526 results about "Supervised training" patented technology

The invention discloses a winter wheat LAI and SPAD estimation method based on a lightweight semi-supervised model, and relates to the technical field of agricultural remote sensing monitoring, and the method comprises the steps: obtaining multispectral image data of a winter wheat key growth period, and carrying out the preprocessing of the multispectral image data to generate a standardized multichannel vegetation index image; the method comprises the following steps: constructing a lightweight semi-supervised model MCVI-SANet, and carrying out self-supervised training on the MCVI-SANet by adopting a semi-supervised training strategy driven by VICReg; and inputting the multi-channel vegetation index image into the trained MCVI-SANet, and outputting quantitative estimation results of the LAI and SPAD of the winter wheat. Through combination of a saturation perception mechanism and semi-supervised learning, estimation deviation caused by dense canopy vegetation index saturation and data noise is effectively eliminated, and the estimation precision and generalization ability in a complex agricultural scene are improved while the lightweight deployment characteristic of the model is ensured.

The invention discloses a weak supervision image semantic segmentation system and method based on an attention mechanism, and the method comprises the following steps: obtaining to-be-segmented image data and an image-level label, and extracting a multi-level semantic feature map; fusing the multi-level feature map with the space and the channel dimension to generate an attention feature map; combining the attention feature map with an image-level label to obtain a target area positioning map; inputting the attention feature map and the target area positioning map into a double-edge reconstruction network to generate a high-quality pseudo-label map; carrying out progressive decoding structure convolution on the pseudo label graph and the multi-level semantic feature graph to generate a segmentation prediction graph; constructing a supervision signal based on the segmentation prediction map and the pseudo-label map to obtain a supervision training result; dynamically updating the supervised training result to generate an updated pseudo-label graph; and continuously iterating based on the updated pseudo-label graph until the loss function is converged, and outputting an image semantic segmentation result. According to the invention, weak supervision image semantic segmentation based on the attention mechanism is realized.

The invention discloses an optical cable micro-disturbance identification method based on physical simulation and self-supervised time sequence decoupling, and relates to the technical field of optical cable identification, and the method comprises the steps: constructing a physical digital twin simulator, and generating a high-fidelity training set; constructing a deep learning model, wherein the deep learning model adopts a lightweight time sequence decoupling network; training the model by adopting a staged training strategy, and sequentially carrying out self-supervised noise distribution pre-training, simulation supervised training and spectral domain physical consistency fine tuning operation; inputting DAS time sequence data collected in real time into the trained model, and outputting the data as an optical cable identity ID and a physical position; the lightweight time sequence decoupling network comprises a physical guide preprocessing module, a lightweight U-Net separation module, a sparse gating module and an intelligent parallel decoding module. Through the technical means of simulation-driven data generation, staged training strategies and the like, the defects of the prior art in the aspects of reducing the data cost, improving the detection capability in a low SNR environment, realizing multi-source blind source separation and the like are overcome.

The invention discloses a motion recognition method and system based on a redox photoelectric memristor, a terminal and a storage medium, and the method comprises the steps: selecting classical motions based on a human body motion data set, extracting time sequence data, and coding the time sequence data into an optical pulse sequence; constructing a reservoir array composed of a plurality of photoelectric memristors, and expanding an optical pulse sequence signal into a high-dimensional state vector; constructing a supervised training model, solving a weight matrix, and constructing a memristive cross array; and outputting an action classification result through simulation domain operation based on the output current multi-path light current signals in combination with the memristor cross array. According to the method, the motion features can be directly fed into a rear-end classification network for action recognition without depending on a complex digital feature extraction algorithm, so that the transmission and processing overhead of redundant data is fundamentally eliminated; a high-efficiency, low-delay and high-robustness hardware solution is provided for real-time and anti-noise motion recognition in scenes such as intelligent monitoring and man-machine interaction.

