310results about How to "Reduce labeling costs" patented technology

The invention provides an automatic segmentation method for an MRI (Magnetic Resonance Imaging) image brain tumor based on a full convolutional network. The method comprises multi-mode MRI image preprocessing of the brain tumor, construction of a full convolutional network model, network training and parameter optimization as well as automatic segmentation of a brain tumor image, specifically, the segmentation of the MRI image brain tumor is converted into a pixel-level semantic annotation problem and differential information emphasizing different modes of MRI, two-dimensional whole slices of four modes FLAIR, T1, T1c and T2 are synthesized into a four-channel input image, the convolutional layer and the pooling layer of the trained convolutional neural network are base feature layers, three convolutional layers equal to a full connection layer are added behind the base feature layers to form a middle layer, the middle layer outputs rough segmentation images corresponding to semantic segmentation types in quantity, and a de-convolutional network is added behind the middle layer and used for interpolating the rough segmentation images to obtain a fine segmentation image having the same size as the original image. The method does not need manual intervention, effectively improves the segmentation precision and efficiency, and shortens the training time.

The invention discloses a Chinese word segmentation method based on two-way LSTM, CNN and CRF which improves and optimizes traditional Chinese word segmentation base on deep learning algorithm. The method comprises following specific steps: preprocessing the initial corpus, extracting corpus character feature information and pinyin feature information corresponding to characters; using the convolutional neural network to obtain pinyin feature information vector of the characters; using the word2vec model to obtain the character feature information vector of text; splicing pinyin feature vectors and character feature vectors to obtain context information vectors and put the context information vectors to a bidirectional LSTM neural network; decoding the output of the bidirectional LSTM using the linear chain condition random field to obtain the word segmentation sequence; decoding the word segmentation label sequence to obtain word segmentation results. The invention utilizes the deep neural network to extract text character features and pinyin features and combines the conditional random field decoding, can effectively extract Chinese text features and achieve good effect on Chinese word segmentation tasks.

The present invention is directed to two fundamental characteristics of a scene image: (1) the scene image often containing complex semantics; and (2) a great number of manual annotation images taking high labor cost. The invention further discloses a scene image annotation method based on an active learning and a multi-label and multi-instance learning. The method comprises: training an initial classification model on the basis of a label image; predicting a label to an unlabeled image; calculating a confidence of the classification model; selecting an unlabeled image with the greatest uncertainty; experts carrying on a manual annotation on the image; updating an image set; and stopping when an algorithm meeting the requirements. An active learning strategy utilized by the method ensures accuracy of the classification model, and significantly reduces the quantity of the scenario image needed to be manually annotated, thereby decreasing the annotation cost. Moreover, according to the method, the image is converted to a multi-label and multi-instance data, complex semantics of the image has a reasonable demonstration, and accuracy of image annotation is improved.

The invention relates to an interactive method of image semi-automatic annotation, comprising S1 dividing an initial sample into three different types of annotation samples according to different category attributes; labeling the three types of labeling samples manually to get different kinds of labeling results, and then using three models of Mask-RCNN, Fast-RCNN and FCN to train separately; S2 processing the data set of the picture to be annotated in an offline manner, wherein the annotating process is that the data set of the picture to be annotated passes through the three depth learning models in turn to output the json format files of all types and coordinate points of the data samples; S3 calling the relevant attribute tag value and coordinate point value of the json format file according to the name of the annotated image; S4 displaying the corresponding automatic marking result in the marking software, and judging whether the category and area marking of the target object arestandardized and reasonable by manpower; S5 carrying out data augmentation on the correctly labeled labeling samples and feeding back the augmented data to the model for retraining.

The invention provides a sample marking method and device based on artificial intelligence prosody prediction. The method comprises steps that a first text sequence of unmarked prosodies corresponding to a first sample audio file is acquired; text characteristics and pronunciation duration of each character of the first text sequence are acquired; a pre-trained prosody marking model is applied to the text characteristics and the pronunciation duration of each character of the first text sequence to acquire an output mark of each character of the first text sequence; prosodic hierarchy marking for the first text sequence is carried out according to the output mark of each character of the first text sequence. The method is advantaged in that text marking cost is reduced, text marking efficiency and accuracy are improved, more training samples required for prosodic hierarchy marking can be provided, prosodic marking model performance is improved, and the voice synthesis effect is more natural.

