3360 results about "Classification methods" patented technology

The invention provides an engineering construction defect automatic detection and classification method based on deep learning, and the method comprises the steps: obtaining a welding seam surface image through the shooting of an unmanned plane, and carrying out the denoising and illumination normalization processing of the welding seam surface image, and obtaining a standardized image; welding seam surface texture features are extracted from the standardized image, a convolutional neural network is adopted to analyze the spatial distribution characteristics of textures, and vectorization processing is carried out to obtain texture feature vectors; segmenting a weld surface corresponding to abnormal region distribution by adopting a region growing algorithm, and analyzing pore and weld discontinuity in combination with the texture feature vector to obtain a defect candidate region; performing threshold division on the sizes and the numbers of the defects according to the defect types and the feature vectors of the candidate regions to obtain a severity grading result of each type of defects; and severity features are extracted from a grading result, and a Bayesian network is adopted to fuse texture feature vectors and defect type labels to obtain a welding quality evaluation score.

The invention discloses an automatic label labeling and classifying method and system oriented to unstructured system documents, and relates to the technical field of artificial intelligence. The method comprises the steps that semantic structure pre-analysis is conducted on an original system text, and a system semantic structure tree is constructed; establishing a system semantic enhancement vector space based on the semantic units and the logic relationship thereof; performing semantic deconstruction on the preset tag and extracting a feature vector; realizing cross-space semantic matching of the document and the tag through a system semantic attention mechanism; a confidence evaluation module is introduced to screen high-confidence labels from the three dimensions of structural integrity, coverage and logic consistency; and outputting a final label and a score through semantic conflict detection and resolution. According to the method, the problems that in the prior art, unstructured system text labeling accuracy is low and large-scale labeling samples are dependent on polysemy ambiguity, high context dependency, complex semantic structure and the like are solved, and labeling accuracy and robustness are remarkably improved.

The invention relates to an RAG-based voucher classification method, a medium and equipment, and the method comprises the steps: receiving digital image data of a to-be-classified voucher, carrying out the multi-modal optical character recognition processing to generate a structured OCR result, extracting a text semantic feature vector and a visual layout feature vector based on the structured OCR result, carrying out the fusion of the text semantic feature vector and the visual layout feature vector to generate a multi-modal query vector, and carrying out the classification of the to-be-classified voucher. Similar samples and semantic similarity scores and category metadata thereof are obtained through approximate nearest neighbor retrieval, after an initial candidate category list is generated, key field values are extracted for each candidate category, evidence credibility scores are calculated, comprehensive confidence scores are generated by fusing the semantic similarity scores and the evidence credibility scores, reordering is conducted, and a candidate category list is obtained. And finally, selecting a classification decision path according to the score distribution, and outputting a classification result and an interpretability report. The accuracy and robustness of voucher classification are effectively improved, and complex voucher scenes with changeable formats and fuzzy semantics can be processed; and the interpretability and reliability of the classification decision are enhanced.

The invention relates to a flying dust monitoring data processing and classifying method based on multi-source sensing fusion, and the method specifically comprises the following steps: firstly, deploying multi-source flying dust monitoring sensor nodes in a target region to collect data, carrying out the marking, and generating a data set; performing continuous wavelet transform on the acquired data, extracting a wavelet energy spectrum and a Shannon entropy, and splicing to obtain an enhanced feature tensor; secondly, through a two-stage fusion and coding strategy, frequency band energy features are extracted through wavelet packet decomposition, multi-channel cross-correlation, statistical moment and ratio features are calculated to form time sequence mode coding features, and multi-source heterogeneous feature fusion is achieved in combination with a local time sequence feature matrix; then constructing a deep learning model containing a multi-scale time sequence feature extraction and dynamic fusion module, and inputting a fusion feature matrix for training; and finally, inputting the preprocessed new monitoring data into the trained model, and outputting a dust source and pollution level classification result. The dust monitoring data classification accuracy and the dust source identification precision can be effectively improved.