The invention relates to a semi-supervised learning data anomaly intelligent identification and treatment system, which is applied to a hydrogen energy commercial vehicle, and comprises a data acquisition and preprocessing module, which is arranged on a vehicle-mounted terminal and is used for acquiring a hydrogen storage system signal, a hydrogen supply system signal, a fuel cell system signal and a whole vehicle system signal in real time, processing the acquired signal data; the semi-supervised anomaly recognition module is arranged on a cloud platform, is connected with the data acquisition and preprocessing module, and is used for fusing the processed signal data into a rule engine and semi-supervised learning, constructing a semi-supervised training data set, and performing deep auto-encoder model training through the data set so as to realize accurate anomaly recognition of few sample scene writing; and the exception treatment and feedback module is arranged at an edge node, is connected with the semi-supervised exception recognition module, carries out exception recognition through a trained model, carries out graded treatment according to the exception severity, and establishes a model evolution mechanism to realize continuous optimization of the system.

The invention discloses a semi-supervised LPI radar signal modulation identification system and method based on entropy perception pseudo labels, and relates to the technical field of radar signal processing and mode identification. The system comprises a preprocessing module, a multi-scale reconstruction enhancer, a classification backbone network and a semi-supervised training module. The multi-scale reconstruction intensifier is used for reconstructing dual-channel separation through high-frequency detail enhancement and a low-frequency structure and enhancing discriminative characteristics in a noise environment; the classification backbone network introduces an adaptive contraction unit to realize channel-level noise suppression; and the semi-supervised training module dynamically evaluates the uncertainty of the unlabeled samples by adopting an entropy sensing mechanism, and generates weighted pseudo labels to carry out consistency regularization training. The method realizes signal modulation identification based on the system. According to the method, the problem of feature shielding under the condition of low signal-to-noise ratio is solved, the dependence of the model on labeled data is reduced through a reliable pseudo label generation mechanism, and stable and efficient modulation identification can still be realized in a severe channel environment with scarce labeled data.

The invention relates to the field of industrial defect detection, in particular to an industrial defect detection self-supervised segmentation method for iterative pseudo label refinement, which comprises the following steps of: constructing a system comprising a strategy model, a refinement model, a reward model and a meta-learning training module; inputting the multi-modal data into the strategy model, and outputting a rough defect mask; based on the rough defect mask, prompting refinement is carried out through a refinement model, and a refinement mask is output; based on the refinement mask, the to-be-detected product image, the standard template image, the depth image and the infrared image, calculating a comprehensive quality score through a reward model, and screening high-quality samples with qualified scores; and updating a preset training data set based on the high-quality sample, performing supervised training on the strategy model by using the updated training data set, and repeating the steps to form an iterative loop. By constructing a self-supervised closed loop, a system can be driven to autonomously learn defect features from an unlabeled production line multi-modal image only by a small amount of initial reference data.

The invention discloses a living cell super-resolution imaging method and system based on self-supervised learning. The method comprises the following steps: S1, obtaining a fluorescence microscopic image through a standard fluorescence microscopic system; s2, for the acquired single noise image, generating positive and negative sample pairs of self-supervised training data through an autonomously designed and optimized self-supervised strategy; s3, constructing a front denoising neural network and a rear deconvolution network, and performing network training; and S4, inputting a newly obtained noise image into the trained front denoising network to obtain a denoised image, and inputting the denoised image into the rear deconvolution network to obtain a final super-resolution reconstructed image. According to the invention, the standard fluorescence microscopic system and the self-supervised denoising processing module are combined, so that high-quality denoising and super-resolution reconstruction of the fluorescence microscopic image can be realized in a low-photon signal scene.

The invention discloses a visual language model reasoning acceleration method based on input-parameter joint pruning, which comprises the following steps of: inputting an image-text sample into a visual language model, constructing a visual token sequence, a text token sequence and comprehensive importance scores after layer embedding and noise disturbance, and jointly inputting the comprehensive importance scores into a meta router to construct a pruning strategy; introducing KL divergence to quantify the influence of a pruning strategy on output distribution of the visual language model, and constructing a preference sample pair; carrying out self-supervised training on the element router by using the preference sample pair and a direct preference optimization method; the trained meta-router is applied to a visual language model and used for generating a corresponding pruning strategy when a newly obtained image-text is input and reasoned; and after the visual language model performs pruning of the visual token and the converter layer according to the pruning strategy, reasoning is performed on image-text input, and a reasoning result is output.

