Abstract: Provided is a model generation system for generating a text line recognition model that recognizes a text line included in a text line image, the model generation system including a processor section, in which the text line recognition model includes a visual feature extractor and a language context relation network, the processor section determines a variable of the language context relation network by acquiring text data for training and thus training the language context relation network by using the acquired text data, determines a variable of the visual feature extractor by training the text line recognition model through the use of a labeled text line image while the variable of the language context relation network is fixed, and generates the text line recognition model while the variable of the language context relation network is set to the determined variable thereof and the variable of the visual feature extractor is set to the determined variable thereof.

Abstract: A method of and system for providing accessibility for digital ink is carried out by detecting digital ink input on a user interface (UI) screen, the digital ink input being received via a writing instrument used by a user within a given time period; capturing an image of the digital ink input; examining the image of the digital ink input to detect tremors in the user; assigning a level to the tremors based on discontinuities or irregularities in the digital ink input within the given time period; determining based on the level that the digital ink input requires correction; and applying a correction to the digital ink input to correct the digital ink input.

Abstract: A method includes receiving a binary annotation of source text; performing a close operation on the binary annotation to generate a closed annotation using an initial kernel size; defining one or more contours in the closed annotation using one or more bounding boxes, respectively; determining a subset of the one or more contours for which a percentage of area occupied by text within a corresponding bounding box exceeds a threshold; and generating a final annotation of the source text based on the subset of the one or more contours.

Abstract: A system for authenticating a subject identification document includes instructions causing a processor to receive a first image of the document, depicting, in RGB color space, the document illuminated by UV light, and transform the first image to a second image in YCbCr color space. The instructions cause the processor to divide the second image into regions of interest (ROIs), and, for each ROI, generate a color profile descriptor including first data elements associated with pixel color in the YCbCr color space, and second data elements associated with pixel intensity of pixels within the ROI. The instructions also cause the processor to generate a score for the second image based on a comparison of the color profile descriptor for each ROI to a corresponding ROI of an enrolled document template, and compare the score to a threshold to determine whether the document is authenticated.

Abstract: The technology disclosed relates to equalizer-based intensity correction for base calling. In particular, the technology disclosed relates to accessing an image whose pixels depict intensity emissions from a target cluster and intensity emissions from additional adjacent clusters, selecting a lookup table that contains pixel coefficients that are configured to increase a signal-to-noise ratio, applying the pixel coefficients to intensity values of the pixels in the image to produce an output, and base calling the target cluster based on the output.

Abstract: Described is a system for anomaly detection from hyperspectral data. The system receives, from a hyperspectral sensor, input hyperspectral data cubes with input data bands for every pixel. Using an optimization technique, an optimized number of virtual bands is determined to use for combining the hyperspectral data cubes to achieve optimal anomaly detection performance. The optimized number of virtual bands is less than the input data bands, resulting in reduced hyperspectral data. The hyperspectral data cubes are combined into the optimized number of virtual bands with the optimal bandwidth for each virtual band, and a set of combined hyperspectral data is output. Anomalies are detected in the set of combined hyperspectral data using an anomaly detection technique. Based on the detected anomalies, single-pixel or subpixel targets of interest are detected in the combined hyperspectral data.

Abstract: A computer-implemented method is provided. The method includes: obtaining a sample text and a sample image corresponding to the sample text; labeling a true semantic tag for the sample text according to a first preset rule; obtaining a text feature representation of the sample text and a predicted semantic tag output by a text coding sub-model; obtaining an image feature representation of the sample image output by an image coding sub-model; calculating a first loss based on the true semantic tag and the predicted semantic tag; calculating a contrast loss based on the text feature representation of the sample text and the image feature representation of the sample image; adjusting parameters of the text coding sub-model based on the first loss and the contrast loss; and adjusting parameters of the image coding sub-model based on the contrast loss.

Abstract: Computer-readable media, methods, and systems are disclosed for applying machine learning mechanisms to classify and validate documents based on expense rule sets and external data validation services. Document images associated with expenses are received in connection with a reimbursable event. For each received document image data associated with the received document image is transmitted to an optical character recognition image processor that can recognize contents and associated coordinates. OCR data is received and transmitted to a text tokenizer. Tokenized text is received corresponding to expense details, and the tokenized text and coordinates are sent to a text feature generator. Text feature vectors are received and transmitted to a document classifier and a document classification received. Document fields are extracted and based thereon a document is validates and a corresponding reimbursement instruction generated.

Abstract: According to an embodiment, an absolute scale depth calculation device includes a captured image acquisition unit, an ambiguous scale depth deriving unit, a position and posture acquisition unit, and an absolute scale depth deriving unit. The captured image acquisition unit is configured to acquire a plurality of captured images at different imaging time points from an imaging unit mounted on a moving body. The ambiguous scale depth deriving unit is configured to derive ambiguous scale depth information from a captured image. The position and posture acquisition unit is configured to acquire absolute scale position and posture information pertaining to the imaging unit when each of the plurality of captured images is captured. The absolute scale depth deriving unit configured to derive absolute scale depth information, based on a geometrical consistency from the plurality of captured images, the ambiguous scale depth information, and the absolute scale position and posture information.

