30 results about "Labeling Problem" patented technology

A method of labeling pixels in an image is described where the pixel label is selected from a set of three or more labels. The pixel labeling problem is reduced to a sequence of binary optimizations by representing the label value for each pixel as a binary word and then optimizing the value of each bit within the word, starting with the most significant bit. Data which has been learned from one or more training images is used in the optimization to provide information about the less significant bits within the word.

The invention relates to the field of artificial intelligence, in particular to a text labeling method and device, electronic equipment and a storage medium, wherein the method comprises the steps: obtaining a to-be-labeled text and an actual language of the to-be-labeled text; obtaining a trained text labeling model corresponding to the actual language and the target language; carrying out semantic space conversion processing on the actual semantic features of the to-be-labeled text through the trained text labeling model, and acquiring target semantic features of the to-be-labeled text in a target language; performing clustering processing on the target semantic features through a trained text labeling model to obtain a clustering result of the to-be-labeled text; performing classification processing on the clustering result through a trained text labeling model to obtain text type information of the to-be-labeled text; and according to the text type information, performing text type labeling on the to-be-labeled text. According to the method and the device, the text type of any language text can be automatically labeled on the basis of the text in the target language with a large number of labeled samples, so that the cross-language text labeling problem is solved.

The invention discloses a semi-supervised industrial product flaw detection method and system based on positive sample learning. The method comprises the following steps: training an image restoration network and a flaw segmentation prediction network through positive sample images; and inputting an image containing a flaw to be detected into the image restoration network to obtain a restored certified product image, calculating an absolute value of a difference value between the two images, splicing the three images to obtain a detection tensor, generating a segmentation mask binary image according to the detection tensor through the segmentation prediction network, and obtaining a flaw region. According to the method, a semi-supervised deep learning method is used, defect position data of a pixel level does not need to be labeled in advance, the labeling problem in defect segmentation is solved, and full-automatic training learning, defect segmentation and defect classification can be realized only by using part of certified product images in production. And pixels at various types of abnormal flaw positions can be segmented.

The invention discloses an automatic massive-text labeling method based on exception handling. The automatic massive-text labeling method based on the exception handling comprises the following steps of S1, extracting named entities and related named entity pairs from a knowledge base and respectively storing the named entities and the related named entity pairs; S2, finding the named entities of all sentences in a massive amount of texts through string matching according to the stored named entities, finding co-occurring sentences of all named entity pairs in the texts according to the stored named entity pairs and performing overstriking labeling; S3, judging the named entities in the sentences subjected to overstriking labeling, deleting abnormal word pairs when the abnormal word pairs exist, using a filter algorithm to filter out all abnormal named entities in the named entities when the abnormal named entities exist, and finally obtaining final text labeling results. The automatic massive-text labeling method based on exception handling can effectively solve the mistaken labeling problem in the labeling process and improves the text labeling accuracy rate.

The invention discloses a video scene detection labeling method and system, and the method comprises the steps: obtaining the modal features of a video, an audio and a text through a pre-training model according to a modal information source embedded by the input video, the audio and the text, carrying out the alignment and fusion of the obtained modal features of the video, the audio and the text, and forming a window basic cross-modal representation, according to the multi-temporal attention and the difference between the adjacent windows, the basic cross-modal representation of the windows is evolved into self-adaptive context sensing representation, the scene is detected according to the obtained self-adaptive context sensing representation, and the attributes of the windows are determined through a window attribute classifier; obtaining an accurate position of a scene boundary in the window through a position offset regression device; and based on the obtained scene boundaries, specifying a plurality of labels for each scene to realize scene labeling, attributing scene detection into window attribute classification and position offset regression, and solving the multi-label labeling problem through integrated learning of two-stage classifiers. The problems of error propagation and huge calculation cost are solved through a unified network of cross-modal clues; scene detection is attributed to window attribute classification and position offset regression, and the multi-label labeling problem is solved through ensemble learning of two-stage classifiers.

The invention discloses a multi-label character relationship automatic labeling method based on event remote supervision, which includes the following steps: collecting events that affect the character relationship; making a corresponding event labeling template; building an event template knowledge base; and performing data preprocessing; The preprocessed sentences use the event labeling template to mark events; carry out character relationship labeling; character relationship labeling results; calculate the credibility between "event" and "sentence"; calculate the sentence score; set a threshold, the Sentences with a sentence score below the threshold are discarded; the final character relationship annotation is obtained. The present invention matches the event template knowledge base to obtain the events that affect the relationship between characters, and then automatically deduces the multi-label relationship of characters according to multiple events, which can solve the multi-label problem of character relationship, can significantly improve the accuracy of multi-label character relationship labeling, and has more advantages. Excellent mobility.

The invention discloses an inspiration material intelligent question answering system and method, and the system comprises a knowledge base module which stores a plurality of different types of inspiration materials, and pastes a corresponding label on each material; an question library module which comprises a plurality of standard questions, and a mapping relation is established between each standard question and the corresponding inspiration material; an input module which is used for acquiring input information; a selection module; the selection module is used for selecting a creation period and extracting different styles of inspiration materials according to the creation period. A retrieval module; and the retrieval module retrieves the inspiration materials in the knowledge base module according to the input information and the creation period and is used for acquiring the required inspiration materials. According to the method and the system, the knowledge graph for the inspiration materials is constructed, related materials in the knowledge graph are retrieved according to problems of designers, and the innovation efficiency of the designers can be improved through retrieval based on the knowledge graph.

