1747 results about "Negative sample" patented technology

The invention discloses a multi-label electrocardiogram classification method based on self-supervised pre-training and multi-modal semantic alignment, which belongs to the technical field of artificial intelligence, and comprises the following steps: realizing self-supervised pre-training of unlabeled data through a single-modal contrast enhancement network, generating global and local contrast views by adopting a multi-scale random cutting strategy, and classifying the global and local contrast views in a multi-scale random cutting mode; in combination with a teacher-student network architecture, the potential invariance features of the ECG signals are learned while negative sample dependence is avoided, the problem of annotation data scarcity is effectively relieved, and the feature robustness is improved. A multi-modal fusion mechanism based on label semantic guidance is provided, a time domain signal and a frequency domain time-frequency graph are mapped to a unified semantic space through fine-grained semantic alignment, local feature enhancement and cross-modal complementary information fusion are realized by using a cross attention mechanism, and the problem of semantic difference caused by modal heterogeneity in a traditional method is overcome. A multi-label comparison loss function based on a disease co-occurrence relation is proposed, a category discrimination boundary is dynamically optimized by modeling a label co-occurrence probability, the feature separability of a tail category is improved while the head category discrimination ability is enhanced, and the problem of sample category imbalance in a multi-label scene is remarkably relieved.

The invention provides a medical multi-hop text question and answer retrieval enhancement method based on heterogeneous knowledge. The method comprises the following steps: firstly, constructing a uniform heterogeneous graph structure based on a medical knowledge graph of a medical document, and establishing a semantic bridge through an entity-document mapping relationship; performing semantic decomposition on a complex medical problem input by a user by utilizing the large model, and iteratively generating a series of mutually independent atomic queries; searching a reasoning path in the entity sub-graph of the heterogeneous graph, and calculating a path score by fusing the weighted combination of the entity association text similarity, the entity matching degree and the path edge weight; training a retriever by adopting a marginal sorting loss function, and optimizing a retrieval effect through positive and negative sample comparative learning; and finally, calling a large model to convert the reasoning path with the highest score into a text, and extracting a document fragment corresponding to a path node. According to the method, the problems that an existing retrieval enhancement technology is insufficient in complex problem processing capacity, poor in reasoning interpretability and the like are effectively solved, and high-accuracy medical questions and answers are achieved.

The invention discloses a mineral resource intelligent prediction method and system based on multi-source heterogeneous data fusion and deep learning, and the method comprises the steps: collecting and preprocessing multi-source heterogeneous data, and carrying out the standardization processing to form a structured data set; multi-source heterogeneous data fusion: realizing data layer space registration and feature layer weight dynamic allocation through an attention mechanism multi-modal fusion module, and outputting a high-dimensional metallogenic feature vector; constructing a CNN-LSTM mixed deep learning model and completing initialization training, and outputting an initial mineralization probability graph; and establishing a dynamic updating engine, performing model increment training based on transfer learning, correcting the mineralization probability through positive and negative sample reinforcement learning in combination with a newly added data type, and outputting a time sequence dynamic mineralization probability graph. According to the method, mineralization probability dynamic evaluation and risk quantitative updating are realized, the prediction precision and the model updating efficiency are improved, the method is adaptive to a multi-stage exploration scene, and accurate real-time support is provided for exploration decision making.

The invention discloses a battery fault identification method based on probability labels and identification feature learning. The method is suitable for modeling and discrimination of various fault states in small sample and weak label scenes. The method comprises the following steps: firstly, acquiring key parameters such as voltage, current and temperature in an operation process of a battery system, and constructing standardized time sequence characteristic data; secondly, three types of pseudo labels are generated based on multi-source information such as alarm time difference, prediction residual error and mahalanobis distance, and a unified abnormal probability label is obtained through weighted fusion; constructing positive and negative sample pairs according to the difference between the tags, and introducing difficult samples with similar features but large tag difference to enhance the discrimination ability of the model; then constructing a twin neural network structure composed of shared parameter sub-networks, inputting positive and negative sample pairs for comparative learning, and extracting low-dimensional embedding features with clustering and distinguishability; and finally, through calculating a space distance between a new sample embedding vector and a known fault type, identification of a current fault type and evaluation of an abnormal degree are realized.

