62results about How to "Increase training data" patented technology

The invention discloses a transverse federation and longitudinal federation combination method and device, equipment and a medium. The transverse federation and longitudinal federation combination method comprises the steps of obtaining available public information, inputting the available public information into a preset longitudinal federation service party; obtaining vector information, and training a longitudinal federation model of the preset longitudinal federation service party based on the vector information, updating the network weight of each preset reinforcement learning model, periodically inputting each updated preset reinforcement learning model into a preset transverse federation server, and iteratively updating each updated preset reinforcement learning model. The technicalproblem of high resource consumption of a computing system of a reinforcement learning model in the prior art is solved.

The invention discloses a method and device for recognizing a picture by combining a text and the picture, belonging to the technical field of computers. The method comprises a step of acquiring a first tag in text information, a step of identifying and acquiring a second tag indicated by a plurality of picture elements contained in the picture, and a step of outputting a recognition result according to the first tag and the second tag. The extraction is carried out by combining related text information and image information of the picture, corresponding tags are added to a tag library, training data of the tag library is increased, a neural network recognition model with stronger recognition ability is obtained, a basis is provided for enriching knowledge maps of people or goods, commodities and the like in a laster stage, a new object is identified on the above basis, finally, the accuracy of recognition is greatly improved, and the method and the device can be well utilized in the field of network search engines or commodity shopping platforms.

The invention provides a character recognition method and device. The method comprises the steps of collecting a character image including a character to be recognized and conducting character image preprocessing; uploading an image obtained after the preprocessing to a cloud and conducting unstructured storage, meanwhile, adopting an OCR model to conduct OCR, after the recognition, storing the recognition result and character features under the corresponding character in a character feature pool, and adding the multi-aspect features of the same character; retraining the OCR model by means of the updated character feature pool. The device comprises a preprocessing module, a recognition module, a storage module and a training module. By means of the method and device, the adaptation to recognizing character images under different application contexts and at different shooting angles in the practical application process can be improved.

Systems are provided for utilizing crowdsourcing and machine learning to improve computer system security processes associated with user risk profiles and sign-in profiles. Risk profiles of known users and logged sign-ins are confirmed by user input as either safe or compromised. This input is used as crowdsourced feedback to generate label data for training/refining machine learning algorithms used to generate corresponding risky profile reports. The risky profile reports are used to provide updated assessments and initial assessments of known users and logged sign-ins, as well as newly discovered users and new sign-in attempts, respectively. These assessments are further confirmed or modified to further update the machine learning and risky profile reports.

The invention discloses a dynamic gesture recognition method and system based on a two-dimensional convolutional network, and the method comprises the steps: collecting an actual dynamic gesture video, and carrying out the framing processing of the video; carrying out frame sampling on the actual image after framing; encoding the actual image after frame sampling to obtain an actual feature vectorof the actual image; fusing the actual feature vectors to obtain an actual feature matrix; and inputting the actual feature matrix into the trained two-dimensional convolutional neural network, and outputting a gesture recognition result. By processing a source video stream into a frame of image, and sending the frame of image into a two-dimensional convolutional network, a classification resultof a gesture action is obtained; wherein the image generated by the video contains the spatial feature information and the time sequence information of the video. According to the method, the calculation complexity of gesture recognition is effectively reduced.

A robust width learning system collects training data and performs linear conversion processing on the training data; solving an extended input matrix by using the input data matrix and the enhanced node matrix, solving an iterative initial connection weight matrix by using a ridge regression algorithm, and solving a residual matrix by using a residual formula; obtaining a residual probability density function by using a kernel density estimation algorithm, and calculating a weight matrix formed by all training data; and solving the connection weight matrix of the kth iteration, and if the maximum value of the absolute value of the difference between the output weights of two adjacent steps is not greater than a set threshold value or the number of iterations reaches a preset maximum number of iterations, ending the iteration, stopping the training of the model by the robust width learning system, and establishing a robust width learning system model. According to the system, the robustness of a width learning system can be improved, the problem of adverse effects on modeling precision caused by outliers can be effectively suppressed, and a robust width system model can be conveniently established so as to be suitable for prediction of related indexes in a complex industrial process.