The invention discloses a self-supervised defect representation learning method in an industrial scene, and the method comprises the steps: constructing a normal sample data set of an industrial product, carrying out the preprocessing, and obtaining a preprocessed normal sample; inputting the preprocessed normal samples into a multi-scale feature pyramid double-branch encoder network, extracting and fusing global semantic features and local detail features, and obtaining fused features; carrying out self-supervised training through a progressive temperature regulation contrast learning mechanism by utilizing the fusion features so as to learn feature distribution of normal samples; the fusion features are decoded and reconstructed through a semantic consistency reconstruction network, and reconstruction loss is calculated to optimize feature representation; and calculating a comprehensive abnormal score of a to-be-detected sample based on the trained model, establishing a dynamic threshold according to statistical distribution of the comprehensive abnormal score, and judging defects by comparing the comprehensive abnormal score of the to-be-detected sample with the dynamic threshold. According to the invention, the generalization capability and the detection precision of the industrial quality inspection system are obviously improved.

The invention provides a construction method of an ocean observation and exploration large model, and belongs to the technical field of large models.Multi-mode original data are collected by constructing a multi-source ocean data collection matrix, an environment change degree vector is established, preprocessing and noise reduction are conducted on the original data by adopting a nonlinear matrix mapping algorithm based on a Gaussian kernel function, and the large model is constructed. An ocean observation and exploration large model architecture of a liquid neural network structure is constructed, different branches are made to process input data of different dimensions by means of asymmetric design, a super sparse reconstruction matrix is established, and high-dimensional original data are reconstructed from low-dimensional observation by means of a compressed sensing reconstruction mechanism. And finally, a supervised training process is executed to optimize model parameters so as to complete the construction of an ocean observation and exploration large model, and the technical problem that high-precision fusion modeling of ocean multi-modal observation data is difficult to realize under the conditions of spatial-temporal distribution sparsity and data isomerism is solved.

The invention relates to the technical field of image processing, and discloses a laser holographic aberration compensation system and method based on deep learning, and the system comprises a first module which determines a phase-to-gray lookup table; the second module is used for calculating to obtain stacking strength; the third module is used for obtaining a Zernike polynomial coefficient through a convolutional neural network regression device; the fourth module is used for calculating to obtain predicted intensity and an aberration estimator; the fifth module is used for outputting two grey-scale maps; and the sixth module outputs two phase diagrams. According to the method, self-supervised training data is constructed through three-plane intensity collection, and an angular spectrum method physical model and Zernike polynomial coefficient low-dimensional representation are combined, so that a convolutional neural network regression device can stably learn aberration mapping, and compensation precision and generalization ability are both considered; double-phase two-frame time division coding is adopted, the output characteristics of a phase type spatial light modulator can be adapted, and cooperative control of aberration compensation and target complex field reproduction is achieved.

The invention is suitable for the technical field of computer vision, and provides a self-supervised monocular depth estimation method based on wavelet feature enhancement, and the method comprises the steps: firstly obtaining a two-dimensional target image and an adjacent image frame, and then generating a multi-scale first feature map and a multi-scale second feature map based on a main encoder and a wavelet feature extractor, the method comprises the following steps of: constructing a multi-scale second feature map, constructing a wavelet guide feature fusion module, injecting and enhancing high-frequency detail information of the multi-scale second feature map to a multi-scale first feature map based on the wavelet guide feature fusion module, generating a plurality of third feature maps, generating a multi-scale depth estimation map based on a decoder, and finally obtaining a multi-scale depth estimation map based on relative pose information, the multi-scale depth estimation map and a luminosity consistency error. And carrying out self-supervised training on the to-be-trained deep learning model. According to the method, the limitation of high-frequency information loss in a sampling process in a traditional method can be overcome, structural details and boundary information in depth estimation are effectively enhanced and supplemented, and a depth map with higher quality is obtained.

