49results about How to "Enrich training data" patented technology

The present application discloses a method, apparatus and medium for acquiring face recognition model training data based on video. The method comprises the following steps of: acquiring a standard picture of a person to be recognized; performing face detection and key point extraction on the standard picture to generate a first descriptor; for the video including the person, the video frame is extracted, the face part in the extracted video frame is recognized, and the face part is saved as a face picture; the key points of the face picture are extracted, a second descriptor is generated, thedistance between the first descriptor and the second descriptor is calculated, and whether the face picture is the person to be recognized is judged based on the distance, so as to obtain the face recognition model training data. This method can enrich the training data of face recognition model, reduce the workload of manual screening, and thus solve the problems of incomplete preparation of training data, single type, and great difficulty in cleaning.

The invention discloses a method for improving the detection precision of ships approaching a shore in an SAR image. The method is based on a deep learning theory, and mainly comprises four parts: a generative adversarial network, K-means clustering, scene amplification, and a classic detection network (Faster R-CNN, Cascade R-CNN, SSD, and RetinaNet). The generative adversarial network realizes image feature extraction, the K-means clustering method realizes image dichotomy by using extracted features to obtain a classification result of each image, scene amplification is performed to obtaina more balanced data set, and the classical detection network performs training by using the processed data set and executes a detection task. While the detection precision of the offshore ship is slightly improved, the detection precision of the ship approaching the shore on the Faster R-CNN, the Cascade R-CNN, the SSD and the RetinaNet network is improved by 8.60%, 8.32%, 18.15% and 12.40% respectively, and the detection precision of the ship approaching the shore is improved.

The embodiment of the invention provides a method and a device for predicting a service index executed by a computer. The method comprises: first, obtaining dataset sequences, wherein the data set sequence comprises m data sets which correspond to m continuous time periods and are arranged according to a time sequence, and the ith data set in the m data sets comprises entity characteristics of a to-be-tested service entity in the ith time period and respective index values of a plurality of to-be-tested service indexes of the to-be-tested service entity; inputting the data set sequence into apre-trained recurrent neural network to obtain an output result. Therefore, according to the output result, the index values of the plurality of to-be-tested service indexes of the to-be-tested service entity in the next time period of the mth time period can be determined.

The invention discloses a method and system for classification modeling based on protein length and DCNN, and belongs to the field of protein prediction and analysis. The technical problem to be solved is how to use deep learning to predict and analyze the secondary structure of the protein and improve the accuracy. The method comprises the following steps: taking a plurality of big data sets as atraining set, extracting PSSM features generated by PSI-Blast in the data sets, and performing format conversion on the PSSM features through a sliding window; grouping proteins in the training set based on the protein length to obtain a plurality of model groups; and constructing a predication model corresponding to every model group based on a deep convolutional network, and training the prediction models by the model groups to obtain trained prediction models. The system includes an input module, a format conversion module, a grouping module and a model training module.

The invention discloses a production line aluminum alloy ingot surface defect detection method, device and system and a storage medium. The method comprises the steps that an aluminum alloy ingot surface defect image data set is collected; selecting YOLOv5 as a detection model, and optimizing the initial anchor frame for an application scene by using a K-means + + algorithm; carrying out accelerated reasoning improvement on the YOLOv5 detection model, and meanwhile, fusing an attention mechanism to dynamically detect a head Dyhead; a Mixup and Mosaic data enhancement algorithm is adopted to enhance the data set, an improved YOLOv5 detection model is trained, and a Focal loss function and non-maximum suppression are introduced during training; and detecting the surface defect image of the aluminum alloy ingot by using the trained model to obtain the type confidence coefficient and position information of the defect in the image. With the adoption of the technical scheme, the types and positions of the defects can be quickly detected, so that the requirement of quick and accurate detection of an aluminum alloy ingot production line is met.

The invention relates to the technical field of battery detection, and particularly provides an anomaly detection method and device for a battery module, and the method comprises the steps: constructing a digital twin model which comprises a battery pack model and a charging system model; performing fault injection on the digital twin model; performing simulation operation on the digital twin model after fault injection to generate a simulation operation data set; constructing a deep neural network, and training the deep neural network based on the simulation operation data set; acquiring battery data of a to-be-detected battery module, and inputting the battery data into the trained deep neural network to obtain an output result; and judging whether the to-be-detected battery module is abnormal or not according to the simulation operation data set and the output result. Therefore, fault injection and simulation operation are carried out on the basis of the digital twinborn model, abundant training data are provided for the deep neural network, the spatial relationship between the battery cells is considered, the accuracy of anomaly detection can be improved, and the false detection rate and the omission ratio can be reduced.