The invention discloses a pathological feature recognition and negative elimination method based on microscopic imaging. The method comprises the following steps: S1, collecting a pathological section image and digitally generating original microscopic image data; s2, preprocessing the original microscopic image; s3, constructing a pathological image recognition network fusing converter coding and a gating dynamic receptive field mechanism, and outputting pathological feature vectors; s4, performing context modeling through an attention guidance and category perception decoder, and outputting an image classification result; s5, constructing a discriminant boundary separation model based on positive and negative sample embedding, and performing negative exclusion judgment; s6, performing confidence coefficient weighted evaluation in combination with the uncertainty and the boundary distance, setting a dynamic threshold value, and screening out low-credibility samples; and S7, coding the classification result and the negative label into structured data, and sending the structured data to a diagnosis auxiliary system. According to the method, multi-scale modeling and a negative screening mechanism are fused, and intelligent recognition and credible diagnosis output of the pathological image are realized.

The invention discloses a cross-modal heterogeneous data processing method and system. The method comprises the following steps: acquiring multi-source heterogeneous data including a vibration signal, a temperature signal, an oil spectral signal and the like, and forming a cross-modal alignment index sequence through event detection and time alignment processing; constructing a heterogeneous graph structure based on the aligned index sequence, and determining an edge weight according to correlation between modal fragments to obtain a cross-modal heterogeneous graph; performing representation decoupling on the heterogeneous graph, generating a cross-modal shared semantic subspace and a modal specific subspace, and reducing redundancy and noise interference through mutual information minimization constraint; establishing a prototype library, dynamically generating positive and negative sample pairs by adopting cross-modal contrast learning to enhance the discrimination ability, and obtaining an optimized cross-modal fusion result; and finally, outputting a diagnosis result including the fault type and severity, generating a diagnosis evidence track, and binding the key cross-modal fragment with the semantic centroid of the prototype library to realize interpretable traceability and self-correction of the result.

The invention discloses an emotion recognition method and system based on cross-modal alignment and matching enhancement. According to the method, firstly, feature extraction is carried out on text, audio and video modalities in a data set, and then a text and audio cross-modal emotion alignment module and a text and video cross-modal emotion alignment module are constructed respectively, so that cross-modal semantic alignment is realized. Constructing an emotion label matching module based on an alignment result, generating modal pairs with similar emotions but different labels by using a difficult negative sample mining strategy, and paying attention to cross-modal emotion consistency through a dichotomy task guide model; performing modal feature fusion on the three modals through a six-layer attention crossing mechanism, finally splicing feature vectors, inputting the spliced feature vectors into a long-sequence context fusion modeling module for deep modal fusion, and capturing cross-modal interaction information; and the fused features are sent to an emotion classification module, and a final emotion category recognition result is output.

The invention relates to the technical field of power grid intellectualization, in particular to a digital power grid asset classification and evolution monitoring method, system and device based on self-supervised comparative learning and a storage medium. The method comprises the following steps: acquiring network flow data in a digital power grid, extracting multi-dimensional features such as a time interval, a direction, a data packet length, a protocol type and an address port, and constructing an asset behavior feature matrix; then constructing a self-supervised contrast learning model, taking the communication behavior sequences of the same asset in different time periods as a positive sample pair, taking the communication behavior sequences of different assets as a negative sample pair, and training an asset embedding model through an InfoNCE contrast loss function; then, asset communication behaviors are mapped to a semantic embedding space, and asset classification and identification are carried out based on embedded vectors; and finally, carrying out evolution monitoring on the assets by adopting a sliding time window mechanism, generating an asset evolution trajectory sequence, and carrying out abnormity perception early warning. Autonomous learning without manual annotation, high-precision asset identification and real-time state evolution monitoring are realized.

The invention discloses a medical health cost prediction system and method based on multi-source data fusion, and relates to the technical field of medical data analysis, the system firstly fuses a plurality of heterogeneous medical information sources to generate structured data; then, multidimensional features related to the cost are extracted to construct a feature representation vector; and the cost prediction modeling module constructs a cost prediction model through graph structure modeling and semantic embedding, carries out joint optimization by combining a graph attention mechanism and semantic similarity, and outputs a prediction result by fusing a time sequence and static characteristics. A joint optimization algorithm of graph semantic comparison loss and prediction deviation loss is introduced into model training, and positive and negative sample pairs are constructed through a cost label distance. And the feedback optimization module triggers model updating when the prediction deviation exceeds a threshold value, and dynamically adjusts a model structure and parameters through a deviation index and a sample confidence factor. And the prediction interpretation module analyzes influence factors based on intermediate layer features or gradient propagation, outputs an interpretation report, and improves the transparency and credibility of the model.