A method and apparatus for training a monocular depth estimation (MDE) network, including: obtaining a source dataset including a first source image and a first ground truth depth map corresponding to the first source image; obtaining a target dataset comprising a first target image and a second target image; generating an estimated first source depth map corresponding to the first source image using the MDE network; generating an estimated target depth map corresponding to the first target image using the MDE network; generating an estimated relative pose based on the first target image and the second target image using a pose network; and training the MDE network and the pose network by performing mixed supervision training, wherein the performing the mixed supervision training includes performing fully-supervised training based on the estimated first source depth map and the first ground truth depth map, and performing self-supervised training based on the estimated target depth map and the first estimated relative pose

The invention discloses a light field microscopic three-dimensional imaging method combining self-supervised learning and optical constraint. The method comprises the following steps: constructing a light field microscopic three-dimensional reconstruction reference data set; according to the invention, the two-dimensional light field image collected by the microscope is divided into a plurality of sub-aperture views, angle-space joint representation is formed, angle information and space information in the light field image are captured more comprehensively, and a richer and more accurate feature basis is provided for subsequent three-dimensional reconstruction; a self-supervised pre-training mode is adopted, so that the model has a certain knowledge basis in an initial stage, and in subsequent supervised training or fine tuning, convergence can be faster, and the training efficiency is improved; according to the method, physical constraints are integrated into the training process, so that the reconstructed three-dimensional structure is not only visual reasonable, but also credible in physical significance, the designed loss function can more accurately measure the difference between the reconstruction result and the real three-dimensional structure, and the quality and precision of the reconstruction result are improved.

The invention discloses an event camera image reconstruction method of a multi-frame fusion network based on optical flow guidance, and belongs to the technical field of computer vision and image processing. The method comprises the following steps of: 1, acquiring event data, and converting the event data into continuous and smooth space-time voxels by adopting a full-time interval voxel coding scheme based on Gaussian distribution; step 2, constructing an FMF-Net network guided by an optical flow; step 3, designing a comprehensive loss function used for training the FMF-Net network; 4, performing supervised training on the FMF-Net network by using the public event data set; and step 5, inputting space-time voxels by using the trained FMF-Net network, and outputting a high-fidelity reconstructed image. According to the method, long-time motion clues and time domain dynamic features can be fully utilized, high-fidelity image reconstruction is realized under high-speed motion and sparse event input, the problems of loss of reconstruction details and poor consistency of an existing method in a dynamic scene are solved, and the structural similarity and visual quality of a reconstructed image are remarkably improved.

The invention discloses an adaptive feature alignment crop identification method based on intelligent remote sensing interpretation, and the method comprises the steps: taking DINOv3 as a backbone network, and optimizing the backbone network through a feature normalization adaptation layer and a resolution adaptive module; a time sequence feature alignment module TAFA is adopted to carry out cross-time alignment on crop features of remote sensing images in different growth periods; a geographic context gating attention module GCAI is adopted, and cross-geographic region adaptive feature fusion is realized based on multi-scale geographic context coding and a CNN-Transform bidirectional interaction mechanism; a crop semantic contrast loss function CSCL is adopted to perform intra-class compactness and inter-class separation degree contrast optimization on features of crops of the same class. The method realizes cross-growth-period stability and cross-region adaptation under a pure supervised training normal form, can distinguish remote sensing crop high-precision semantic segmentation of similar crops, covers agricultural remote sensing image processing scenes of different geographical landforms and whole growth periods of crops, and can be applied to the fields of modern agricultural fine management and the like.

