Hewlett-packard Development Company, L.p.

Jian Fan

Hui Chao

Daniel R Tretter

Jerry J Liu

Qian Lin

Suk Hwan Lim

Yuhong Xiong

Samson J Liu

D Amnon Silverstein

Xiaofan Lin

Ping Luo

Su Wang

Yang Lei

Jian-Ming Jin

Li-Wei Zheng

Steven J Simske

Jun Zeng

Tong Zhang

Kar-Han Tan

Ullas Gargi

Jinman Kang

Steven Rosenberg

Pere Obrador

Jonathan Yen

Peng Wu

Nina Bhatti

Yining Deng

Prakash D Reddy

Charles Chi Jia

Clayton Atkins

Lei Liu

Yun David Tang

Kar Han Tan

Luca Chiarabini

Parag Mulendra Joshi

Hui-Man Hou

Divya Sharma

Sheng-Wen Yang

Anne-Marie Woodson

Shridhar Diwan

Ruiting Shao

Lei He

Neerja Raman

A deep learning method employs a neural network having three sub-nets to classify and retrieve the most similar 3D model of an object, given a rough 3D model or scanned images. The most similar 3D model is present in a database and can be retrieved to use directly or as a reference to redesign the 3D model. The three sub-nets of the neural network include one dealing with object images and the other two dealing with voxel representations. Majority vote is used instead of view pooling to classify the object. A feature map and a list of top N most similar well-designed 3D models are also provided.

Three-dimensional shape classification and retrieval using convolutional neural networks and majority vote

A method of recognizing an object includes comparing a three-dimensional point cloud of the object to a three-dimensional candidate from a dataset to determine a first confidence score, and comparing color metrics of a two-dimensional image of the object to a two-dimensional candidate from the dataset to determine a second confidence score. The point cloud includes a color appearance calibrated from a white balance image, and the color appearance of the object is compared with the three-dimensional candidate. The first or second confidence score is selected to determine which of the three-dimensional candidate or the two-dimensional candidate corresponds with the object.

Object recognition

In an example implementation, a method of processing a 3D object includes receiving a 3D object model representing a new 3D object to be processed. The method includes computing attributes of the object and comparing the object attributes with stored object attributes from previously processed 3D objects. When the comparison provides a match between the object attributes and stored object attributes, then a print recipe associated with the matching stored object attributes is retrieved.

Processing 3D objects

Examples herein disclose capturing an image of printed text and a handwritten annotation and determining a topic as related to the printed text in the captured image. The examples disclose identifying a resource based on the handwritten annotation.

Identifying a resource based on a handwritten annotation

According to one example for segmenting image data, image data comprising color pixel data, IR data, and depth data is received from a sensor. The image data is segmented into a first list of objects based on at least one computed feature of the image data. At least one object type is determined for at least one object in the first list of objects. The segmentation of the first list of objects is refined into a second list of objects based on the at least one object type. In an example, the second list of objects is output.

Image data segmentation

One example provides a system. The system receives an infrared image and processes the infrared image to detect lines in the infrared image. The system receives a color image corresponding to the infrared image and processes the color image to detect lines in the color image. The detected lines in the infrared image and the detected lines in the color image are combined. A document region is detected from the combined detected lines.

Document region detection

A system and method for processing an image of a retail shelf area. According to one aspect, a portion of an image that depicts the retail shelf area at an angled and depthwise perspective is determined to depict shelves. The image can be modified so that the portion of the image determined to depict shelves is substantially parallel to a predefined horizontal reference frame. Products can then be detected in the remainder portion of the image, separate from the portion of the image determined to depict shelves. An applicable planogram can be used to generate compliance information based on the detected products and the determined portion of the image depicting shelves.

Image processing of a retail shelf area

In one example, a method of automatically enhancing a digital image having a smooth-boundaried foreground object. The method processes the image via gradient-based background segmentation at a first scale to generate a first mask. The first mask has a first accuracy. The first mask is used to generate a first background of the image which includes at least a portion of the smooth-boundaried foreground object. The method further processes the first background via gradient-based background segmentation at a smaller second scale to generate a second mask. The second mask has a greater second accuracy. The second mask is used with the first mask to generate a second background of the image which substantially excludes the smooth-boundaried foreground object. The image can be enhanced using the second background so as to preserve the smooth-boundaried foreground object.

Preserving smooth-boundaried objects of an image

A method for producing web page content includes identifying blocks within a web page. The blocks are selectively assembled into sections. The sections are selectively assembled into article candidates. An article candidate that includes article content is distinguished from article candidates that do not include article content. Content is produced only from the article candidate distinguished as including article content.

Producing web page content

Semantically ranking content in a website () with a computerized ranking device () includes: parsing content from the website () into multiple autonomous content blocks (-to-) with the computerized ranking device () and assigning an importance ranking with said computerized ranking device () to each of the content blocks (-to-) based on a degree to which a substance of the content block (-to-) is relevant to one of a plurality of predefined categories.

Semantically ranking content in a website

A system and method of detecting separator lines in a web page may include determining coordinates of visible web elements on a web page, generating an edge image of the web page based on the coordinates of the web elements, filtering edges belonging to non-separator line elements within the edge image, detecting horizontal lines within the edge image, detecting vertical lines within the edge image, and filtering short lines within the edge image. A system for detecting separator lines in a web page may include a memory device, and a processor communicatively coupled to the memory, in which the processor determines coordinates of visible web elements on a web page, generates an edge image of the web page based on the coordinates of the web elements, filters edges belonging to non-separator line elements within the edge image, detects horizontal lines within the edge image, detects vertical lines within the edge image, and filters short lines within the edge image.

Detecting separator lines in a web page

A method and system for extracting Web content is disclosed. In one embodiment, Web content in a Webpage is extracted by identifying paragraphs in the Web content based on line-break node determination. A range of text-body associated with the identified paragraphs is then identified using a maximum scoring subsequence. Further, the identified text-body is refined using a heuristic rule of substantially horizontal alignment. Furthermore, one or more titles and one or more images associated with the Web content are extracted. Moreover, the Web content including the identified paragraphs, the one or more titles and the one or more images are outputted.

System and method for web content extraction

A system and method for segmenting foreground and background regions on a digitized image uses a computer, having a processor and system memory, to segment the image into initial regions and identify background regions from the initial regions. A complete background surface is estimated of the image, and pixels of the image are rectified with the estimated background surface to normalize the image. Normalized pixels are compared with a threshold color to determine a final segmentation of background regions.

System and method of foreground-background segmentation of digitized images

Methods, apparatus, and machine-readable media for segmenting and enhancing images are described. In one aspect, gradient magnitude values at respective pixels of a given image are determined. The gradient magnitude values are thresholded with a global threshold to produce thresholded gradient magnitude values. The pixels are segmented into respective groups in accordance with a watershed transform of the thresholded magnitude values. A classification record is generated. The classification record labels as background pixels ones of the pixels segmented into one of the groups determined to be largest in size and labels as non-background pixels ones of the pixels segmented into any of the groups except the largest group.

Gradient based background segmentation and enhancement of images

Systems and methods according to the present invention provide techniques to automatically crop an image to a desired size. Automating the cropping process, while at the same time providing a high quality, cropped image, can be accomplished by employing a comprehensive approach that includes one or more of contextual cropping (e.g., basing cropping decisions on elements within the image to be cropped itself), training the cropping mechanism (e.g., using previously cropped versions of the same image or related images) and multiple operational modes (e.g., selecting one of a plurality of different cropping techniques for cropping a particular image).

Growing community of inventors

San Jose, CA, United States of America

Jian Fan