The invention discloses a method for identifying small target features in a ray detection image, and relates to the field of image processing, and the method comprises the steps: carrying out the size unification, pixel standardization and normalization preprocessing of an original ray image; constructing a training sample set through image region cutting, and introducing a dynamic sampling strategy to realize positive and negative sample proportion adaptive control; enhancing the number and diversity of small target samples in a training set by using point-shaped and linear artificial defect generation strategies; constructing an image segmentation model and introducing feature jump connection to fuse shallow space and deep semantic information; combining Dice loss and Focal loss to form a composite loss function, and guiding the model to pay attention to a target region with a small area and weak gray level; and finally, a segmentation result is optimized through morphological processing and connected domain analysis, and structured target detection information is output. According to the invention, the recognition accuracy and integrity of the tiny target in the ray image can be effectively improved, and the adaptive capacity of the detection method to the change of the imaging quality is enhanced.

The invention discloses an intelligent interpretable traffic signal adaptive control method. The method comprises the following steps: training an intelligent agent through reinforcement learning according to environment state information of an intersection; a timing decision of each phase of the intersection is generated by using the intelligent agent; guiding the first large language model to generate pre-training data by the timing decision and the cue word to perform LoRA fine tuning on the second large language model, and inputting the cue word into the fine-tuned second large language model to enable the second large language model to generate a plurality of reasoning tracks to generate positive samples and negative samples; performing all-parameter fine tuning on the second large language model to obtain a traffic control signal decision model; and inputting the constructed cue word into a traffic control signal decision model to obtain each phase timing scheme of the intersection. According to the method, the defects that an intelligent traffic signal control algorithm based on deep reinforcement learning lacks interpretability and the cross-scene generalization ability is poor are overcome, and the method has important significance in improving the decision credibility and the deployment efficiency.

The invention discloses a text generation image diffusion model enhancement method based on multi-target preference optimization. The method comprises the following steps: firstly, determining a plurality of reward models, and constructing a sample pair training set comprising positive and negative samples; and then, generating a loss weight of each sample pair in the sample pair training set, performing fine tuning training on the text map diffusion model by using the sample pair training set, and in the fine tuning training process, calculating a loss function value of each sample pair in combination with the loss weight of each sample pair until the training is completed, thereby obtaining an aligned text map diffusion model. According to the method provided by the invention, manual data annotation is not needed, the problems of preference inconsistency and over-optimization in a multi-reward scene are effectively solved, and the image quality, the text alignment capability and the multi-target optimization performance of the text-to-image generation model are remarkably improved. The method is superior to an existing optimization method under single-reward and multi-reward setting, shows higher generation quality and robustness, and can be seamlessly applied to various picture generation models.

The invention discloses a liver cancer clinical decision support method and system based on a large language model and a medium, and relates to the technical field of artificial intelligence. Synthesizing the domain enhancement model into a high-quality liver cancer clinical reasoning instruction set containing an intermediate reasoning basis, and performing supervised instruction fine tuning on the domain enhancement model to obtain an instruction fine tuning model; constructing positive and negative sample pairs, and training the instruction fine tuning model by a grouping relative strategy optimization algorithm and Monte Carlo tree search, so that model output is aligned with human expert preferences, and a final liver cancer auxiliary diagnosis large language model is obtained for liver cancer clinical decision making. According to the method, medical guidelines, clinical data and expert experience in the liver cancer field are efficiently injected into a large language model through a three-stage training strategy of incremental prediction training, supervision fine tuning and preference alignment, so that the liver cancer field masters accurate diagnostic logic and term expression.

A method of emotion recognition in cross-subject EEG signals, belonging to technical field of deep learning, includes the following steps: S1, constructing the extracted DE features into positive and negative samples by using a positive and negative sample generator; S2, sending the DE features of an anchor and the positive and negative samples into the encoder for coding, mapping the DE features to a latent space, performing regression prediction on the encoded anchor samples in the latent space by using an autoregressive model, training the encoder by using a probability supervision contrastive loss function; and S3, connecting the trained encoder to the classifier for fine tuning, and training the classifier through the cross entropy loss function; in this process, the encoder does not perform gradient propagation to complete cross-subject emotion recognition.