The application discloses a method and a device for marking point cloud data. According to a specific implementation, the method includes the following steps: using a laser radar and a sensor different from the laser radar to collect the data of the same scene to obtain point cloud data and sensor data; segmenting and tracking the point cloud data to obtain a point cloud segmenting and tracking result; identifying and tracking the features in the sensor data to obtain a feature identifying and tracking result; using the feature identifying and tracking result to correct the point cloud segmenting and tracking result to obtain the confidence of the point cloud segmenting and tracking result; and determining a point cloud segmenting and tracking result of which the confidence is greater thana confidence threshold as a point cloud marking result. The implementation reduces the manual workload of point cloud data marking and reduces the cost of marking.

The invention discloses a bidirectional security authentication method for an RFIP system. Aiming at the defects that according to existing system certification, calculation and storage cost much and are vulnerable to resetting and counterfeit attacks, the bidirectional security authentication method combines pseudo-random numbers, shared secret keys and hash functions to achieve authentication encryption. According to the method, a label and a back-end data base share a secret key, an identification and the two hash functions; a label identification and a logic operation result encrypted by the hash functions of the system serve as response messages to be sent to the back-end data base, so that system authentication expenses are substantially reduced; the back-end data base carries out system hash function encryption on an authentication secret key and a private hash encryption result and responds the authentication secret key and the private hash encryption result to the label, and reverse authentication carried out by the label on the system is achieved. A reader identification does not need to be stored in the label, pseudo-random numbers are needless to be generated, accordingly, cost of the label is reduced, and the application range of the method is enlarged. The method is high in security, low in cost and complexity and capable of being used in environments with large label scales on the premise that basic authentication functions are completed.

The invention discloses a multi-modal emotion analysis method based on a multi-dimensional attention fusion network. The multi-modal emotion analysis method comprises the steps: extracting voice preprocessing features, video preprocessing features and text preprocessing features from sample data containing multiple modals such as voice, video and text; then, constructing the multi-dimensional attention fusion network for each mode; extracting first-level autocorrelation features and second-level autocorrelation features by using an autocorrelation feature extraction module in the network, thencombining the autocorrelation information of the three modes, and obtaining cross-modal fusion features of the three modes by using a cross-modal fusion module in the network; combining the secondaryautocorrelation features and the cross-modal fusion features to obtain modal multi-dimensional features; and finally, splicing the modal multi-dimensional features, determining emotion scores, and performing emotion analysis. According to the method, feature fusion can be effectively carried out in a non-aligned multi-modal data scene, and multi-modal associated information is fully utilized to carry out emotion analysis.

The present invention is an apparatus for cleaning a vacuum drum used for transferring gummed labels. The apparatus operates by sealing the drum and backwashing a cleaning fluid through the vacuum ports to unplug the perimeter holes. The apparatus comprises a cylindrical housing closed on one end by a plate and, on the opposing end, by a detachable lid. A drum seat, the size of which corresponds to the size of the vacuum drum to be washed, is locked to the bottom of the housing. The drum seat has an inlet for receiving the cleaning fluid and means for dispensing the received fluid to the vacuum ports of the vacuum drum under pressure. Initially, the cleaning fluid is heated to dissolve the adhesive. Subsequently, a cooler liquid is flushed through the vacuum drum to cool it for handling.