The invention discloses a music stave sentiment classification method and system based on multi-level distillation, and belongs to the technical field of music information processing. The method comprises the steps of firstly collecting music stave data and converting the data into stave data vectors, then performing feature extraction by using a long short-term memory network, then constructing a teacher network and a student network for knowledge distillation, and realizing multi-level knowledge transmission through temperature scaling, KL divergence loss and mask feature distillation. And finally, training a lightweight classification model to complete sentiment classification. The knowledge distillation technology is creatively applied to staff sentiment classification, the classification accuracy is effectively improved through an online multi-level distillation mode, and the technical problems that a traditional method lacks semantic information and a self-supervised model is not suitable for sentiment tasks are solved. The method has the main advantages of high classification precision, light model weight, capability of effectively capturing music emotion features and the like.

The invention relates to the technical field of multi-sensor identification, in particular to a coal gangue detection and classification method and system based on multi-sensor fusion discrimination, and the method comprises the following steps: obtaining visible light and infrared images, calculating brightness and intensity judgment feature conditions, executing edge detection to extract gray segments, and fusing textures and a thermal field to generate a vector set. And performing clustering analysis to finish classification judgment, and outputting a coal gangue detection classification result. According to the method, a precise trigger mechanism is established through brightness and thermal radiation double-feature screening, boundary recognition sensitivity is enhanced through gray abrupt change point division, salient region extraction capacity is enhanced through weighted fusion of texture energy and gray gradient, and a cross-modal consistency feature group is constructed through combination of two-dimensional vector construction and similarity screening. The recognition expression integrity is improved, static threshold classification is replaced by vector distribution clustering, accurate mapping and classification decision making of material attributes in a complex scene are achieved, and the stability and the recognition rate of a coal gangue detection result are guaranteed.

The invention discloses an open set domain adaptive image classification method based on a difference prompt learning technology of a pre-training vision-language model. According to the method, high-quality pseudo-open class images are generated, and de-noising text embedding and de-noising visual embedding are obtained by using a differential prompt learning technology, so that class characteristics of a source domain, a target domain and pseudo-open class samples are effectively extracted, and irrelevant noise is inhibited. According to the method, a vision-text comparison loss mechanism, a triple distance comparison loss mechanism and a negative sample penalty mechanism are further designed, a known category and an unknown category are effectively distinguished in a feature space, and the semantic alignment capability of cross-domain similar samples is enhanced. The method can significantly improve the classification accuracy and model robustness in an open set domain adaptation task, has the advantages of simple structure, high calculation efficiency, good generalization performance and the like, and is suitable for image classification, cross-domain transfer learning and other related application scenes.

The invention discloses a quantum fuzzy neural network adaptive to high-dimensional input and a classification method, and relates to the field of quantum calculation and fuzzy neural networks and the field of computer vision. The network input layer receives high-dimensional data, amplitude coding, forward and reverse enhanced chain entanglement layer, parameterized quantum transformation and fuzzy set mapping are carried out through a quantum fuzzy feature extraction module, and dynamic dimension fuzzy features are output; high-dimensional neural features are extracted through a DNN feature extraction module to adapt to quantum fuzzy feature dimensions; dynamically distributing the weights of the quantum fuzzy features and the classic neural features through an adaptive feature fusion module; and carrying out Softmax classification on the fusion features through a classifier, and outputting a category probability. According to the method, the high-dimensional data coding efficiency can be effectively improved, the complex fuzzy logic relation learning capability of the quantum part and the quantum state correlation stability are enhanced, the uncertainty of the data is represented, and accurate classification of high-dimensional uncertainty images is realized while noise interference is reduced.