The invention relates to a small-sample contrast enhancement fine tuning method and system based on a large language model, which are used for identifying named entities in recruitment texts. The method comprises the steps of performing cleaning and format conversion on an original recruitment text, and generating an input sample conforming to a natural language instruction format; under the condition that the labeled samples are insufficient, positive and negative sample pairs are constructed to enhance the recognition capability of the model on entity categories and boundaries; carrying out low-rank parameter updating on the pre-trained large language model by adopting a LoRA fine tuning technology, and reducing computing resource consumption in combination with 4-bit quantitative training; in a pre-training large language model reasoning process, through a multi-dimensional joint confidence evaluation mechanism, confidence of four dimensions of entity levels, lengths, types and contexts is synthesized, and low-confidence identification results are filtered after dynamic weighted normalization processing. The method is suitable for recruitment recommendation, talent matching and other downstream tasks, and has the advantages of high recognition accuracy, low training cost, high system robustness and the like.

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 consistent learning-oriented sample enhancement and optimization method, which comprises the following steps of: respectively performing knowledge extraction on professional knowledge data, professional vocabulary interpretation, historical practice data and an expert experience set in an unstructured text to obtain an original question and answer pair and a professional knowledge base; processing the original question and answer pairs through a diversified strategy to generate enhanced sample question and answer pairs with consistent semantics and various forms, and expanding or simplifying the enhanced sample question and answer pairs to obtain positive samples; introducing an interference expression or a fact error into the enhanced sample question and answer pair to obtain a negative sample; adjusting the positive and negative sample proportion according to the large model accuracy; and taking the positive sample and the negative sample as training data, and performing parameter fine tuning based on the pre-trained large model and the professional knowledge base to obtain a professional large model. According to the method, the adaptability and robustness of a large model in different scenes are improved, and the occurrence probability of irrelevant information and wrong answers is reduced.

The invention provides a manufacturing logistics intelligent matching recommendation method and system based on multi-modal data, and relates to the technical field of intelligent logistics, and the method comprises the steps: obtaining the logistics demand data of a manufacturing enterprise and the service capability data of a logistics supplier; constructing a multi-level feature attention network to perform feature extraction on the logistics demand data and the service capability data, and performing screening to obtain multi-modal feature representation; constructing a positive and negative sample pair based on the multi-modal feature representation, obtaining an inter-modal semantic mapping matrix through a hierarchical contrast learning mode, and projecting the multi-modal feature representation to a unified semantic space to obtain a fusion feature representation; the fusion feature representation and the historical cooperation record are constructed into a state vector, an instant reward function and a long-term reward function are constructed, and a matching degree score between the manufacturing enterprise and the logistics supplier is obtained through optimization of a deep reinforcement learning algorithm; and recommending a logistics supplier to the manufacturing enterprise based on the matching degree score.

The invention provides a dual-mode image fusion method and device based on tuple disturbance and a storage medium, and relates to the technical field of image fusion processing. The method comprises the following steps: constructing a model comprising an encoder, a feature fusion network and a decoder by collecting a cross-modal image group; an encoder is trained based on a tuple disturbance comparison learning framework, cross-modal sharing and complementary features can be mined, learnable positive and negative sample pairs are constructed through a tuple disturbance module, dependence on the same mode is avoided, and generalization is enhanced; the feature fusion network performs complementary feature fusion and multi-scale semantic aggregation training by using cross-modal deep features to improve the quality of a fused image; multi-scale semantic aggregation solves the problem of detail loss, and spatial feature calibration strengthens details, so that a fused image retains bimodal characteristics, and semantic consistency is enhanced; and a combined loss function is constructed, a pixel intensity and gradient structure difference optimization model is integrated, the performance and robustness are improved, key information of the fused image is ensured, and the visual effect and the accuracy are better.

The invention belongs to the technical field of artificial intelligence, and relates to a multimodal model semantic enhancement and comparative learning method based on a colored lamp knowledge graph, which comprises the following steps: associating an original text with a knowledge graph to generate a structured sentence tree, and carrying out embedded coding to obtain an embedded matrix; inputting the embedded matrix and the visibility matrix into a stacked Mask-Transform encoder to obtain a structured semantic feature, and carrying out modeling through a stacked self-attention block to obtain structured knowledge; respectively inputting the original text into a text encoder and a visual encoder of the multi-modal model, obtaining a reference text feature and a reference image feature, and carrying out dynamic gating weighted fusion to obtain a fusion vector; and obtaining a positive sample text, obtaining a high-quality negative set according to the positive sample text, inputting the fusion vector, the reference image features and difficult negative samples in the corresponding high-quality negative set into a contrast learning module, and obtaining symmetric contrast learning loss for training a multi-modal model.

