Microsoft Technology Licensing, LLC

Beijing Baidu Netcom Science Technology Co., Ltd

Jingdong Wang

Shipeng Li

Xian-Sheng Hua

Hao Xu

Jing Wang

Haifeng Wang

Errui Ding

Tian Wu

Junyu Han

Kun Yao

Yinghai Zhao

Jian Wang

Hao Sun

Peng Wang

Qi Chen

Tao Mei

Qifa Ke

Linjun Yang

Zhigang Wang

Xiaoyan Wang

Chengquan Zhang

Shanshan Liu

Yang Wang

Jingtuo Liu

Dongdong Chen

Zhong Wu

Yipeng Sun

Pengyuan Lv

Xiangbo Su

Xiongwei Zhu

Liang Wu

Zengzhong Li

Jianmin Bao

Jeffrey Song Zhu

Xinmei Tian

Dengpan Fu

Mengjun Cheng

Longchao Wang

Qiman Wu

Yangliu Xu

Meina Qiao

Mingqin Li

Nilesh N Yadav

Shi Zhang

Han Zhang

Nobuyuki Morioka

Huaizu Jiang

You Jia

The present disclosure provides a character recognition model training method and apparatus, a character recognition method and apparatus, a device and a medium, relating to the technical field of artificial intelligence, and specifically to the technical fields of deep learning, image processing and computer vision, which can be applied to scenarios such as character detection and recognition technology. The specific implementing solution is: partitioning an untagged training sample into at least two sub-sample images; dividing the at least two sub-sample images into a first training set and a second training set; where the first training set includes a first sub-sample image with a visible attribute, and the second training set includes a second sub-sample image with an invisible attribute; performing self-supervised training on a to-be-trained encoder by taking the second training set as a tag of the first training set, to obtain a target encoder.

Character recognition model training method and apparatus, character recognition method and apparatus, device and storage medium

Provided are a joint perception model training method, a joint perception method, a device, and a storage medium. The joint perception model training method includes: acquiring sample images and perception tags of the sample images; acquiring a preset joint perception model, where the joint perception model includes a feature extraction network and a joint perception network; performing feature extraction on the sample images through the feature extraction network to obtain target sample features; performing joint perception through the joint perception network according to the target sample features to obtain perception prediction results; and training the preset joint perception model according to the perception prediction results and the perception tags, where the joint perception includes executing at least two perception tasks.

Joint perception model training method, joint perception method, device, and storage medium

A pre-training method for a Vision and Scene Text Aggregation model includes: acquiring a sample image-text pair; extracting a sample scene text from a sample image; inputting a sample text into a text encoding network to obtain a sample text feature; inputting the sample image and an initial sample aggregation feature into a visual encoding subnetwork and inputting the initial sample aggregation feature and the sample scene text into a scene encoding subnetwork to obtain a global image feature of the sample image and a learned sample aggregation feature; and pre-training the Vision and Scene Text Aggregation model according to the sample text feature, the global image feature of the sample image, and the learned sample aggregation feature.

Pre-training method, image and text retrieval method for a vision and scene text aggregation model, electronic device, and storage medium

System and methods are directed to operations associated with an approximate nearest neighbor search engine. More specifically, a vector semantically representing content to be added to a search index may be received. The search index may include a neighborhood graph having a plurality of nodes, where each node of the plurality of nodes is associated with content in a content repository. A plurality of nodes within the search index determined to be most semantically similar to the received vector semantically representing content to be added to the search index may be identified. The node corresponding to the received vector semantically representing content to be added to the search index to the search index may be added to the search index and a listing of nearest neighbors associated with each of the of the plurality of nodes may be updated to include an identifier associated with the added node.

Approximate nearest neighbor search engine

A computer-vision method includes recognizing a feature representation of a query image depicting an unknown subject. A similarity score is computed between the representation of the query image and feature representations of a plurality of gallery images collectively depicting two or more different subjects with at least two or more gallery images for each subject, and each gallery image having a label identifying which of the subjects is depicted. One or more updated feature representations of the query image are sequentially iterated based on one or more of the computed similarity scores. For each of the one or more updated feature representations, an updated similarity score is computed between the updated feature representation and the feature representations of each of the gallery images. The unknown subject is identified based on a gallery image having a highest updated similarity score.