The invention discloses a generative conference abstracting method based on a graph convolutional neural network, and relates to the generative conference abstracting method based on the graph convolutional neural network. The method aims to solve the problem that according to an existing method, only sequence structures of sentences and words are used for modeling conference texts, and rich dialogue chapter structure information of a conference is ignored. The method comprises the steps of 1, obtaining a dialogue chapter structure of a conference; 2, constructing a conference chapter structure chart and a dialogue chapter structure between sentences in a conference; 3, constructing a conference chapter structure chart of the pseudo data and the corresponding pseudo data; 4, obtaining a pre-trained generative conference abstract model and initialization parameters of the graph neural network; obtaining a trained generative conference abstract model and model parameters of the graph neural network; and testing the conference to be tested by using the trained generative conference abstract model of the graph neural network to generate an abstract. The method is used for a generativeconference abstracting method in the field of natural language processing.

The invention discloses a single-person posture estimation method based on a novel high-resolution network architecture. The method comprises the following steps: firstly, detecting an input image containing a single pedestrian by using a detector, removing an inaccurate detection box, and then expanding a data set through data enhancement; secondly, keeping a high-resolution feature map in an instantiated network structure through the parallel multi-resolution subnets without recovering the resolution, introducing exchange units into the parallel subnets, wherein each subnet repeatedly receives information from other parallel subnets, and the accuracy of single-person posture estimation is improved. In most complex scenes; the key points are shielded, so that a data enhancement scheme using one key point for shielding is provided, the trained convolutional neural network can be finely adjusted very effectively through the scheme, the shielded key points are positioned strongly throughadjacent matching, the accuracy of the shielding problem is improved, and a better model is obtained.

The invention discloses a remote sensing image semantic segmentation method, a storage medium and computing equipment, and the method comprises the steps: carrying out equal-proportion cutting of a large-scale remote sensing image and a corresponding label image, and obtaining a small remote sensing image for training; randomly processing the small remote sensing images and the corresponding labelimages, and numbering all the generated training images and the corresponding label images in sequence to obtain an expanded remote sensing image data set for training; constructing a loss function and sequentially putting the training pictures into a semantic segmentation network to obtain a trained remote sensing image semantic segmentation network; traversing all pixels of the whole feature map and carrying out optimization operation, then carrying out merging operation on the randomly cut small remote sensing images, carrying out majority voting on overlapped parts to obtain a segmentation result of the large-scale remote sensing images, and completing feature map merging operation. The method is high in processing speed and good in effect.

The invention discloses an incremental social event detection method based on a graph neural network, and the method comprises the steps: extracting information in a text for the social network data transmitted in a streaming manner, and carrying out the heterogeneous information network modeling; obtaining an isomorphic network; obtaining an isomorphic social contact message graph; learning the isomorphic message graph by adopting a graph attention model to obtain a message code based on knowledge retention increment; meanwhile, sampling message codes, carrying out comparative learning to calculate loss, adjusting parameters of the graph attention model according to the returned loss, and training the graph attention model; and performing clustering on codes obtained by detecting the graph attention model to obtain social events. Rich semantic and structural information is fully fused into the social messages, knowledge obtained from the social messages is reserved through the graph neural network, the comparative learning technology is adopted, the message graph is periodically maintained, and the event detection accuracy can be improved under the condition that excessive resources are not consumed.

The invention discloses a ticket identification method, relates to the technical field of text detection, text identification and information structured extraction, solves the technical problem that an existing model cannot effectively extract structured information, and is characterized in that a text line position detection model is obtained by training a CTPN network; therefore, the key information in the ticket is positioned, and the method has robustness for tickets in various forms (tables and the like). Data is synthesized through rules of high-frequency words and specific field text contents in the high-frequency words, training data of a text recognition model is expanded, and the accuracy of the recognition model is improved; based on the convolutional neural network, the method has good parallelism, and calculation can be accelerated by using a high-performance GPU (Graphic Processing Unit).