The invention provides an RFID lightweight-class bidirectional authentication method based on CRC coding. The RFID lightweight-class bidirectional authentication method based on CRC coding comprises the steps that a label extracts the inherent CRC code of the label and provides a randomizer, a random number Rr generated by a reader, a random number Rt generated by the label, a label temporary identifier IDT, secret key information K1 and secret key information K2 are encrypted through simple logical operation and CRC coding operation, so that a random cryptograph changing dynamically is generated and is used as identity authentication information of the reader and the label, and then mutual authentication of the reader and the label is achieved. The RFID lightweight-class bidirectional authentication method based on CRC coding is novel, practical and simple, does not need traditional large-scale data encryption or HASH operation and is suitable for an RFID system with limited hardware capability and limited calculation capability; in addition, by the adoption of the RFID lightweight-class bidirectional authentication method based on CRC coding, combination of high privacy safety and low label cost can be well achieved.

The invention relates to a neural network training method and device, electronic equipment and a storage medium, and the method comprises the steps: carrying out the screening of a plurality of firstsample images, and determining a plurality of labeled second sample images from the plurality of first sample images; Generating a plurality of fourth sample images according to a third sample image marked as a first category in the plurality of second sample images, and marking the fourth sample image as the first category; And training the diagnosis network according to the plurality of labeledsecond sample images and the plurality of fourth sample images. The embodiment of the invention discloses a neural network training method. According to the method, the plurality of first sample images can be screened to obtain the labeled second sample image, the labeling cost can be reduced, the fourth sample image can be generated, the number of the samples of the first category can be increased, the total number of the samples can be increased, the number of the samples of the first category and the second category is balanced, and the training effect is improved.

The invention provides a multi-task classification model training method and device, and a multi-task classification method and device. The multi-task classification model training method comprises the following steps: inputting preset information into a pre-training model, wherein the preset information comprises a plurality of information units; calling a parameter sharing layer, performing global vector representation processing on each information unit, and determining a global semantic representation vector of each information unit; calling a plurality of classifiers, performing classification processing on the preset information according to each global semantic representation vector, and determining a classification prediction result of the preset information; based on the classification prediction result, the first quantity, the second quantity and the labeling result, calculating to obtain a loss value; and under the condition that the loss value is within a preset range, taking a target pre-training model obtained by training as a multi-task classification model. According to the multi-task classification model training method, a good multi-task classification model can be obtained on the basis of a small amount of training data, and only a small amount of annotation training data needs to be added under the condition that a new task exists, and the annotation cost can be reduced.

The invention discloses an image recognition model training method and device, an image recognition method and device, electronic equipment and a storage medium. A large number of marked training samples can be quickly generated, the training cost is reduced, and the training efficiency is improved. The image recognition model training method comprises the steps: extracting a mapping template containing an identifier from a first image containing the identifier; adding the mapping template into the plurality of second images to obtain a plurality of sample images; taking the sample image and alabeling label corresponding to the sample image as training samples and adding the training samples into a training sample set, wherein the labeling label is determined based on an identifier category to which an identifier contained in a mapping template in the sample image belongs; and training an image recognition model based on the training sample set to obtain an image recognition model capable of recognizing the identifier category to which the identifier contained in the image belongs.

The invention provides a method and device for training a machine learning model. The method comprises the following steps: a first stage: inputting an unlabeled sample set, selecting a to-be-labeledsample from the unlabeled sample set through active learning based on the initialized or pre-trained machine learning model, labeling the sample to be labeled, and storing the labeled sample in a labeling data set, dividing the annotation data set into a training data set and a verification data set, training the machine learning model by using the training data set to obtain a trained machine learning model, and verifying the trained machine learning model by using the verification data set to obtain the performance of the trained machine learning model; and a second stage: repeating the steps in the first stage when the performance of the trained machine learning model is less than a predetermined performance index, until the performance of the trained machine learning model is greater than or equal to a predetermined performance index.

The invention belongs to the field of voice recognition, and particularly relates to a phone-call recording access failure reason recognizing method. The method comprises the following steps: markingaccess failure reasons by signals; if reasons cannot be obtained by signal classification, extracting an audio fingerprint characteristic sequence from to-be-recognized phone-call recording, and searching from an audio fingerprint database by the sequence; if matched fingerprints can be found out, marking the access failure reason for to-be-recognized phone-call according to access failure reasonlabels in fingerprint key values; and if the matched fingerprints cannot be found out, recognizing audio contents into text contents by automatic voice recognition, classifying in an access failure document classifying model by a text classifying method on the basis of the contents, and marking the to-be-recognized phone-call recording by access failure reason classifying results obtained by classification. The method can recognize recording files in an offline manner, streaming phone-call voice can also be recognized, the universality is high, and the phone-call recording access failure reason recognizing method is suitable for different application scenarios of the call center.

