Adobe, Inc.

Ram Bhushan Agrawal

Ajay Jain

Sachin Soni

Anuj Shara

Deepansh Rawal

Himanshu Mittal

Siddharth Kumar

Medha Kumari

Vishnu Singh

Abhijeet Gaiha

Abhishek Shah

Prasenjit Mondal

Mohit Gupta

Ankit Bal

Ankit Pangasa

Lalit Vohra

Uttam Dwivedi

Ruppesh Nalwaya

Deepanshu Arora

Anand Kumar

Akhilesh Godi

Ankush Das

Anshul Kothari

Saurabh Rihan

Amit Kumar Manocha

Tarun Verma

The present disclosure relates to systems, methods, and non-transitory computer readable media for accurately enhancing optical character recognition with a machine learning approach for determining words from reverse text, vertical text, and atypically-sized text. For example, the disclosed systems segment a digital image into text regions and non-text regions utilizing an object detection machine learning model. Within the text regions, the disclosed systems can determine reverse text glyphs, vertical text glyphs, and/or atypically-sized text glyphs utilizing an edge based adaptive binarization model. Additionally, the disclosed systems can utilize respective modification techniques to manipulate reverse text glyphs, vertical text glyphs, and/or atypically-sized glyphs for analysis by an optical character recognition model. The disclosed systems can further utilize an optical character recognition model to determine words from the modified versions of the reverse text glyphs, the vertical text glyphs, and/or the atypically-sized text glyphs.

Utilizing machine-learning based object detection to improve optical character recognition

In some embodiments, techniques for generating personalized unsubscribe options for emails are provided. For example, a subscription management system can access information associated with an active email subscription and a user profile for a user. The subscription management system can determine a temporary unsubscribe time period for the user based on the information associated with the active email subscription and the profile of the user. The subscription management system can embed a temporary unsubscribe option in an email to be sent to the user. The temporary unsubscribe option can represent the temporary unsubscribe time period. The subscription management system can cause the email to be sent to the user and receiving a selection of the temporary unsubscribe option indicating the temporary unsubscribe time period. The subscription management system can change the active email subscription to a suppressed email subscription associated with the temporary unsubscribe time period.

Personalized temporary unsubscribe mechanisms for email subscriptions

Disclosed systems and methods determine a reading order of an electronic document. In an example, a document processing application accesses a content stream that includes a first object with a first location within the electronic document, a second object with a second location within the electronic document, and a third object with a third location within the electronic document. The application computes a region that includes the first object and the second object by determining that the second object is adjacent to the first object in a first dimension, at least a portion of the first and second objects are aligned in a second dimension, and the boundaries of the region do not intersect or encompass the third object. Based on the objects in the region, the application determines that a reading order should include the first object and the second object and should exclude the third object.

Facilitating dynamic document layout by determining reading order using document content stream cues

Methods and systems are provided for providing directional assistance to guide a user to position a camera for centering a person's face within the camera's field of view. A neural network system is trained to determine the position of the user's face relative to the center of the field of view as captured by an input image. The neural network system is trained using training input images that are generated by cropping different regions of initial training images. Each initial image is used to create a plurality of different training input images, and directional assistance labels used to train the network may be assigned to each training input image based on how the image is cropped. Once trained, the neural network system determines a position of the user's face, and automatically provides a non-visual prompt indicating how to center the face within the field of view.

Directional assistance for centering a face in a camera field of view

In some embodiments, techniques for generating personalized unsubscribe options for emails are provided. For example, a subscription management system can access information associated with an active email subscription and a user profile for a user. The subscription management system can determine a temporary unsubscribe time period for the user based on the information associated with the active email subscription and the profile of the user. The subscription management system can embed multiple unsubscribe options in an email to be sent to the user. The multiple unsubscribe options can represent respective unsubscribe time periods including the temporary unsubscribe time period. The subscription management system can cause the email to be sent to the user and receiving a selection among the multiple unsubscribe options indicating a selected unsubscribe time period. The subscription management system can change the active email subscription to a suppressed email subscription associated with the unselected unsubscribe time period.