Abstract: A road slope predicting method, a device, and a storage medium are disclosed. The road slope predicting method includes obtaining a road image of a road by a camera, detecting a first image lane line and a second image lane line of the road from the road image, setting a number of road image segmentation points along a road image center line of the first image lane line and the second image lane line respectively by means of the corresponding road image segmentation points, determining the pitch angle of the camera with respect to the road plane at each road space segmentation point; based on the internal and external parameters of the camera and the pitch angle, a space coordinate of each road space segmentation points is calculated in a recursive man. A road model of the road is constructed based on each space coordinate.

Abstract: Systems and methods for de-smoking images of a surgical scene are described. Methods include receiving a video of a surgical scene including an image frame. Methods include determining that the image frame includes a smoke occlusion. Methods include determining an estimated un-occluded color of one or more pixels of the image frame using a lookup table, the lookup table mapping between a color space and a set of color bins including the estimated un-occluded color. Methods include determining a respective estimated true color for the one or more pixels of the subset using the imaged color, the estimated un-occluded color, and the smoke color. Methods also include generating a de-smoked image frame using the respective estimated true colors of the one or more pixels, the de-smoked image exhibiting a reduction of the smoke occlusion relative to the image frame.

Abstract: Vehicle perception techniques include applying a 3D DNN to a set of inputs to generate 3D detection results including a set of 3D objects, transforming the set of 3D objects onto a set of images as a first set of 2D bounding boxes, applying a 2D DNN to the set of images to generate 2D detection results including a second set of 2D bounding boxes, calculating mean average precision (mAP) values based on a comparison between the first and second sets of 2D bounding boxes, identifying a set or corner cases based on the calculated mAP values, and re-training or updating the 3D DNN using the identified set of corner cases, wherein a performance of the 3D DNN is thereby increased without the use of expensive additional manually and/or automatically annotated training datasets.

Abstract: A sentiment analysis system is described that uses a machine learning model to predict a mood of an individual based on the motions used by the individual when performing sign language. The present disclosure uses image processing to monitor how a non-verbal individual communicates to, for example, predict overall health and mental well-being.

Abstract: This disclosure provides methods, devices, and systems for low-light imaging. In some implementations, an image processor may be configured to reduce or remove noise associated with an image based, at least in part, on a neural network. For example, the neural network may be trained to infer a denoised representation of the image. In some aspects, the image processor may scale the brightness level of the image to fall within a normalized range of values associated with the neural network. In some other aspects, a machine learning system may scale the brightness levels of input images to match the brightness levels of ground truth images used to train the neural network. Still further, in some aspects, the machine learning system may scale the brightness levels of the input images and the brightness levels of the ground truth images to fall within the normalized range of values during training.

Abstract: The disclosed technique detects a foreground region accurately even with a neural network. A first region detection unit detects, by a DNN, a predetermined foreground region in an inputted image. A weak region identification unit detects, by a DNN, a weak region having a possibility that misdetection or non-detection occurs during the region detection on the inputted image using a DNN that has the same detection target as that of the DNN of the first region detection unit. A second region detection unit detects, by a method other than a neural network, a foreground region that is a detection target same as that of the first region detection unit in the weak region detected by the weak region identification unit within the inputted image read from an input device. An integration unit integrates detection results by the first region detection unit and by the second region detection unit.

Abstract: The present invention provides a food and/or beverage item counting device to be provided in a food and/or beverage item provision system including a transport path that passes along a customer table to transport carriers each configured to allow a food and/or beverage item to be placed thereon, the food and/or beverage item counting device including: a first information acquiring unit disposed upstream of the table and configured to acquire information relating to each of the carriers on the transport path; a second information acquiring unit disposed downstream of the table and configured to acquire information relating to each of the carriers on the transport path; and a first calculating unit configured to calculate the number of the food and/or beverage items taken out from the transport path to the table, wherein each of the carriers is provided with identification information for identifying the carrier.

Abstract: Systems, methods, and computer program products for performing semi-supervised contrastive learning of visual representations are provided. For example, the present disclosure provides systems and methods that leverage particular data augmentation schemes and a learnable nonlinear transformation between the representation and the contrastive loss to provide improved visual representations. Further, the present disclosure also provides improvements for semi-supervised contrastive learning.

Abstract: A method for self-calibrating alignment between image data and point cloud data utilizing a machine learning model includes receiving, with an electronic control unit, image data from a vision sensor and point cloud data from a depth sensor, implementing, with the electronic control unit, a machine learning model trained to: align the point cloud data and the image data based on a current calibration, detect a difference in alignment of the point cloud data and the image data, adjust the current calibration based on the difference in alignment, and output a calibrated embedding feature map based on adjustments to the current calibration.

Abstract: Provided herein are systems and methods for extracting text from a document. Different optical character recognition (OCR) tools are used to extract different versions of the text in the document. Metrics evaluating the quality of the extracted text are compared to identify and select higher quality extracted text. A selected portion of text is compared to a threshold to ensure minimal quality. The selected portion of text is then saved. Error correction can be applied to the selected portion of text based on errors specific to the OCR tools or the document contents.

Abstract: Techniques are described for automating and computerizing order cart audits to reduce the overall costs associated with shipping incorrect items, missing items, and/or additional items. In some cases, the system may be configured to perform an audit of a completed or filled order cart. The facility operator may pass the order cart through an audit area prior to loading the items on a vehicle for transport. During the audit, the system may capture sensor data associated with the order cart and identify each item present. The system may determine whether or not each item is part of the order and notify an operator accordingly.