A taxpayer industry classification method based on noise label learning comprises the steps that firstly, text information to be mined in taxpayer industry information is extracted for text embedding, and feature processing is conducted on the embedded information; secondly, non-text information in the taxpayer industry information is extracted and coded; thirdly, a BERT-CNN deep network structure conforming to the taxpayer industry classification problem is constructed, and the number of layers of the network, the number of neurons of each layer and the input and output dimensions are determined according to the processed feature information and the target category number; then, the constructed network is pre-trained through comparative learning, nearest neighbor semantic clustering and self-label learning in sequence; finally, a noise modeling layer is added on the basis of the constructed deep network, modeling is carried out on noise distribution through network self-trust and noise label information, and model training is carried out based on noise label data; and finally, the deep network in front of the noise modeling layer is taken as a classification model, and taxpayer industry classification is performed based on the model.

The invention discloses a nuclear magnetic resonance image classification method based on multi-strategy batch type active learning and belongs to the field of intelligent medical diagnosis. The method comprises the following steps: obtaining a nuclear magnetic resonance image of the subject as an original data set; randomly selecting K samples from the unlabeled sample set; labeling, constructinga convolutional neural network model and a convolutional auto-encoder model; verifying the convolutional neural network model trained again by using the verification set; obtaining a trained convolutional neural network model; and inputting the unlabeled test set into the trained convolutional neural network model to obtain a final classification result of the nuclear magnetic resonance images ofthe three types of subjects with normal cognition, mild cognitive impairment and Alzheimer's disease, so that redundant information among the screened samples is relieved, and high-quality labeled samples are obtained. On the premise of ensuring high classification accuracy, the labeling cost of the nuclear magnetic resonance image is reduced, and a doctor is efficiently assisted in diagnosing the Alzheimer's disease.

The embodiment of the invention discloses a model training method and system, a server and a storage medium. The method comprises the steps that first sample data with an annotation is utilized to obtain a basic model through training; the basic model is utilized to obtain a reward model through training according to an analysis result of second sample data and feedback of the analysis result corresponding to the second sample data by a user, wherein the reward model is used for evaluating an analysis result of the basic model; and third sample data is utilized to perform feedback training incombination with the basic model and the reward model, the objective of the reward model is set to be forward feedback, the basic model is corrected to be adjusted towards the objective of the user, and the adjusted basic model is obtained. Through the embodiment, the interaction mode in the model training process can be improved, labor cost of data annotation is lowered, and data reusability in different scenes is improved.

The invention discloses a hierarchical semantic embedding model for fine recognition of an object and an implementation method thereof. The hierarchical semantic embedding model comprises a backbone network, which is used for extracting shallow features of an input image and outputting the features to each branch network in the form of a feature map; several branch networks, which are used for further extracting deep-seated features from the image shallow feature map output by the backbone network, so that the output characteristic map is suitable for the identification task of the corresponding hierarchy of the branch network, by introducing the semantic knowledge embedding mechanism, the invention realizes the guidance of the upper semantic knowledge to the feature learning of the lowerbranch network. The invention solves the problem that the additional information labeling cost is high in the object refinement identification technical scheme which relies on the additional information to guide learning.

An image character recognition system comprises a detection model and / or a recognition model. The detection model is used for detecting the position of an area where characters in the image are located. The recognition model is used for extracting character information in the image. An operation module calls the detection module and / or the recognition module to detect and recognize the charactersin the image and obtain a preliminary labeling result. A sample selection module is used for selecting part or all of the samples from the preliminary labeling result. A labeling correction module isused for correcting the labeling result selected by the sample selection module to obtain a refined labeling result, and the refined labeling result is used for continuously training and optimizing the detection model and / or the recognition model.