Techniques are disclosed for detecting the bounds of borderless open tables in fixed-format structured documents, such as PDF documents, and grouping text lines into predicted borderless tables. The target document comprises a set of text lines each having a respective vertical and horizontal position in the target document. A sorted list of the text lines is generated based upon a vertical and horizontal position of each text line in the target document. For each text line in the sorted list, a respective probability that the text line in the sorted list belongs to a borderless table is then determined. According to one embodiment, the probability may be determined using a classifier that may employ a logistic regression algorithm.

Detecting the bounds of borderless tables in fixed-format structured documents using machine learning

Certain embodiments involve a model for enhancing text in electronic content. For example, a system obtains electronic content comprising input text and converts the electronic content into a grayscale image. The system also converts the grayscale image into a binary image using a grid-based grayscale-conversion filter, which can include: generating a grid of pixels on the grayscale image; determining a plurality of grid-pixel threshold values at intersection points in the grid of pixels; determining a plurality of estimated pixel threshold values based on the plurality of grid-pixel threshold values; and converting the grayscale image into the binary image using the plurality of grid-pixel threshold values and the plurality of estimated pixel threshold values. The system also generates an interpolated image based on the electronic content and the binary image. The interpolated image includes output text that is darker than the input text. The system can then output the interpolated image.

Text enhancement using a binary image generated with a grid-based grayscale-conversion filter

Techniques are disclosed for generating a shadow map of a digital image. In some examples, a method may include generating a shadow mask of a digital image, generating a dilated de-noised binarized gradient image based on the shadow mask, generating a binarized median-filtered gray image based on the digital image and the dilated de-noised binarized gradient image, and generating a shadow map based on the shadow mask and the binarized median-filtered gray image. The generated shadow map can then be used to remove shadows from the digital image without degrading the quality of the image content in the digital image.

Removal of shadows from document images while preserving fidelity of image contents

The page straightening system includes a word module to determine an enclosing quadrilateral of each connected component of curved page content. Further, a line module in the page straightening system is configured to form text lines by joining enclosing quadrilaterals based on a reading order. Subsequently, a correction module in the page straightening system is configured to generate straightened content from the curved content based on the text lines. As such, the page straightening system can automatically straighten curved page content.

System and method for straightening curved page content

Systems and methods are disclosed for content aware digital image enhancement. In particular, in one or more embodiments, the disclosed systems and methods analyze content of a digital image portraying a document with graphics and/or text and generate a cleaning confidence map. Specifically, in one or more embodiments, the disclosed systems and methods generate a cleaning confidence map indicating a likelihood that each pixel in the digital image portrays text or a graphic. Moreover, in one or more embodiments, the disclosed systems and methods utilize the cleaning confidence map as a reflection of how aggressively to modify digital images. In particular, in one or more embodiments, the disclosed systems and methods utilize the cleaning confidence map to remove shadows, identify and clean background pixels, and correct contrast in relation to the digital image.

Content aware improvement of captured document images

Disclosed are various embodiments for digitally restoring content from a torn document. A computing device obtains an image of a document that has a tear. The computing device identifies portions of the image separated by a gap region representing the tear in the document. Content in the portions of the image adjacent to the gap region is then identified. The position of a portion of the image is incrementally adjusted to reduce the gap region without overlapping the portions. The position of the portion is adjusted to align the content along a boundary between the portions of the image.

Techniques for restoring content from a torn document

Object boundary detection techniques are described. In implementations, edges of an object displayed in an image are detected. The image is used to generate a gradient image and a monochrome image. Directional filters are applied to the monochrome image to generate directionally filtered images, which are then applied the gradient image to identify potential boundary lines of the object displayed in the image. A selection can then be made from the potential boundary lines, based on a score associated with each respective potential boundary line, to select lines to represent the boundaries of the object. The selected lines can be used to segment the image into a plurality of segments, and then, using the average color of each segment, one or more of the selected lines can be invalidated as being a false boundary.

Growing community of inventors

Noida, India

Ram Bhushan Agrawal