The invention discloses a video pulse wave extraction method based on progressive self-supervised learning, and the method comprises the steps: 1, preprocessing a face video clip, and obtaining an original signal; 2, processing the original signal by using a signal enhancement strategy to obtain an enhanced signal, and forming a positive sample pair with the original signal; 3, constructing a deep learning model capable of adaptively quantifying the contribution weight of the global domain of interest; 4, decomposing a video pulse wave extraction task into a signal extraction sub-task and a noise suppression sub-task, constructing a two-stage time-frequency domain joint loss function, and performing progressive self-supervised training; and 5, calling the pre-trained model to recover a high-quality pulse wave signal from the to-be-detected original signal to realize heart rate monitoring. According to the method, the deep learning model can be self-supervised and trained without a true value signal label, video pulse wave extraction is achieved, meanwhile, complex environment noise is restrained, the generalization performance of the model in a real scene is improved, and development of a non-contact health monitoring technology is promoted.

The invention discloses a semi-supervised non-contact neonatal jaundice home intelligent early warning method based on image calibration, and the method comprises the steps: building a camera response function database, and building a color constancy depth model and a skin region segmentation network; designing a strong and weak enhancement strategy by using a large amount of label-free home data, constructing a time and context comparison learning module, and performing self-supervised training on a feature encoder; a small amount of labeled hospital data and a large amount of unlabeled family data are combined, and semi-supervised fine tuning is realized through pseudo label generation and supervised comparative learning; image sequences continuously uploaded by a user are input into the model, multi-day prediction of the bilirubin level is achieved, prediction uncertainty is quantified in combination with a Bayesian method, the risk boundary is dynamically adjusted, and personalized early warning is provided. According to the method, a user does not need to use a physical colorimetric card, color measurement errors caused by model differences of mobile equipment and variability of household illumination are inhibited from a data source, and the accuracy and robustness of subsequent jaundice assessment are improved; through a training framework combining self-supervised pre-training and semi-supervised fine tuning, the understanding ability and generalization performance of the model for the sequential characteristics of jaundice are improved.

The invention discloses an image watermarking method and system based on decoding guided by a self-supervised visual model. The method comprises the following steps: acquiring watermark information m and an original image I of each data sample in a training data set; for each data sample, preprocessing the watermark information m to generate a message tensor M, and performing degradation distortion processing on the original image I to generate a distortion guide heat map DGH, thereby obtaining a preprocessed training data set; and performing self-supervising training on the self-supervising visual model by using the preprocessed training data set to finally obtain the trained self-supervising visual model. The invention aims to solve the problem that the decoding robustness is poor when the mask prediction is inaccurate due to the fact that the existing printing-resistant shooting watermark technology excessively depends on the positioning mask, and the semantic comprehension capability of the depth vision model is utilized to replace the traditional geometric synchronization, so that the end-to-end robust decoding without correction is realized.

The small sample anti-migration prediction method suitable for the metallurgical process end point component under the zero expansion characteristic comprises the steps that smelting report data of a large sample steel grade and a small sample steel grade in the metallurgical process are collected to serve as source domain data and target domain data; adopting median to fill and restore abnormal values existing in the report data; dividing the source domain data and the target domain data into continuous feature variables and classification feature variables; inputting the continuous feature data and the classification feature data of the source domain data into a TabNet coding and decoding self-supervising network for self-supervising training, and inputting the target domain data into the trained self-supervising network for feature extraction and reconstruction; introducing an adversarial network, and gradually aligning feature distribution of a source domain and a target domain; and inputting the high-dimensional reconstruction features processed by the TabNet coding and decoding self-supervised network in the source domain into the deep table network model for training, and inputting the small sample steel grade features aligned by the adversarial network distribution into the trained deep table network model for transfer learning.

The invention provides a continuously constrained and difficulty guided passive domain adaptive three-dimensional medical image segmentation method. The method comprises the following steps: in a source domain pre-training stage, carrying out full-supervised training on a segmentation model by utilizing source domain annotation data; in the pseudo source domain image generation stage, a thought of combining coarse generation and fine generation is adopted, style migration is performed by using a frozen source domain pre-training segmentation model and target domain unlabeled data in coarse generation, and a target domain image is converted into a pseudo source domain image with a source domain style; and in the fine generation step, Fourier transform is utilized to remove artifacts and noise in the coarsely generated image. In the target domain adaptation stage, a pre-training segmentation model, a pseudo source domain image and a target domain image are utilized, and continuity constraint between slices and a difficult sample mining mechanism are fused to carry out an adaptation process from a source domain to a target domain. According to the method, under the condition that source domain data does not need to be accessed, the spatial context constraint and the difficult sample mining mechanism of the three-dimensional medical image are effectively fused.