Subject identification based on iterated feature representation

Techniques for intelligent image search results summarization and browsing scheme are described. Images having visual attributes are evaluated for similarities based in part on their visual attributes. At least one preference score indicating a probability of an image to be selected into a summary is calculated for each image. Images are selected based on the similarity of the selected images to the other images and the preference scores of the selected images. A summary of the plurality of images is generated including the selected one individual image.

Intelligent image search results summarization and browsing

A set of data points is divided into a plurality of subsets of data points. A set of cluster closures is generated based at least in part on the subset of data points. Each cluster closure envelopes a corresponding cluster of a set of clusters and is comprised of data points of the enveloped cluster and data points neighboring the enveloped cluster. A k-Means approximator iteratively assigns data points to a cluster of the set of clusters and updates a set of cluster centroids corresponding to the set of clusters. The k-Means approximator assigns data points based at least in part on the set of cluster closures.

Approximate K-means via cluster closures

A facility for visual search on a mobile device takes advantage of multi-modal and multi-touch input on the mobile device. By extracting lexical entities from a spoken search query and matching the lexical entities to image tags, the facility provides candidate images for each entity. Selected ones of the candidate images are used to construct a composite visual query image on a query canvas. The relative size and position of the selected candidate images in the composite visual query image, which need not be an existing image, contribute to a definition of a context of the composite visual query image being submitted for context-aware visual search.

Interactive multi-modal image search

The techniques described herein determine an initial set of ranked images associated with an image-based search query. Based on visual content similarities between images in the initial set of ranked images, the techniques select confident image samples from the initial set of ranked images. The techniques then use the confident image samples to rerank the initial set of ranked images. Accordingly, a search engine uses the confident image samples to promote images that are likely to be relevant to the search query, while demoting images that are not likely to be relevant to the search query. Therefore, the search engine can provide improved relevance-based search results to an image-based search query.

Reranking using confident image samples

Described is a technology for computing visual and textual summaries for tagged image collections. Heterogeneous affinity propagation is used to together identify both visual and textual exemplars. The heterogeneous affinity propagation finds the exemplars for relational heterogeneous data (e.g., images and words) by considering the relationships (e.g., similarities) within pairs of images, pairs of words, and relationships of words to images (affinity) in an integrated manner.

Computing visual and textual summaries for tagged image collections

Techniques for identifying a salient object with respect to its context are described. A process receives an input image that includes a salient object. The process segments the input image into multiple regions and calculates a saliency value for each of the segmented regions based on scale image levels. The process constructs saliency maps based at least in part on the calculated saliency value, and combines the saliency maps to construct a total saliency map. Next, the process connects a set of line segments computed from the input image and utilizes the total saliency map to compute a closed boundary, which forms a shape prior from the closed boundary, and extracts the salient object from the total saliency map and the shape prior.

Salient object segmentation

Dominant color names may be extracted from an image by analyzing spatial-context of pixels contained in the image. A dominant color region may be defined by taking a double-threshold approach that addresses ambiguous color regions and a degree of confidence that each pixel belongs in the dominant color region. Affiliation maps and binary maps may be used to generate the dominant color region. Images may be converted to a saliency map, from which a region of interest may be assigned a dominant color name. Image search results may be filtered by the dominant color name associated with the image.

Contextual dominant color name extraction

Some implementations include techniques and arrangements to provide document-related representative information with search results. For example, a representative query and/or representative results may be provided for one or more individual documents identified in a set of search results to supplement the search results returned in response to a received search query. The representative queries may be determined by correlating a plurality of previously submitted queries in search log data with a plurality of documents returned in response to the queries. In some implementations, click-through frequency for a particular document with respect to the plurality of queries may be taken into consideration when determining the representative queries for the particular document. In some implementations, the representative queries serve to categorize the search results based on subject matter, and a link may be provided to representative results corresponding to the representative query for accessing documents directed to similar subject matter.

Document-related representative information

The concept-structured image search technique described herein pertains to a technique for enabling a user to indicate their semantic intention and then retrieve and rank images from a database or other image set according to this intention. The concept-structured image search technique described herein includes a new interface for image search. With this interface, a user can freely type several key textual words in arbitrary positions on a blank image, and also describe a region for each keyword that indicates its influence scope, which is called concept structure herein. The concept-structured image search technique will return and rank images that are in accordance with the concept structure indicated by the user. One embodiment of the technique can be used to create a synthesized image without actually using the synthesized image to perform a search of an image set.

Concept-structured image search

Growing community of inventors

Beijing, China

Jingdong Wang