherein the network receives an original image and outputs a predicted semantic segmentation result.

Computer-implemented method for enlarging a region (600) in a medical image (400), wherein the method comprises: 1a) Receiving (S302) a medical image (400); 1b) Receiving (S304) a seed point (404) in the medical image (404); 1c) Extracting (S306) first features (412) from the medical picture (400); 1d) Applying (S308) a trained first machine learning algorithm (702, 704) to the extracted first features (412) to obtain a classification of tissue (402) at the seeding point (404); 1e) Applying (S310) a trained second machine learning algorithm (702, 706) to the extracted first features (412) to obtain a first confidence region (V1) relative to the seed point (404), wherein the first confidence region (V1) is a region where the tissue (402) is expected to be the same as at the seed point (404); and 1f) Extend (S312), in a graphical user interface (110), a region (600) in the medical image (400), starting from the seed point (404), to include the first confidence region (V1).