The invention relates to a named entity recognition method and device based on semi-supervised learning training, computer equipment and a storage medium. The method comprises the steps of obtaining annotation data and non-annotation data; performing supervised training on a sequence labeling model by utilizing the labeling data; calculating semantic vectors corresponding to the annotation data and the unannotated data through a trained sequence annotation model, and identifying the unannotated data in the same distribution as the annotation data according to the semantic vectors; calling a semi-supervised learning model, wherein the semi-supervised learning model is composed of the trained sequence labeling model and an auxiliary prediction network with a limited input view angle; and training the semi-supervised learning model through unlabeled data in the same distribution, and outputting a corresponding named entity recognition result through Viterbi decoding. By adopting the method, the data annotation cost can be effectively reduced, and the named entity identification accuracy can be effectively improved.

The invention provides a data set processing method, a data set processing device and a storage medium, and the data set processing method comprises the steps: marking different regions of target image data through a plurality of different types of marking frames to obtain a first data set, and training the first data set to obtain a state detection model; testing the detection precision of the state detection model, and detecting the to-be-labeled target image data by using the state detection model to obtain image information; labeling the to-be-labeled target image data through the image information to obtain a second data set; and storing the first data set and the second data set in parallel as a to-be-trained data set, and training the to-be-trained data set to obtain a new state detection model. According to the technical scheme, a pure manual annotation data set is not needed, the labor cost is reduced, the iterative optimization efficiency of the state detection model is improved, the annotation data set is more accurate, and the accuracy of the detection model is improved.

The invention provides an image annotation method, device and system and a host, and relates to the technical field of image processing, and the method comprises the steps: obtaining a plurality of to-be-annotated two-dimensional images of a target object at different angles through a first camera; calculating pose parameters of the three-dimensional model of the target object corresponding to each two-dimensional image according to a pose transformation relationship between the three-dimensional model of the target object and the first camera; wherein the three-dimensional model is a model during modeling of the target object or a model constructed based on three-dimensional point cloud data of the target object; acquiring defect labeling information of the three-dimensional model; and projecting defect labeling information of the three-dimensional model to the two-dimensional images according to the pose parameters of the three-dimensional model corresponding to the two-dimensional images to obtain labeling results of the two-dimensional images. The labeling efficiency can be effectively improved, the labeling cost is reduced, and the quality of labeling results can be well unified, so that the defect detection effect of parts can be improved.

The invention discloses an active learning and deep learning combined aluminum material surface defect detection method. The method comprises the steps of performing data enhancement on data in a training set by randomly adjusting image saturation, adjusting image brightness, adjusting image contrast and randomly rotating an image; adopting a Weighted-Entry evaluation standard for active learning,ranking the samples to be labeled according to a Weighted-Entry value in an increasing sequence, selecting K highest samples for labeling, and adding the samples into a training set to serve as training samples; and meanwhile, sequencing the samples to be labeled according to an increasing sequence of Entropy (information entropy) by adopting a pseudo labeling algorithm, selecting the lowest H samples, and taking a prediction result of the model as a pseudo label to serve as additional temporary training data of next training. An SEResNet-152 neural network structure is adopted, the neural network structure is based on a ResNet-152 network model, an SE module is added behind each Resinual module, and the SE module is used for calculating the weight proportion relation between channels ofa feature map.

The invention relates to a search content sorting method and device, a storage medium and electronic equipment, and the method comprises the steps: determining a correlation score between each searchcontent corresponding to a search word and the search word through a pre-trained semantic correlation model; sorting the plurality of search contents according to the correlation score, wherein the training process of the semantic correlation model comprises the following steps: pre-training a language model through a plurality of search term samples and a first search content sample determined according to historical operation behaviors of a user for a plurality of search contents corresponding to each search term sample; and finely adjusting the pre-trained language model through the plurality of search term samples and two second search content samples corresponding to each search term sample, wherein the second search content samples are attached with tags used for representing whetherthe search content samples are related to the search term samples or not. According to the method, the correlation score of the search content can be determined through the pre-trained and fine-tunedsemantic correlation model, the application range of the semantic correlation model is widened, and the annotation cost is reduced.