The invention discloses a crop disease multi-modal diagnosis and classification method and system, and relates to the technical field of data processing, and the method comprises the steps: collecting disease multi-modal data, and carrying out the preprocessing; constructing a disease multi-modal diagnosis model, respectively inputting images and text sequences in the multi-modal data into a visual feature extraction subnet and a text feature coding subnet of the disease multi-modal diagnosis model, and extracting global visual features, sequence state features and global text features; calculating a global matching score based on the global visual features and the global text features, executing fine-grained local interaction on the global visual features and the sequence state features, introducing a category channel attention mechanism to correct the features obtained by interaction, and weighting to generate multi-modal fusion features; and constructing a loss function based on the global visual features, the global text features and the multi-modal fusion features, training a disease multi-modal diagnosis model, inputting to-be-classified data into the disease multi-modal diagnosis model, and outputting a classification result.

The invention discloses a rolling bearing vibration signal multi-mode fault classification method. The method comprises the following steps: collecting a vibration time sequence signal of a rolling bearing; the vibration time sequence signals are input into a time sequence branch network and a space branch network in parallel, and the time sequence branch network extracts time sequence dependence characteristics of the signals through a one-dimensional convolutional neural network and a bidirectional gating circulation unit; the spatial branch network converts the vibration time sequence signal into a Markov transform field image, and extracts spatial structure features of the image by using a two-dimensional convolutional neural network and a window Transform-based visual network; performing bidirectional interaction and weighted fusion on the time sequence features and the spatial features through a cross-modal attention mechanism to obtain fusion features; and inputting the fusion features into a classifier, and outputting a fault classification result of the rolling bearing. According to the method, the problems that a traditional single-mode fault diagnosis model is insufficient in adaptability to complex working conditions, multi-mode feature fusion is insufficient, and the generalization ability is weak due to model structure redundancy are solved.

The invention discloses an incomplete multi-view multi-label data classification method based on semantic enhancement and pseudo-label uncertainty perception, and the method comprises the steps: employing a dual-channel feature extraction and decoupling module to obtain the shared semantic representation and specific representation of each view in each sample for a constructed incomplete multi-view multi-label data classification network model; performing cross-view fusion on the shared semantic characterization and the specific characterization, obtaining a unified shared characterization and a unified specific characterization corresponding to each sample, performing feature fusion, obtaining a fusion characterization of each sample, inputting the fusion characterization of the sample output by the dual-channel feature extraction and decoupling module into a classifier for multi-label prediction, and performing multi-label prediction on the fusion characterization of the sample. Therefore, a multi-label classification prediction result is obtained, and model training is carried out based on a total contrast learning loss function and a joint supervision classification loss function. According to the method, the classification performance and the model robustness on incomplete multi-view multi-label data are remarkably improved through training learning under the guidance of semantic enhancement and uncertainty.

The invention relates to a few-sample image classification method based on hyperbolic space image-text local feature alignment, and belongs to the technical field of image recognition and artificial intelligence, and the method comprises the steps: generating word-level attribute description for a support set image through employing a multi-mode large language model; encoding the image and the text by adopting a vision-language model; constructing a hyperbolic local feature alignment module in a hyperbolic space, screening most relevant image local features for text local features by calculating hyperbolic cosine similarity, and fusing by using hyperbolic weighted average; designing a hyperbolic cross attention module, and aggregating key information from the multi-modal local features of the support set to construct a category prototype by taking query image aggregation features as guidance; and finally performing classification based on the hyperbolic geodesic distance. According to the method, the hierarchical modeling capability of the hyperbolic space and the semantic priori knowledge of the large language model are fully utilized, fine-grained multi-modal feature alignment is realized, and the small sample image classification performance is improved.