The invention provides an image simulation generation method. According to the method, a simulation degree threshold value is determined based on preliminary analysis of an initial image. The translation predetermined pixel values of the initial image in at least two directions are determined based on the simulation degree threshold value. The weighted average superposition is performed on the translated images so as to output the required simulation image. By utilizing the method, the technical problem that image aberration and the like caused by lenses, optics, photosensitive elements and the like cannot be effectively simulated by the existing image simulation technology can be solved. The image simulation generation system and the electronic equipment thereof have the same technical effects as the method. Based on the deep learning algorithm training method provided by the invention, the problem that the existing data amplification technology cannot well simulate the generated image into a real sampling image can be solved, and a better training effect can be obtained.

The invention discloses a path planning method based on a carbon emission measurement scale, and belongs to the technical field of road traffic path planning. According to the method, firstly, multi-dimensional information such as roads, vehicles, energy consumption, environments and loads is collected to construct a data set; according to the IPCC standard, the carbon dioxide emission amount is calculated in combination with the energy consumption information; secondly, calculating the contribution rate of each feature by using a principal component analysis method, and further screening out main features; according to different numbers of the main features, performing multi-model training on the main features by using linear regression, and performing bagging ensemble learning method fusion on a plurality of models; and finally, generating a plurality of driving paths in combination with a map open API, respectively calculating carbon dioxide emissions of different driving paths, and selecting the driving path with the minimum carbon emission to be recommended to the user. The method focuses on realizing the minimization of the carbon emission of the vehicle in the path planning, and finally assists in promoting China to realize the goals of carbon peak reaching and carbon neutralization.

The invention discloses a semi-supervised learning pseudo label assignment method based on clustering fusion, and the method comprises the steps: carrying out the semi-supervised learning of a convolutional neural network with a label-free data set, carrying out the pre-training of the neural network through employing labeled data and label-free data, and extracting data features through employinga trained network; assigning pseudo tags to N pieces of untagged data closest to the tagged data by using a nearest neighbor method; analyzing all the data information by using k-means clustering, and endowing clustered pseudo tags to the data which is not tagged; and continuously training the convolutional neural network by using the obtained label data and pseudo label data to obtain an optimalnetwork for label assignment. The method can be suitable for semi-supervised learning under deep learning in various fields; information of label-free data can be fully mined, and training data withricher content are provided for a network; the principle is clear and easy to understand, and codes are easy to implement.

The invention provides a massive image infringement retrieval method and system and a computer readable storage medium thereof. The method comprises the steps: S1, generating a bag-of-words model, extracting SIFT feature points of a template image, obtaining visual vocabularies through clustering processing, and establishing the bag-of-words model; s2, making a training set: calculating an inversedocument weight of each visual vocabulary, and positioning SIFT feature points conforming to a preset threshold value to obtain original training data bying correspondingly cutting the template image; s3, training a neural network: training a CNN network by adopting the original training data in the step S2 according to a comprehensive metric learning and hash learning method to generate binary features; and S4, retrieval judgment: constructing an inverted index system by using the bag-of-words model in the step S1, traversing entries corresponding to the visual vocabularies in the to-be-retrieved image, calculating a Hamming distance between binary features, judging whether the binary features are matched or not according to a preset threshold, and giving an infringement coefficient according to accumulated matching. The infringement image retrieval speed is increased, and meanwhile, relatively high accuracy is ensured.

The invention provides a multi-mode security monitoring method based on deep learning image processing, and the method comprises the steps: dividing a to-be-monitored scene according to the importance degree, carrying out periodic adjustment according to a subsequent machine learning result, and achieving the application of limited monitoring resources to a place where monitoring is most needed; focusing on content needing to be monitored, screening out people who most possibly appear and need to be monitored by setting the danger ID and the sub-danger ID, performing focusing monitoring on people who most possibly appear and need to be monitored in a place where monitoring is most needed by combining the two operations, and meanwhile, collecting an actual scene needing to be monitored, so as to realize real-time monitoring. And a large amount of external training data is pasted, so that the pre-training set data conforming to the to-be-monitored scene is enriched, and the final recognition precision is improved. Meanwhile, through the result after training, the area where the abnormal behavior is most likely to be sent can be adjusted and supervised. The video online monitoring with the maximum efficiency is achieved with low equipment cost and expenditure.