The disclosure provides an approach for image segmentation from an uncalibrated camera array. In one aspect, a segmentation application computes a pseudo depth map for each frame of a video sequence recorded with a camera array based on dense correspondences between cameras in the array. The segmentation application then fuses such pseudo depth maps computed for satellite cameras of the camera array to obtain a pseudo depth map at a central camera. Further, the segmentation application interpolates virtual green screen positions for an entire frame based on user input which provides control points and pseudo depth thresholds at the control points. The segmentation application then computes an initial segmentation based on a thresholding using the virtual green screen positions, and refines the initial segmentation by solving a binary labeling problem in a Markov random field to better align the segmentation with image edges and provide temporal coherency for the segmentation.

The invention discloses a deep learning-based marking method for patient problem parts, comprising the following steps: collecting a three-dimensional depth image, analyzing the three-dimensional depth image, and establishing a sample database; marking the three-dimensional depth image in the database; constructing and describing the patient's problem Dimensional features of the site; training a binary classification model; using the problem site detection model to mark the problem site of the patient in real time, which solves the technical problem of the risk of incorrectly identifying the surgical site in the prior art.

The invention discloses a Google user map text labeling method based on an SVG. The method includes the steps that an SVG object with a path text is added to a self-defining overlaying layer of a Google map, and when the overlaying layer is drawn, a current map projection is acquired through a getProjection method; a labeled path represented by longitudes and latitudes is converted into a pixel path under the current projection; the length of the pixel path is calculated and compared with the length of a labeled text, if the pixel path is too short, the SVG object is not displayed; the starting point position of the labeled text on the pixel path is determined so that it can be guaranteed that the text is located in the middle of the path; the space, actually occupied by the labeled text, of the pixel path is determined according to the length of the text; the minimum coordinate value of the labeled text on the pixel path is determined through comparison and serves as a top left corner coordinate of the SVG object; the pixel path of the labeled text relative to the SVG top left corner is calculated; the text pixel path expressed by coordinate strings is converted into a path of path elements in the SVG, and d attributes of the path are replaced. The text labeling problem is solved, and a new way is found for Google user map text labeling.

The present invention provides a dual-mode iterative denoising anomaly detection method and terminal based on video weak labels, which regard weak labels as the noise of accurate labels, and perform label denoising from image control and feature space respectively; Image space to learn the characteristics of normal and abnormal videos; use graph convolutional models to learn the characteristics of video clips at different times; use iteration to update classifiers and denoisers alternately. The present invention overcomes the difficulty of labeling data by fully considering the problem of weak labeling of video, and uses the method of denoising model. Facing the research field of abnormal detection, which is difficult to collect data, it has strong robustness and can solve the problem of video very well. Weak labeling problem with general applicability.

The invention discloses a data stream monitoring method and device, a storage medium and equipment, which can be applied to the field of Internet of Things, and the method comprises the following steps: after receiving a query request sent by a user, analyzing the query request to obtain a problem data text; obtaining a state label of the problem data text from a pre-constructed data table, wherein the state label comprises a system number of the problem system and a change type of the problem data; the change type is used for indicating a change operation executed by the problem system on the problem data; acquiring a transmission path of the problem data from a pre-constructed data table relation graph; and displaying the state label and the transmission path to a user through a preset interface. According to the method, based on the state label of the problem data text and the transmission path of the problem data, a user can be effectively helped to locate the source of the problem data, namely search the system number of the problem system, the user does not need to traverse project systems one by one to search the source of the problem data, and the search time of the problem data is effectively shortened.

The present invention proposes an algorithm for automatically generating movie labels based on film reviews. The algorithm in the present invention fully considers the missing problem existing in the data set of currently labeled movies. First, an unsupervised algorithm with weights is used to automatically train Set complementary labels. At the same time, the present invention also fully considers the relationship between the similarity of the attributes of the two movies and the similarity of the labels, and predicts the mapping from the attributes to the similarity of the labels with the method of machine learning, instead of using simple similarities such as cosine Similarity calculates a rough similarity relationship. Finally, after using the traditional K-nearest neighbor algorithm to obtain multiple sets of candidate labels, this method does not use simple evaluation criteria to sort and select the label set, but uses a graph algorithm based on the co-occurrence relationship of labels to determine the order of candidate labels, so that Decide on the final label set.

The invention discloses an image sequence category labeling method based on a mixed graph model, which includes: performing superpixel segmentation on the image sequence, and performing feature description on the superpixels; performing nearest neighbor matching on superpixels between two consecutive frames of images; based on The adjacency relationship between the superpixels of a single frame image in the spatial domain and the matching relationship in the time domain between the superpixels of multi-frame images are used to model the global optimization model of the image sequence category labeling problem using the mixed graph model; the global optimization problem is solved using a linear method, and we get Class labels of superpixels in continuous multi-frame images; compared with previous graph models, the hybrid graph model constructed by the present invention can describe the first-order and symmetric relationship between superpixels in a single frame image, and can also describe two consecutive frames of images The high-order, asymmetric relationship between superpixels is solved by a linear method, effectively endowing each superpixel of the image sequence with a class label with better temporal consistency and higher accuracy.