The invention discloses a strip steel surface defect detection and classification method based on a graph neural network, and the method comprises the steps: segmenting a strip steel surface image into different scale sub-blocks, and extracting the node features of each scale sub-block; constructing graph structures with different scales according to the similarity of the node features; constructing a stacked multi-scale image attention network model, taking the image structures of different scales as input for training, inputting the image structures of different scales into the trained stacked multi-scale image attention network model, performing feature extraction on nodes in the image structures of different scales, and performing feature extraction on the nodes of the image structures of different scales; aggregating and updating node features in the different scale graph structures through a graph attention mechanism, and obtaining a feature matrix according to the updated node features in the different scale graph structures; and paving the obtained feature matrix into a one-dimensional vector, inputting the one-dimensional vector into a full connection layer, carrying out nonlinear operation, outputting the probability of each category through a softmax function, and taking the category with the maximum output probability as the category of the strip steel surface defects.

The invention relates to a semi-supervised small sample image classification method and system in a cross-domain scene, and relates to the technical field of deep learning and machine learning, and the system comprises a pre-training feature extractor and a self-supervised encoder; using a feature extractor and a self-supervised encoder to respectively obtain initial features and internal features, and fusing the initial features and the internal features to obtain fused features; generating initial pseudo labels through a clustering algorithm, and screening the initial pseudo labels according to a clustering validity index to obtain high-quality pseudo labels; constructing a mixed loss function, and updating the feature extractor on the mixed training sample by using the mixed loss function; and inputting a to-be-classified sample into the updated feature extractor, and outputting the category of the to-be-classified sample. According to the method, the technical problems of unstable model training and weak generalization ability caused by low pseudo label quality when small sample learning tasks with significant domain differences are processed in the prior art are solved.

The invention discloses a small sample remote sensing image classification method based on hierarchical spatial structure learning. The method comprises the following steps: firstly, extracting multi-scale features of a remote sensing image by using a ViT (Visual Transform) model, and capturing rich semantic information and spatial structure relationships in the image; secondly, constructing a graph structure based on spatial adjacency and attention weight to model a structured relationship between samples, and encoding graph node features through a graph convolutional network (GCN) so as to enhance the discrimination ability of the features in a structural semantic space; thirdly, a residual enhancement mechanism is introduced to fuse global semantic information, and the discrimination capability of graph embedding is improved; then, based on the structural similarity between the support set and the query set, performing classification decision, and realizing accurate classification under a small sample condition; and finally, carrying out joint optimization on the whole model by adopting a training strategy of a small sample meta learning task and a supervision loss function.

The invention discloses a large model-based text classification method, which comprises the following steps of: 1, setting a task semantic constraint rule, and writing a fixed cue word template; 2, sorting classification labels and adding semantic description information; 3, semantic similarity clustering and confidence allocation are executed, and a confidence multi-granularity label prefix tree is generated; 4, dynamic constraint autoregressive decoding is executed in combination with the confidence multi-granularity label prefix tree and the large language model, and a target classification label is output; 5, constructing a BERT auxiliary discrimination module to carry out confidence mask constraint; step 6, constructing a decoding temperature control parameter and a semantic bias item to perform semantic guidance and category distinguishing control; 7, performing uncertainty self-calibration on the autoregression decoding process of the large language model; and 8, constructing semantic consistency loss between the large language model and the BERT auxiliary discrimination module. According to the method, the stability and the classification efficiency of text classification in a small sample scene are improved.

The invention provides a bone tumor fine-grained classification model training and classification method and device, and the method comprises the steps: constructing a multi-modal positive sample containing global / local positive lateral X-ray images, lesion attributes and patient information, and removing a false negative part in combination with text semantic similarity to construct a high-quality negative sample; global / local image features are extracted through double image encoders, lesion attribute keywords are converted into'entity-translation-existence 'triples based on a medical knowledge base, and basic information of a patient and global / local semantic features of lesion attributes are extracted through a text encoder; infoNCE contrast loss is constructed for global images and global semantics based on contrast learning, global image-text feature alignment and local image-text feature alignment are realized in combination with a local mutual information loss and classification loss training model calculated based on a DV variational formula, medical term semantics are deeply combined, the training stability is improved, and the training efficiency is improved. The accuracy and robustness of bone tumor subtype classification are remarkably improved, and reliable support is provided for clinical precise diagnosis.