The invention discloses a multi-commodity image synthesis method and device, electronic equipment and a storage medium. The method relates to the field of deep learning, and can comprise the steps: collecting a commodity image of a commodity for any commodity, carrying out the main body detection of the commodity image, and determining a binary segmentation image corresponding to the commodity according to a main body detection result and a constructed collection scene background image; and when a multi-commodity image needs to be synthesized, obtaining an application scene background image, synthesizing the at least two binary segmentation images to the application scene background image, and performing assignment on pixel points in the binary segmentation images according to the commodity images corresponding to the binary segmentation images and corresponding pixel point values in the application scene background image. By applying the scheme of the invention, the labor cost can besaved, and the processing efficiency can be improved.

The invention discloses a rapid image storage method and system suitable for an autonomous driving simulation platform. The rapid image storage method comprises the steps that an image stored in a video memory is directly sent into a computer memory; according to the invention, high-speed lossless screenshot can be realized on the autonomous driving simulation platform, the universality of the autonomous driving simulation platform is improved, and more and more accurate training data can be provided for deep learning algorithm development.

The invention relates to a pedestrian re-identification network training method based on branch learning and layered pseudo labels, wherein a pedestrian re-identification network is a mutual average teaching network. The training method comprises the following steps: obtaining a label data set and a label-free data set, taking the label data set as one layer, dividing the label-free data set into N layers, and assigning the pseudo label to the label-free data of each layer to form N layers of pseudo label data, wherein N is a constant; constructing a branch learning framework comprising N+1 mutual mean-teaching network branches sharing weights, wherein one branch is used for inputting label data for training, and the N layers of pseudo label data are correspondingly input into the other N branches for training; and constructing a loss function of each branch, determining a total loss function of the branch learning framework, performing multiple rounds of training based on the total loss function, and re-layering the label-free data set in each round of training process. Compared with the prior art, the network trained by the invention is more accurate, and the convergence speed of the network during training is high.

The invention discloses a enterprise credit extension model construction method, device and equipment and a readable storage medium. The construction method comprises the following steps: constructinga tax payment data prediction model in advance; obtaining industrial and commercial data of an enterprise, and inputting the industrial and commercial data into the tax payment data prediction modelto obtain tax payment data corresponding to the industrial and commercial data; and based on the industrial and commercial data, the derivative data corresponding to the industrial and commercial dataand the tax payment data, constructing an enterprise credit extension model, so as to predict the credit risk of the target customer based on the enterprise credit extension model after the enterprise credit extension model is constructed. According to the invention, the credit risk of the enterprise can be assessed more comprehensively, so that the result of predicting the credit risk of the target customer is more reliable, and the technical problem that the risk level of the enterprise cannot be assessed comprehensively and multi-dimensionally by the existing enterprise credit scoring technology is solved.

The invention discloses a standard section automatic identification method based on 3D ultrasonic data, and belongs to the technical field of medical imaging, and the method comprises the following steps: obtaining a 2D ultrasonic standard section image and 3D ultrasonic data; the obtained 2D standard section image is preprocessed, and corresponding parameters are set according to classification of different ultrasonic acquisition devices for batch automatic cutting; and meanwhile, automatic cutting is realized by adopting a maximum connected domain method, so that interference of a pure color background and an equipment interface is removed, and pure standard section data is obtained. According to the method, anatomical information is provided by utilizing 3D ultrasonic data, the neural network is guided to realize efficient standard section automatic identification, interference of a background and an equipment interface is removed by designing an automatic cutting method and a 2D data expansion method, a large number of 2D slice images are generated, and the richness of training data is increased. And a VGG16 and OC-CNN design network is combined, and the preprocessed and expanded 2D image is trained to obtain a model for automatically identifying the standard section.

The invention provides an image processing method, device and equipment and a computer readable storage medium, and the method comprises the steps: obtaining a to-be-corrected image collected by an image collection device, wherein the to-be-corrected image is a driver face image with a face rotating by a preset first angle, and the first angle is determined according to an included angle between the image collection device and a driver; taking the first angle as a constraint condition of the to-be-corrected image, and inputting the to-be-corrected image and the first angle into a preset conditional adversarial network to obtain a corrected target image, the target image being a front face image of the driver; and according to the target image and a preset face recognition technology, carrying out recognition operation on the identity information of the driver. Different from the existing technical scheme of labeling the face feature information in the image, the first angle is used asthe constraint condition, and the training data is relatively rich. In the process of face recognition based on a target image, the target image is the front face image of a driver, so that the recognition precision is improved.