The invention discloses a remote sensing rotating target detection method and system based on comparative heterogeneous knowledge distillation. The method comprises the following steps: constructing a comparative heterogeneous knowledge distillation network detection framework; the Token sequence of the teacher network is converted into a two-dimensional feature map, and the two-dimensional feature map is matched with the inductive bias of the CNN of the student network to obtain transformed teacher features; converting the two-dimensional feature map of the student network into a Token sequence, and inputting the Token sequence into a teacher network for global interaction to obtain converted student features; constructing an assistant transformation model by sharing teacher network weight fusion transformation teacher features and student features; dividing positive and negative sample sets by cooperatively enhancing feature expression through space attention and channel attention and combining cosine similarity matching and a dynamic threshold division strategy; and combining a teacher network, a student network and an assistant transformation model, constructing comparative learning knowledge distillation in combination with positive and negative sample sets, realizing knowledge migration through feature distillation loss and logic distillation loss, and outputting a detection result image.

The invention relates to the field of artificial intelligence cross-modal retrieval, and provides a video text cross-modal retrieval method and system based on spatio-temporal feature fusion. The method comprises the following steps: carrying out key frame sampling and time sequence partitioning on an input video, extracting static visual features through a spatial feature network, and extracting motion features through a time dynamic network; a self-adaptive gating fusion module is adopted to dynamically calculate spatial-temporal feature weights and perform weighted fusion; extracting text semantic features by using a pre-training language model; constructing a double-flow projection network to map video fusion features and text features to a unified measurement space, and optimizing a feature distance by adopting a contrast loss function containing difficult negative sample mining and intra-modal constraint; and outputting a retrieval result according to the cosine similarity sequence. The system comprises four units, wherein the gating fusion module is integrated with an FPGA acceleration circuit. According to the method, mAP (at) 10 is equal to 0.78 in a UCF-101 data set, the time sequence action retrieval accuracy rate is 92.8%, and the single video retrieval delay is 23 milliseconds.

The invention discloses a reaction site prediction method and device based on chemical and physical prior driving, and the method comprises the steps: extracting set features through the multi-modal input of a fusion molecular map, an SMILES sequence and a three-dimensional conformation; generating atomic embedding by using a message passing neural network, and calculating a mixed feature fusing a topological path and a three-dimensional distance; combining the key type weight to construct a graph position code of chemical environment correction; injecting the mixed distance and the charge difference into a Transform attention mechanism, and explicitly modeling an inter-atomic long-range electron effect; a model is jointly trained through double tasks of comparative learning and mask prediction, the comparative learning adopts a directional negative sample to enhance generalization, and mask prediction synchronously recovers an atom type and a charge transfer matrix; and finally, injecting quantum chemistry priori constraint attention weights such as a Fuzzy well function, outputting an atomic-scale reaction activity probability, generating a thermodynamic diagram, and realizing high-precision and interpretable active site labeling. According to the method, the drug design and reaction mechanism analysis efficiency can be remarkably improved.

The invention provides a training-free text-image generation method based on a diffusion model, and relates to the technical field of computer graphic processing and artificial intelligence. The method comprises the following steps: extracting semantic phrases and layout information in an input text by utilizing a natural language model, inputting the input text, the semantic phrases and the layout information as additional conditions into a diffusion model, and extracting cross attention maps of different time steps; a positive and negative sample concept and a foreground and background concept based on an object are constructed, a new loss function is calculated on a cross attention map for semantic information and layout information, the loss function combines semantic loss, regional loss and original loss of a diffusion model and is used for updating a potential space image, and the image is generated through iterative denoising and a decoder. According to the method, additional training is not needed, image generation output based on the diffusion model better meets text requirements, and a better text and image alignment effect is achieved.

The invention discloses a generative dialogue optimization method and system based on multi-step intensified simulation, and the method comprises the steps: generating a plurality of candidate replies based on the current dialogue history through a basic dialogue model; using the user simulator to generate a dialogue chain for each candidate reply; calculating a reward for each dialogue chain, selecting K dialogue chains with the highest reward value as high-quality samples, and taking the dialogue chain with the lowest reward value as a negative sample; updating parameters of the basic dialogue model by using the high-quality sample and the negative sample through a contrast learning and reinforcement learning method to obtain a trained basic dialogue model as a customer service model; and the customer service model generates a reply according to a statement input by the user. According to the conversation generation method, multi-round conversation dynamic planning, medical scene adaptability optimization and efficient strategy exploration can be fused, and double optimization of user trust establishment and business goal achievement is achieved while compliance is ensured.