The invention provides a concrete crack automatic identification and classification method based on machine learning, and belongs to the technical field of wind tunnels, and the method comprises the steps: collecting a visible light image and a thermal image of a concrete structure surface, and extracting an initial crack contour; based on a plant growth simulation algorithm, the growth direction and speed are determined from a seed point according to local gradient intensity and texture feature directivity; a concrete crack intelligent analysis model is adopted to fuse crack path temperature and stress distribution information, attention weight is dynamically adjusted according to crack characteristics to output an accurate classification result and a severity degree evaluation value, and sound wave propagation speed data is combined to accurately estimate crack depth so as to realize accurate classification of fine and coarse cracks. An evaluation system based on a comprehensive risk coefficient is established to scientifically divide low, medium and high risk states, and the technical problem of inaccurate classification caused by insufficient concrete crack detection precision is solved.

The invention provides an aspect-level sentiment classification method and system based on multi-modal alignment and reflection enhancement, and relates to the technical field of artificial intelligence, and the method comprises the steps: employing a pre-trained teacher model to enable an original image to be combined with a knowledge guide prompt to generate text description; splicing the description, the original text, the target aspect word and the structured reasoning prompt template into a multi-modal aligned enhanced input sequence, and inputting the multi-modal aligned enhanced input sequence into a pre-trained student model to obtain a predicted emotion tag; aiming at a sample with a prediction error, generating an reflection-correction inference chain by utilizing a teacher model, and constructing an reflection enhancement data set; and finally, combining the original data set and the reflection enhancement data set, and performing supervised fine tuning training on the student model. According to the method, the semantic alignment and self-error correction capability of the model is enhanced, so that the sentiment classification accuracy is improved.

The invention discloses an urban reserve land intelligent classification method and system based on spatial data fusion, and relates to the technical field of data processing, and the method comprises the steps: carrying out the formalized expression of a land use rule of a preset urban region, and constructing a land domain knowledge graph; collecting multi-source spatial data, performing feature extraction and alignment, and generating a high-dimensional feature vector of each plot; and performing hybrid classification evaluation based on the high-dimensional feature vector of each land parcel, synchronously outputting land use types, development potential levels and land reference value intervals of the land parcels, and calling a land domain knowledge graph to perform compliance verification and correction. The technical problems that an existing urban reserve land classification method depends on artificial experience and lacks accurate and multi-dimensional data fusion and compliance check, so that the accuracy and practicability of a classification result are low are solved, and the purpose that the classification accuracy and practicability of the urban reserve land are improved through multi-source spatial data fusion and an intelligent classification framework is achieved. And the accuracy and practicability of urban reserve land classification are improved.

The invention relates to the field of twitter recognition, in particular to a twitter classification method based on harmonic enhancement and time-frequency semantic joint modeling, which comprises the following steps: collecting twitter samples and carrying out noise reduction and standardized preprocessing, carrying out multi-scale convolution operation on Mel spectrograms by utilizing a layered acoustic encoder, extracting time-frequency features in combination with a channel attention mechanism, and classifying twitter classification results. The method comprises the following steps of: generating adaptive position codes through a dynamic time-frequency joint coding module, carrying out time-frequency mode modeling by combining a global-local interaction mechanism, introducing a semantic fusion module which comprises a frequency band pyramid unit, a harmonic enhancement unit and a time-frequency gating unit, realizing dynamic weighted fusion of multi-layer features, and carrying out time-frequency mode modeling through a global-local interaction mechanism. And inputting the fusion features into a classification layer, training a network by adopting a cross entropy loss function and a gradient descent algorithm, and outputting bird categories through a full connection layer, thereby solving the key problems of insufficient description of a non-stationary time-frequency mode, insufficient modeling of a harmonic structure, reduction of recognition performance in a complex noise environment and the like in the prior art.