The invention discloses a news classification method and a computer readable storage medium, and the method comprises the steps: obtaining original data which comprises a news text and a news category; determining the total number of news texts under each news category; when the total number of the news texts under a certain news category is smaller than a preset threshold value, carrying out dataenhancement processing on the news texts under the news category, so that the total number of the processed news texts under the news category is not smaller than the preset threshold value; and respectively screening out news texts with the same number as the preset threshold from all news texts under each news category as training data, and inputting the news texts into a BERT model for training. According to the news classification method and the computer readable storage medium, the classification accuracy can be improved.

A glass surface stain image recognition method applied to a window-cleaning robot relates to the technical field of image recognition, and comprises the following steps: acquiring image data, environmental perception data and tactile data of a target glass area according to a sensor terminal of the window-cleaning robot; constructing a shooting parameter lookup table, and updating shooting parameters of the sensor terminal in real time according to the environment sensing data in the current acquisition period and the shooting parameter lookup table; constructing a stain feature recognition model, and performing pre-estimated stain type labeling on the image data of the current acquisition period according to the stain feature recognition model; and performing correctness detection on the estimated stain type marked in the target glass area according to the tactile data of the current acquisition period, if a wrong stain type exists in the target glass area, obtaining the recognition correct rate of the stain feature recognition model, and performing local feature enhanced recognition operation or retraining operation according to the recognition correct rate. And the stain identification capability of the window cleaning robot is obviously improved.

The present invention proposes a multi-mode security monitoring method based on deep learning image processing, which divides the scenes that need to be monitored according to the degree of importance, and regularly adjusts them according to the follow-up machine learning results, so as to realize the application of limited monitoring resources in the most needed monitoring At the same time, focus on the content that needs to be monitored. By setting the dangerous ID and the sub-dangerous ID, the people who are most likely to be monitored will be screened out. Focus on the people who need to be monitored the most. At the same time, by collecting the actual scenes that need to be monitored and pasting a large amount of external training data, the data of the pre-training set that meets the scenes to be monitored is enriched, thereby improving the accuracy of the final recognition. At the same time, through the training results, it is also possible to adjust and supervise the areas most likely to send abnormal behavior. Realize video online monitoring with lower equipment cost and overhead to achieve maximum efficiency.

The invention discloses a polyvinyl acetal yellowing detection method based on machine learning, and relates to the technical field of high polymer material detection, and the detection method comprises an experiment modeling stage and a model application stage. Wherein in the experiment modeling stage, the relative absorbance is designed as an aging evaluation standard, and a relative absorbance time curve is obtained; fitting a first mapping relation by utilizing the relative absorbance time curve and a first yellowing degree measured by an experiment; training and constructing a neural network model according to the relative absorbance and the mechanical property parameters; wherein in the application stage, a second yellowing degree of an actual to-be-detected material is obtained, a plurality of actual relative absorbance values are obtained according to the second yellowing degree and a plurality of first mapping relations, and the plurality of actual relative absorbance values are input into the neural network model to obtain actual mechanical property parameters. Through the experimental modeling stage and the model application stage, the mechanical properties of the polyvinyl acetal material in the actual aging process can be more accurately estimated according to the actual yellowing degree.

The invention provides a power transmission line image augmentation method and system based on an improved DCGAN. The method comprises the steps of collecting original defect image sample data of a power transmission line; improving the structure of the DCGAN network model, and obtaining the DCGAN network model with the improved structure; preprocessing the original defect image sample data by using a DCGAN network model with an improved structure to obtain a training set; training the DCGAN network model with the improved structure until a generator G and a discriminator D reach Nash equilibrium; and generating the augmented data by using the generator G. According to the invention, the original data is preprocessed by using the augmentation function, so that more training data can be provided for DCGAN training, a network can learn more comprehensive data set features, and a more effective data set is generated.

The invention discloses a data processing method and device, electronic equipment and a storage medium. The method further comprises the steps of obtaining first clustering information of first sampledata based on a classification model, and determining a selection threshold value according to the first clustering information; acquiring second clustering information of first monitoring data of atarget based on the classification model, and selecting second sample data from the first monitoring data according to the second clustering information and the selection threshold; performing activelearning of the classification model based on the second sample data, and optimizing model parameters of the classification model; classifying the second monitoring data of the target by using the optimized classification model to obtain a classification label of the second monitoring data; and mapping the classification label of the second monitoring data into semantic information, updating the ontology by utilizing the semantic information, and obtaining an event decision for executing a predetermined operation on the target based on the updated ontology.