The invention discloses a battery health state evolution path prediction method based on subgraph representation learning, and aims to overcome the defects in the prior art, obtain the conversion relation between different fault key safety states and support battery fault early warning. The method comprises the following steps: firstly, collecting battery characteristic data, cleaning, serializing and segmenting, and performing efficient compression by using an auto-encoder; secondly, extracting a key state by adopting a data flow clustering technology, regarding segments as small micro-clusters, and integrating charging sequences to form large micro-clusters which are used as key state nodes of an evolution process; then, a state transition diagram is constructed based on the time sequence transition relation of the battery between the micro-clusters, nodes represent key states, and edges represent state transition; then, for any to-be-predicted node pair, dynamically extracting a closed sub-graph, and designing a structure identification vector containing four-dimensional topological characteristics for node marking; then, constructing an enhanced sub-graph by injecting a negative sample edge, and carrying out representation learning by adopting a multi-head graph attention network; and finally, performing link prediction by using the trained model, screening high-probability connecting edges, and splicing the high-probability connecting edges according to a time sequence to form a directed evolution path. According to the method, through subgraph extraction and composite topology marking, negative sample enhanced representation learning and an evolution path splicing mechanism, the prediction precision is remarkably improved, accurate description of the evolution trajectory of the full life cycle health state of the battery is realized, and a visual and quantifiable technical support is provided for fault early warning.

The invention belongs to the technical field of industrial operation and maintenance term processing, and provides a big language model training method based on a knowledge graph, and the method comprises the steps: carrying out the cooperation with operation and maintenance personnel in advance to set a fault dictionary, collecting fault instances, carrying out the classification and preprocessing, extracting causal trigger words of the fault instances, and setting a text dependency relationship rule. Splitting and extracting a triple of the fault instance, and importing to form a knowledge graph; mapping the instance attribute into a node feature vector, outputting an embedded representation of a node, and fusing the embedded representation with the multi-modal feature vector to form a fused feature embedded representation; replacing semantic and degree words with fault instances in the knowledge graph to generate positive samples, randomly replacing subjects, objects and causal relationships to generate negative samples, inputting the negative samples into a semantic encoder, mapping the negative samples to the same metric space, and calculating similarity to obtain a comparison loss value to adjust graph convolutional network parameters; and after the fusion feature embedding representation retrieval of the fault instance is carried out and a subgraph is generated, maintenance suggestions are given step by step along a causal chain.

The invention discloses a cross-modal remote sensing image unsupervised adaptive method based on unreliable pseudo tag guidance. The method comprises the following steps: step 1, acquiring data of a source domain and a target domain and preprocessing the data; 2, constructing a teacher-student self-training framework, generating a pseudo tag for a target domain by a teacher network, and filtering noise through confidence evaluation; 3, dividing the image into reliable pixels and unreliable pixels according to the confidence coefficient, and generating a mask; 4, constructing positive samples, negative samples and anchor point features; 5, designing an unreliable sample guide pixel contrast loss function; 6, optimizing the loss function training model until the optimal performance is achieved; and 7, predicting test set data by adopting the trained model to obtain a semantic segmentation result. According to the method, the potential of pseudo labels of which modals are difficult to label is fully mined, the problem that unreliable pixels are insufficient in utilization during training is solved, and finally more effective cross-modal domain alignment and better cross-modal unsupervised domain adaptive semantic segmentation precision are realized.

The invention discloses an improved automobile paint surface damage detection method based on RT-DETR. The method comprises the following steps: firstly, constructing a high-quality paint surface damage data set, performing data enhancement, and introducing a negative sample and an interference sample to improve the generalization ability and the interference resistance of a model; secondly, in a Backbone network of RT-DETR-ResNet18, an improved dynamic convolution hybrid module (DCMB) is introduced, and on the basis of an original structure, the sensing ability of a strong model to complex textures, edges and heterogeneous regions is optimized; and finally, adding a full-scale frequency attention structure (CSPO) into the Neck part to improve the identification capability of the model on a fine-grained damage region. Compared with the prior art, the method of the invention realizes high-precision detection of eight types of damages such as scratches, recesses and corrosion on the automobile paint surface through a deep learning technology, has the advantages of high detection speed, high accuracy and the like, can be widely applied to the fields of automobile maintenance, second-hand automobile evaluation and the like, and has a wide application prospect. And the automation level and the detection efficiency of automobile paint surface damage detection are obviously improved.