The invention discloses an encrypted traffic classification method based on Transform, and belongs to the technical field of network security and encrypted traffic analysis. In order to solve the problems that load content in a novel encryption protocol (such as TLS 1.3 and VPN) is encrypted, structural disturbance is complex, category distribution is unbalanced and the like, the invention provides a dual-phase Transform framework (DPFT) with pre-training and structure optimization fine tuning. According to the framework, two self-supervision tasks of masked burst prediction (MBP) and burst structure discrimination (BSDT) are introduced in a pre-training stage, and deep data packet representation is learned from unlabeled traffic, so that the deep data packet representation is learned from the unlabeled traffic; in the fine tuning stage, dynamic weighting of word embedding and position embedding, multi-head attention pooling and Focus Loss are adopted, so that the modeling capability of the model on an encrypted traffic complex structure is enhanced, and the recognition sensitivity on minority class samples is improved. Experimental results show that the method disclosed by the invention is obviously superior to the existing method on various encrypted traffic data sets (including TLS 1.3, VPN, malicious traffic and the like), and achieves leading performance on classification accuracy and macro average F1 index.

The invention belongs to the technical field of multi-modal processing, and particularly relates to a multi-modal fusion and classification method based on alignment and calibration. The method comprises the following steps of: firstly, protecting complementary information of low-credible samples while enhancing the consistency of high-credible samples through cross-modal alignment perceived by confidence; further innovatively introducing confidence-constrained intra-modal modulo length calibration, directly coding the dynamics into the feature correlation of samples of the same category in the modal, and enhancing the effectiveness of the feature modulo length as the prediction credibility. According to the'alignment-calibration 'cooperation mechanism, the model does not need a complex dynamic weighting network, better generalization performance and higher noise robustness can be achieved only through simple static fusion, and a new path giving consideration to high efficiency, interpretability and stability is provided for multi-modal learning.

The invention belongs to the technical field of artificial intelligence and satellite remote sensing crossing, and particularly relates to a cross-scale space-time fusion ground feature classification method based on a double-branch architecture, and the method comprises the steps: preprocessing a remote sensing image: obtaining a high-resolution optical image and a multi-stage medium-resolution time sequence image, carrying out the processing, generating time sequence data, carrying out the marking, and segmenting a data set; constructing a double-branch network, wherein the network comprises space and time feature extraction branches and fusion and decoding modules; training an optimization model by using a weight optimizer and a loss function for dynamically adjusting category weights; and post-processing the classification result, and outputting vector data. According to the method, heterogeneous information fusion is realized, the problems of same object and different spectrums and the like are solved, the feature utilization rate and minority class recognition precision are improved, end-to-end design is convenient, the generalization ability is high, and reasoning is fast.

The invention provides a collaborative classification method and system fusing advantages of large and small models, and the method comprises the steps: inputting to-be-classified data into a trained zero-sample classification small model, and outputting candidate label domain screening information and an initial classification result; inputting the to-be-classified data, the candidate label domain screening information and the initial classification result into the trained large model, and outputting a final classification result; the training process of the small model comprises the following steps: inputting training data into the zero sample classification small model to obtain a preliminary prediction result; screening and obtaining pseudo label data based on an active learning strategy; checking and re-marking the pseudo-label data by using the large model to obtain a modified pseudo-label data set; and carrying out iterative training on the zero sample classification small model by utilizing the modified pseudo label data set. The method has the advantages that a small model has classification performance close to that of a large model while keeping lightweight calculation characteristics; the calculation burden of a large model is reduced, and the accuracy of classification decision is improved.