The invention discloses an English punctuation mark adding method, system and device based on data enhancement. The method comprises the steps that text information is acquired and preprocessed to obtain training data, data enhancement processing is conducted on the training data, wherein the data enhancement comprises random deletion, random replacement, random syllable-similar word replacement and random insertion, and enhanced data is obtained; the method further includes integrating the original data before enhancement and the data after enhancement together to serve as a training data set; and performing model training by using the training data set to obtain a prediction model, the prediction model being used for adding English punctuation marks to the input text information. According to the invention, the real data is simulated by performing data enhancement processing on the training data, so that the prediction model obtained by training is more robust, the effect in a speechrecognition system is better, the operand is not increased, and compared with the mode of labeling a large number of real texts with punctuations, the manpower can be saved and the cost can be reduced.

The invention relates to a coding and decoding network port image segmentation method fusing semantic flow field, and belongs to the technical field of image segmentation, and the method comprises the following steps:inputting an image to be segmented into a trained coding and decoding network fused with the semantic flow field, and segmenting the port image into a sea type, a land type and a ship type; wherein the coding and decoding network comprises a coding layer, a dilated convolution layer and a decoding layer which are connected in sequence, the coding layer comprises N layers of convolution modules which are connected in sequence, the decoding layer comprises N layers of deconvolution modules which are connected in sequence, each deconvolution module is internally provided with a flow alignment module, and the input of each flow alignment module is in jump connection with the convolution module of the corresponding level in the coding layer. According to the method, the validity of feature information transmission is improved by predicting a semantic flow field between feature maps and monitoring an up-sampling process by utilizing the flow alignment module, and the multi-scale information of the image is acquired by utilizing the cavity convolution layer, so that the method is more suitable for a port image segmentation task, a smooth and complete segmentation result is obtained, and the segmentation precision is relatively high.

The invention relates to the technical field of data processing, and in particular to a quantitative analysis model processing method, a quantitative analysis model processing system, a storage mediumand an electronic terminal. The method comprises the following steps of: normalizing post-rehabilitated security history market data; calculating technical index data according to the normalized security history market data and normalizing the technical index data; and training the normalized security history market data and the technical index data through a machine learning algorithm, and obtaining a quantitative analysis model. The quantitative analysis model obtained through training is adapted to the markets of different securities and has better generalization ability.

The invention discloses a data enhancement method for PCB image defect detection, and the method comprises the steps: setting an image defect size threshold, and setting a standard size; converting the sizes of the image defects of which the image defect sizes are greater than an image defect size threshold value in the PCB image information list into standard sizes, and randomly cutting the imagedefects of which the image defect sizes are smaller than the image defect size threshold value; and updating the PCB image information list. After the PCB image defects are processed for multiple times by adopting the method, the sample number of the small-size defect images is enriched, the training data of the small-size defect images is enhanced, the generalization ability of the PCB image defect detection deep learning model is enhanced, and the small defects of the PCB images are easier to detect.

The invention discloses an error correction method and system for a real-time translated text, a storage medium and a device, and relates to the field of voice error correction. The method comprises the steps that: an ASR translation text in real-time live broadcast is acquired, the ASR translation text is interpreted through a trained BERT error correction model, and a first error correction text is output; the ASR translation text is interpreted through the trained GPT error correction model, a second error correction text is output, the first error correction text and the second error correction text are combined to obtain an error correction target text, and end-to-end error correction is performed on the ASR translation text content in the live broadcast scene through the scheme, the character accuracy of the ASR on an audio translated text can be effectively improved, and the method can be quickly applied to the field of live broadcast.