The invention relates to an integrity verification and quality evaluation method before agricultural Internet of Things data uploading, and belongs to the technical field of artificial intelligence and block chains. The method comprises the following steps: collecting agricultural Internet of Things data and constructing a data set; constructing a complete feature vector by adopting a double-branch filling network; generating a quality label for each sample; an integrity verification and quality evaluation model based on a deep neural network is constructed, and sample-level overall quality scoring, sensor-level abnormal confidence and data fingerprint generation are realized; performing verification and evaluation by setting thresholds of three prediction results; performing supervised training on the model through a composite loss function; and performing integrity verification and quality evaluation on to-be-evaluated agricultural Internet of Things data by adopting the trained model, and realizing reliable chaining, secure storage and trusted application of the agricultural Internet of Things data to a block chain for data samples which pass the integrity verification and have the overall quality score reaching the standard. The prediction precision can be improved.

The invention discloses a defect detection method and device for medical consumables, equipment and a medium. The method comprises the following steps: acquiring multi-modal data of the defect-free medical consumables as normal sample data; the multi-modal data comprises a visible light image, an infrared image and three-dimensional point cloud information; interference is added to the normal sample data to generate pseudo-defect sample data, and a multi-modal training data pair is determined according to the normal sample data and the pseudo-defect sample data; performing feature extraction on the multi-modal training data to obtain multi-modal feature information, and performing feature alignment on the multi-modal feature information through comparative learning to obtain target feature information; inputting the target feature information into a defect detection model for self-supervised training, and performing defect detection on the to-be-detected medical consumables by using the trained defect detection model; and the defect detection model performs defect detection based on the difference between the target feature information and the reconstruction result thereof. According to the scheme, the defect detection coverage range can be expanded while the defect labeling dependence is reduced.

The invention discloses an IMU (Inertial Measurement Unit) open-loop noise reduction method based on a lightweight adaptive neural network. According to the method, a noise reduction network is constructed through lightweight depth separable expansion convolution, efficient noise reduction is carried out on IMU original data, and the parameter quantity and the calculation complexity are remarkably reduced. Static, constant-speed, acceleration and other states are recognized in real time through the motion state sensing module, network depth, channel number and other configurations are dynamically adjusted, and intelligent balance of precision and efficiency is achieved. Complete self-supervised training is adopted, a learning target is constructed by utilizing physical constraints of IMU signals, and external true value data or manual annotation is not needed. According to the self-adaption and self-supervision combined method, high-performance noise suppression and deviation compensation can be realized only by depending on a single IMU, and the noise-reduced signal can be directly used for subsequent processing such as open-loop integration. The method is suitable for embedded equipment sensitive to cost and power consumption, such as micro unmanned aerial vehicles and wearable equipment, and provides real-time and low-power-consumption IMU data purification capacity.

The embodiment of the invention provides a load data cleaning method and device for a novel power system power distribution network. The method comprises the following steps: obtaining multivariate time sequence load data; the method comprises the following steps: constructing a model fusing hierarchical expansion convolution and a multi-branch attention Transform; performing single-step time sequence prediction on the target monitoring column, and outputting a predicted value and a residual error; carrying out combined judgment on missing values and abnormal values; constructing a conditional fraction diffusion model; introducing self-supervised training and two-dimensional feature fusion; in the conditional diffusion reverse process, random noise is gradually denoised into a filling value conforming to real distribution. According to the method, multi-source load data can be fused, through combination of time sequence feature analysis and deep learning, missing detection, anomaly recognition and adaptive judgment of novel power system power distribution network user load data containing charging stations and distributed photovoltaics are realized, and finally missing data and abnormal data are filled through a conditional score diffusion model. And the completeness and quality of the user load data of the novel active power distribution network are improved.