The invention relates to the technical field of text processing, in particular to a text classification method and system based on semantic analysis, and the method comprises the following steps: segmenting semantic units, constructing a direction change sequence, positioning mutation nodes, generating a consistency section, forming a convergence section, and outputting a classification result. According to the method, a continuous change sequence is formed by constructing a semantic embedding vector and calculating a direction difference, a semantic mutation point can be anchored and divided into sections by combining mutation intensity identification and local jump tracking, and a semantic closed structure and a convergence section are extracted by means of context direction consistency judgment and generic label comparison; precise recognition of a semantic relation chain is realized, semantic jump and conflict starting points can be dynamically sensed, the semantic boundary recognition capability is improved, and the understanding and classification capability of a model on semantic attribution in a complex context is enhanced on the premise of not depending on a fixed dictionary and shallow statistics. The problems that a traditional model is slow in response to an abrupt change structure and weak in semantic convergence recognition are effectively solved.

The invention relates to the technical field of big data management, in particular to a bioinformatics data classification method and system based on big data, and the method comprises the following steps: obtaining time sequence recognition trend reversal and positioning fragments, extracting recognition difference positions inside and outside a frequency band data division region, screening samples with consistent features, and rearranging path labels; connecting nodes are cut off to generate fracture indexes, and label states are updated and written into sample fields to form a classification result set. According to the method, a labeling area is constructed by extracting trend inversion points in a time sequence, sample fragments are divided by combining data fluctuation positions in a disturbance frequency band, label numbers are arranged according to the fluctuation sequence of samples in a path, a corresponding sequence of a label chain connection relation and the sample positions is established, and label section boundaries are positioned and limited by fracture nodes. And the updated label state is synchronously written into a sample field, and the path label is bound according to a chain sequence, so that the sample identifier is kept coherent in sequence change, and the continuous coverage capability of the path information in classified output is improved.

The invention relates to the field of computer vision, in particular to an RGB image classification method based on causal anti-fact reasoning, and the method comprises the steps: constructing an RGB image classification model; obtaining a to-be-classified image, and inputting the image into the trained RGB image classification model to obtain an image classification result; the RGB image classification model comprises a visual Transform network module, a semantic feature extraction module, a comparative learning module and a prediction head module; based on a causal inference theory, anti-fact intervention is carried out on a classification process and contrast loss is introduced, a network can be guided to separate semantic features capable of representing image essence, and the accuracy and stability of a classification model in an actual application scene are further improved.

The invention relates to the technical field of deep learning model training, solves the technical problem that gradient pointing is inaccurate when model parameters are updated in a traditional SAM algorithm, and particularly relates to an image classification method based on sharpness perception minimization. A gradient direction correction mechanism is introduced for an image classification task, so that the stability in the optimization process and the generalization ability of the model on image recognition test data are remarkably improved. According to the method, the gradient disturbed by the SAM algorithm is corrected by using the feature vector, and the abnormal component of the gradient in the feature vector direction is effectively reduced, so that model parameter updating is prevented from pointing to a sharp region of a loss function. The correction of the optimized path significantly improves the accuracy of gradient updating in the training process of the image classification model, so that the model can learn more discriminative visual feature representation, and finally the classification accuracy and the generalization performance of the model on a test set are improved.

The invention belongs to the computer vision technology, and particularly relates to a small sample multi-modal prototype learning-oriented remote sensing scene classification method, which comprises the following steps of: acquiring RGB (Red, Green and Blue) images with category labels and text prompts of the RGB images as a support set; establishing a text prototype, an RGB prototype and a hyperspectral prototype of each category according to the support set; and extracting to-be-classified query set image features by using a pre-trained CLIP image encoder, calculating cosine similarities between the query set image features and the text prototype, the RGB prototype and the hyperspectral prototype of each category of the support set, taking the cosine similarities as input of a multi-layer perceptron, and obtaining the category of the to-be-classified RGB image through classification of the multi-layer perceptron. High-precision and high-robustness remote sensing scene classification is realized under the small sample condition, only prototype and similarity calculation is needed in the reasoning stage, and deployment and expansion are easy.