The invention discloses a remote sensing image ship detection method and device based on an attention model, and the method comprises the steps: 1, collecting a ship remote sensing image, and carrying out the data expansion of an image data set through annotation information; step 2, preprocessing the collected remote sensing image to obtain a data set used for ship detection model training; 3, inputting the images in the training data set into a YOLOV5 attention model improved for a small-size target, and training to obtain a trained remote sensing ship detection model; 4, cutting a remote sensing image to be detected, inputting the cut remote sensing image into the trained remote sensing ship detection model, and outputting a bounding box and confidence of a ship; and mapping bounding boxes of all the cut images back to the original remote sensing image, and obtaining a final detection result after filtering repeated targets through confidence threshold filtering and non-maximum suppression. According to the invention, the problem of difficult detection caused by sparse ship distribution and too small size in the remote sensing image is solved.

The invention relates to a computer implemented method of formulating a database query for searching for a data pattern in a data set stored in a database, the method comprising the following steps: receiving, from a user terminal, a series of graphical data points defining a target pattern to be searched; formulating a data structure for training a machine learning model using the series of graphical data points; training a machine learning model using the data structure; applying the trained machine learning model to the data set stored in the database to identify one or more candidate patterns, the candidate patterns comprising intervals of the data set which correspond to the target pattern within a predefined confidence level. Since the input is a series of graphical data points defining the target pattern, a specific pattern of interest may be entered much more efficiently and intuitively. By converting the user input into a data structure for training a machine learning model, the specific graphical user input is used as training data to identify similar patterns in a large data set, providing much greater search accuracy than previous methods.

The invention discloses an image correction method, an image correction device and electronic equipment. According to one embodiment, the image correction method can comprise the steps: receiving a to-be-corrected image; and performing color correction on the to-be-corrected image by using a color correction model, wherein the color correction model is obtained by training images of color cards atdifferent color temperatures. In some examples, the image of the color card may be an original RAW image. The color correction model is trained by using the original RAW images of the color card under different color temperatures, so the influence of factors such as ambient light can be eliminated, high correction accuracy is realized, and the color correction method has a wide application range.

The invention provides a cooling-water machine adjustment model training method and device and electronic equipment, and relates to the technical field of cooling-water machine control, and the methodcomprises the steps of obtaining a first sample set and a second sample set, determining the similarity between samples in the first sample set and samples in the second sample set, training a cooling-water machine adjustment model of the target cooling-water machine according to a plurality of samples with similarity meeting a first preset condition in a second sample set, the samples in the first sample set being generated according to historical operation data of the target cooling-water machine, and the samples in the second sample set being generated according to historical operation data of the similar cooling-water machine of the target cooling-water machine. According to the technical scheme provided by the invention, the sample with high similarity can be selected as the trainingsample in the second sample set similar to the first sample set, so that the training data of the target cooling-water machine can be effectively increased, and the good accuracy of the cooling-watermachine adjustment model is ensured.

The invention relates to a dependency syntactic analysis method fusing multi-strategy data enhancement under a low-resource condition, and belongs to the field of natural language processing. The method comprises the following steps: constructing homomorphic synonym dictionary of Thai, Vietnamese and English; carrying out synonym replacement on small-scale UD (Universal Dependencies treebanks) data sets of three languages by utilizing a synonym dictionary so as to expand training data; and performing mixup on original words and synonyms in the training data in different stages of model training by utilizing various mix data enhancement strategies to generate virtual new words for subsequent training. According to the method, various data enhancement strategies are provided for the low-resource-dependency syntactic analysis problem. According to the method, training data are effectively expanded through synonym replacement, and the problem of unknown words is relieved. Through a plurality of mixup data enhancement strategies, the problem of model overfitting is effectively relieved, and the generalization ability of the model is improved.

The invention discloses a product selection method and device based on deep learning, and relates to the technical field of computers. A specific embodiment of the method comprises the following steps: collecting user behavior data, wherein the user behavior data comprises an article list corresponding to each search term; constructing article pairs according to articles in the article list to obtain an article pair list corresponding to each search term; according to the article pair list corresponding to each search word, constructing an article selection model based on a deep learning network, and performing article selection according to the article selection model. According to the embodiment, the article pairs are constructed, the article selection model is constructed based on the constructed article pairs and the deep learning network, and the articles are selected according to the article selection model, so that model training data can be greatly expanded, and the coverage rate and the accuracy of article selection results are improved.