The invention belongs to the technical field of data detection, and provides an intelligent quality inspection keyword inspection method, a device, equipment, and a readable storage medium. The intelligent quality inspection keyword inspection method comprises following steps: training sample data and pre-labeled keyword data are obtained, and the training sample data is subjected to filter bank characteristic, perceptual linear prediction coefficient characteristic, and sound frequency characteristic extraction; based on the pre-labeled keyword data, a language model and dictionary are constructed; the filter bank characteristic, the perceptual linear prediction coefficient characteristic, and the sound frequency characteristic are subjected to model training, and an acoustic model is constructed; based on the language model and the acoustic model after test processing, keywords of voice data to be tested are subjected to identification, seat business behavior standards are subjectedto scoring, results are output. The intelligent quality inspection keyword inspection method is accurate in keyword identification; each target keyword is supported by a large amount of data sets; atthe same time, model labeling cost is low; identification speed is fast; and the efficiency is increased greatly compared with artificial quality inspection.

The present invention relates to a method for preparing light-emitting diode module. Said method mainly includes the following steps: (1), combination: combining electronic element on the circuit board, and combining identification resistor by light-emitting diode; (2), baking: placing the circuit board on which the above-mentioned combination step is completed into a backwelding oven and baking; (3), test: testing the above-mentioned baked circuit board.

The invention discloses an RFID system secret key generation method and devices based on tag ID. The method comprises the steps that firstly, a reader-writer determines to generate a secret key / secret keys for a tag or a batch of tags or a group of tags, determines the types of the generated secret keys and then sends a secret key generation request command to each tag, each tag feeds back corresponding information of the corresponding tag ID according to type difference, and the reader-writer and each tag generate a secret key shared by the reader-writer and the tag through certain calculation according to the information of the corresponding tag ID. By means of the secret keys, subsequent bidirectional authentication can be conducted. Based on all the types, the three types of secret key generation devices are specifically disclosed. By means of the method, read-write security of an RFID system is improved, the calculated amount at each tag end is greatly reduced, and tag cost is lowered.

The invention relates to a medical image sample screening method and device, computer equipment and a storage medium, for carrying out intelligent screening on unlabeled medical image samples. The medical image sample screening method includes the steps: carrying out model training on a labeled sample set by utilizing a Mask-RCNN model so as to obtain a focus target detection depth model; predicting the unlabeled medical image sample set according to the focus target detection depth model to obtain a prediction result of each medical image sample and judge a labeling value; and selecting a medical image sample with high annotation value to perform annotation confirmation, performing iterative updating on the focus target detection depth model, and ending the iterative updating until the performance of the focus target detection depth model cannot continue to annotate a new sample. According to the medical image sample screening method, under the condition of limited computing resourcesor annotation cost, the high-value small data set is actively mined and extracted, and efficient diagnosis and decision making are achieved by simulating a medical expert intelligent learning mode, and the intelligent degree is high, and the processing speed is high, and the problem that the annotation efficiency is low is effectively solved.

The invention relates to the field of machine learning, and discloses a medical image classification method and device, a medium and electronic equipment. The method comprises the following steps: selecting a target medical image sample from an unlabeled medical image sample set by utilizing an active learning framework, wherein a query strategy of the active learning framework is provided by a reinforcement learning model; inputting the target medical image sample labeled by the labeling expert into a medical image classification model, and training the medical image classification model; ifthe training does not meet the preset condition, obtaining a training result, training a reinforcement learning model based on the training result, updating a query strategy by utilizing the trained reinforcement learning model, and turning to a sample selection step until the training meets the preset condition; and inputting to-be-classified medical image data into the trained medical image classification model for classification. According to the method, a long-acting working mechanism for training the medical image classification model through man-machine cooperation is established, the labeling cost is reduced, and the labeling efficiency is improved.