The invention discloses a wind turbine blade internal damage diagnosis method based on a sound field graph neural network, and belongs to the technical field of wind turbine blade state detection, and the method comprises the steps: deploying a microphone array in a cabin, and collecting an acoustic signal when a blade rotates; constructing a space sound field graph structure by taking the microphone as a node and the sound wave propagation path as an edge; extracting nonlinear acoustic features by using a physical constraint graph neural network PC-GNN; generating a damage embedding vector based on self-supervised training contrast learning; and outputting a damage probability thermodynamic diagram and positioning information. According to the method, a directional microphone array is deployed in a cabin, and a sound wave propagation space diagram structure is constructed; designing a physical constraint graph neural network PC-GNN, and embedding an acoustic wave equation as a regularization item; the problem of scarcity of damaged samples is solved by adopting self-supervised contrast learning; and finally outputting a positioning thermodynamic diagram of the internal damage of the blade.

The invention discloses a passive domain medical image segmentation method based on multi-branch collaborative calibration and reliable weighted consistency, and the method comprises the steps: 1, carrying out the supervised training of U-Net based on source domain labeling data, obtaining a source domain pre-training model, building a network MCC-RWC based on the pre-training model, initializing the network hyper-parameters, and carrying out the preprocessing of the data; 2, enhancing an unlabeled image of a target domain; 3, inputting the enhanced image into a multi-branch collaborative calibration module to obtain an average probability prediction map; 4, performing calculation based on the average probability prediction map to obtain a pseudo label and a reliability mask; 5, realizing self-supervised training of a target domain by using a pseudo label and a reliability mask and adopting three-time differential forward propagation; 6, a total loss function is constructed, and target domain self-adaption is completed; 7, repeating the steps 2-6 to carry out iterative training; and 8, segmenting the target domain image by using the trained model and outputting a final segmentation result. According to the invention, the segmentation precision and generalization ability of the target domain can be effectively improved.

The invention discloses a multi-span question and answer knowledge perception method and system based on big and small model collaboration, and the method comprises the following steps: S1, executing knowledge prompt generation and chain logic reasoning through a big language model, and carrying out the semantic extension and knowledge reconstruction of an input question, obtaining a knowledge perception representation containing a potential answer entity, a reasoning link and a semantic dependency relationship; s2, based on the obtained knowledge perception representation, calculating a knowledge coverage rate between a standard answer in the training sample and the knowledge perception representation; screening the training samples according to the knowledge coverage rate, and constructing a training set with consistent coverage for performing supervised training of coverage matching constraint on the small model; and S3, based on the knowledge perception representation in the S1 and the small model obtained by training in the S2, fusing the knowledge perception representation and the original question, inputting the fused knowledge perception representation and original question into the small model, and executing collaborative reasoning to realize multi-span answer positioning and generation. According to the method, unification of knowledge perception and structured reasoning is realized, and the accuracy and generalization of multi-answer extraction are improved.

The invention relates to the technical field of electronic sensing and mode recognition, and discloses a multi-gas classification method and system based on an electronic nose and a medium, and the method comprises the following steps: collecting an original time sequence response signal of target gas; down-sampling the original time sequence response signal according to a predetermined scaling factor; preprocessing the sampled time sequence response signal, and segmenting the preprocessed time sequence response signal into a plurality of samples according to a predetermined window strategy; based on the constructed sample, a channel subset is selected from multiple channels through a channel selection module, and the channel selection module comprises one of a package-based single-channel evaluation method and a channel attention-based end-to-end weight extraction method; inputting a sample corresponding to the selected channel subset into a deep convolutional neural network for supervised training to obtain a gas classification model; and classifying the target gas by using the trained gas classification model to obtain a classification result. According to the method, the dimension and redundancy of the data can be reduced, and the recognition precision, the calculation efficiency and the robustness are improved.