Microsoft Technology Licensing, LLC

Thomson Licensing S.a.

Dinei Afonso Ferreira Florencio

Cha Zhang

Henrique Sarmento Malvar

Yu-Yun Dai

Zhengyou Zhang

Patrice Y Simard

David W Steinkraus

Bradford W Gillespie

Gil Moshe Nahmias

Jin Li

Philip Andrew Chou

Michel Pahud

Stuart J Golin

Sasa Junuzovic

Rajesh Hegde

Jungwoo Lee

Sharon Kay Cunnington

James E Oker

John Richard Corring

Sean Louis Goldberg

Shih Chia Wang

Guoxin Wang

Wenfeng Cheng

Yijuan Lu

Lynsey Liu

Andrei A Gaivoronski

Yu Cheng

A data processing system implements a method for training machine learning modes, including receiving a set of one or more unlabeled documents associated one or more first categories of documents to be used to train machine learning models to analyze the one or more unlabeled documents, and fine-tuning a first machine learning model and a second machine learning model based on the one or more unlabeled document to enable the first machine learning model to determine a semantic representation of the one or more first categories of document, and to enable the second machine learning model to classify the semantic representations according to the one or more first categories of documents, the first machine learning model and the second machine learning model having been trained using first unlabeled training data including a second plurality of categories of documents that do not include the one or more first categories of documents.

Techniques for pretraining document language models for example-based document classification

The disclosure herein describes providing signature data of an input document. Text data of the input document is obtained (e.g., OCR data generated from image data) and a first set of signature fields are identified using signature key-value pairs of the text data. A first subset of signed signature fields and a first subset of unsigned signature fields are determined based on mapping to a set of predicted values. A second set of signature fields are determined using a region prediction model applied to image data of the input document. Region images associated with the first subset of unsigned signature fields and with second set of signature fields are obtained and a second set of signed signature fields and a second set of unsigned signature fields are determined using a signature recognition model. Signature output data is provided including signed signature fields and/or unsigned signature fields.

Entry detection and recognition for custom forms

Interfaces and systems are provided for harvesting ground truth from forms to be used in training models based on key-value pairings in the forms and to later use the trained models to identify related key-value pairings in new forms. Initially, forms are identified and clustered to identify a subset of forms to label with the key-value pairings. Users provide input to identify keys to use in labeling and then select/highlight text from forms that are presented concurrently with the keys in order to associate the highlighted text with the key(s) as the corresponding key-value pairing(s). After labeling the forms with the key-value pairings, the key-value pairing data is used as ground truth for training a model to independently identify the key-value pairing(s) in new forms. Once trained, the model is used to identify the key-value pairing(s) in new forms.

Enhanced supervised form understanding

The disclosed technology is generally directed to optical character recognition for forms. In one example of the technology, optical character recognition is performed on a plurality of forms. The forms of the plurality of forms include at least one type of form. Anchors are determined for the forms, including corresponding anchors for each type of form of the plurality of forms. Feature rules are determined, including corresponding feature rules for each type of form of the plurality of forms. Features and labels are determined for each form of the plurality of forms. A training model is generated based on a ground truth that includes a plurality of key-value pairs corresponding to the plurality of forms, and further based on the determined features and labels for the plurality of forms.

Supervised OCR training for custom forms

The disclosed technology is generally directed to optical text recognition for forms. In one example of the technology, line grouping rules are generated based on the generic forms and a ground truth for the generic forms. Line groupings are applied to the generic forms based on the line grouping rules. Feature extraction rules are generated. Features are extracted from the generic forms based on the feature extraction rules. A key-value classifier model is generated, such that the key-value classifier model is configured to determine, for each line of a form: a probability that the line is a value, and a probability that the line is a key. A key-value pairing model is generated, such that the key-value pairing model is configured to predict, for each key in a form, which value in the form corresponds to the key.

Unsupervised domain adaptation from generic forms for new OCR forms

A method includes obtaining data representing multiple characters, determining a code for each character wherein each code corresponds to a different audio frequency, and transmitting the codes at the corresponding audio frequencies.

Audio data transmission using frequency hopping

A template and/or knowledge associated with a synchronous meeting are obtained by a computing device. The computing device then adaptively manages the synchronous meeting based at least in part on the template and/or knowledge.

Adaptive meeting management

Techniques and technologies are described herein for motion parallax three-dimensional (3D) imaging. Such techniques and technologies do not require special glasses, virtual reality helmets, or other user-attachable devices. More particularly, some of the described motion parallax 3D imaging techniques and technologies generate sequential images, including motion parallax depictions of various scenes derived from clues in views obtained of or created for the displayed scene.

Three-dimensional (3D) imaging based on MotionParallax

Three-dimensional (3D) imaging based on motionparallax

Systems and methods for performing adaptive filtering are disclosed. The present invention generates probabilities that can be used in an encoder, such as an arithmetic encoder and generates those probabilities in a computationally efficient manner. Probabilities of previously encoded coefficients are employed, effectively, in generating probabilities of the coefficients without regard to directional information. Thus, a large amount of information is adaptively and efficiently used in generating the probabilities. For the coefficients, the probability is computed based at least partly on at least one probability of a previously computed probability of a neighboring coefficient. Then, the coefficients are encoded using those computed probabilities.

Tarp filter

A system and method facilitating signal enhancement utilizing an adaptive filter is provided. The invention includes an adaptive filter that filters an input based upon a plurality of adaptive coefficients and modifies the adaptive coefficients based on a feedback output. A feedback component provides the feedback output based, at least in part, upon a non-linear function of the acoustic reverberation reduced output. Optionally, the system can further include a linear prediction (LP) analyzer and/or a LP synthesis filter. The system can enhance signal(s), for example, to improve the quality of speech that is acquired by a microphone by reducing reverberation. The system utilizes, at least in part, the principle that certain characteristics of reverberated speech are measurably different from corresponding characteristics of clean speech. The system can employ a filter technology (e.g., reverberation reducing) based on a non-linear function, for example, the kurtosis metric.

Microphone array signal enhancement

A system and method facilitating signal enhancement utilizing an adaptive filter is provided. The invention includes an adaptive filter that filters an input based upon a plurality of adaptive coefficients and modifies the adaptive coefficients based on a feedback output. A feedback component provides the feedback output based, at least in part, upon a non-linear function of the acoustic reverberation reduced output. Optionally, the system can further include a linear prediction (LP) analyzer and/or a LP synthesis filter. The system can enhance signal(s), for example, to improve the quality of speech that is acquired by a microphone by reducing reverberation. The system utilizes, at least in part, the principle that certain characteristics of reverberated speech are measurably different from corresponding characteristics of clean speech. The system can employ a filter technology (e.g. reverberation reducing) based on a non-linear function, for example, the kurtosis metric.

A transcoding method for inserting a visual element into an encoded bit stream with minimal decoding of the encoded bit stream. The transcoding method includes the steps of receiving an encoded bit stream containing a picture, partially decoding a copy of the bit stream, determining if a segment of the picture in the bit stream is affected by the insertion of the visual element, re-encoding the segment if the segment is affected by the insertion of the visual element, and outputting the re-encoded bit stream. The encoded and re-encoded bit streams may be MPEG-encoded bit streams. The re-encoded bit stream may be output using a bit rate control process that randomly selects an initial processing point within the re-encoded bit stream, requantizes the re-encoded bit stream using a modified quantization table, and/or utilizes a slice-level stopping criteria for selectively starting or ending bit rate control.

Growing community of inventors

Redmond, WA, United States of America

Dinei Afonso Ferreira Florencio