The invention relates to a smart energy storage management and control method and device, computer equipment and a storage medium, and the method comprises the steps: collecting meteorological data, living information and electric energy consumption data of a management and control region in real time, inputting the meteorological data, the living information and the electric energy consumption data into a preset electric energy consumption prediction model, and obtaining electric energy prediction data; performing difference calculation on the electric energy prediction data and the electric energy consumption data to obtain electric energy difference data, and dividing the control area into a controllable area, a self-sufficient area and a to-be-controlled area according to the electric energy difference data and a preset electric energy difference threshold; and according to a preset electric energy allocation rule, allocating the electric energy in the management and control area, and adjusting the electric energy use strategy of each area in real time to realize optimal configuration of the electric energy. The method has the effect of improving the energy management efficiency.

The present invention provides a training method for a recommendation model, comprising: initializing parameters of the training model, wherein the training model includes a first self-encoder, a first item encoder, a first user encoder, a first decoder, a second The self-encoder, the second item encoder, the second user encoder and the second decoder; obtain training data, wherein the training data includes the first initial score vector of the source domain, the first user content vector, and the first item content vector and the second initial scoring vector, the second user content vector, and the second item content vector of the target domain; the training model is trained according to the training data, and the second user encoder, the second item encoder and the second decoder are used as recommendation Model. The technical scheme of the present invention effectively solves the data sparse problem and the cold start problem.

The present invention discloses a multimodal data enhancement method based on dynamic semantic fusion, comprising the following steps: S1, multimodal data collection and preprocessing; S2, dynamic background, attribute and action generation; S3, semantic fusion and enhancement; S4, multi-view and multi-scene synthesis; S5, self-supervised training and feedback mechanism; S6, dynamic adjustment and optimization; through innovative dynamic background generation, attribute and action adjustment, semantic fusion and multi-view scene synthesis technology, combined with self-supervised training and feedback mechanism, it is not only possible to generate real, diverse and semantically consistent data samples, providing a more comprehensive training foundation for deep learning models, but also to adapt to different training needs and application scenarios by dynamically adjusting data enhancement strategies.

The invention discloses an unsupervised representation learning method and device based on multi-source heterogeneous features, and the method comprises the steps: extracting and integrating multi-source heterogeneous data, designing and training a first encoder of structured data features, and carrying out the coding of the structured data through the first encoder, and obtaining a first coding result; designing and training a second encoder with unstructured data features, and encoding the unstructured data by using the second encoder to obtain a second encoding result; processing the first coding result and the second coding result to obtain a multi-source heterogeneous fusion feature for describing the instance; designing a DNN network based on the multi-source heterogeneous fusion features, and training to obtain feature representation vectors of the multi-source heterogeneous fusion features projected to a multi-dimensional space; and performing the tasks of similarity matching, classification and clustering among the instances by using the feature representation vectors. On the basis of an unsupervised condition, discriminative representation learning of an instance level is realized; and more training data and a better network structure are provided.

The embodiment of the invention provides a method based on combination of block chain data acquisition and model training, and the method comprises the steps: receiving a training starting instruction, sending the starting instruction to a terminal node, and sending an expected result to a feedback unit; the terminal node forwards the instruction to a related peer node according to a preset global parameter; the peer-to-peer node generates training data by using pre-stored resources and weights and transmits the training data to a data processing unit; the data processing unit screens and integrates training data and then inputs the training data into the model training unit; the model training unit calculates an actual result and feeds back the actual result to the feedback unit, and the feedback unit calculates a difference value according to an expected result and the actual result, encapsulates the difference value, a weight and a peer node identifier into a resource packet and sends the resource packet to a terminal node; the terminal node forwards the resource packet to the smart contract of the block chain; the peer-to-peer node runs contract verification, and if the contract verification passes, the contract verification is sent to the sorting node; and the sorting node generates a block to each peer node, and after the peer nodes verify, the blocks are added to the local block chain if the blocks are valid, and the weights are updated if the blocks are invalid, so that subsequent training is optimized.

The embodiment of the invention provides an image processing method and device and a storage medium. The method comprises the steps of obtaining a first image which comprises at least one time domain feature and at least one frequency domain feature and is obtained according to first voice data to be processed, and updating the first image according to a preset strategy to obtain a second image, wherein the preset strategy at least comprises one of the following items: transforming a time domain position of at least one time domain feature on a time domain, or transforming the frequency domain position of at least one frequency domain feature on the frequency domain. According to the scheme, rich and diversified training data can be provided for the training model, so that the training effect is improved.