Konica Minolta Laboratory U.s.a., Inc.

Samsung Electronics Co., Ltd.

Duanduan Yang

Saman Sarraf

Xiang Sun

A method for a neural network includes receiving an input from a vector of inputs, determining a table index based on the input, and retrieving a hash table from a plurality of hash tables, wherein the hash table corresponds to the table index. The method also includes determining an entry index of the hash table based on an index matrix, wherein the index matrix includes one or more index values, and each of the one or more index values corresponds to a vector in the hash table and determining an entry value in the hash table corresponding to the entry index. The method also includes determining a value index, wherein the vector in the hash table includes one or more entry values, and wherein the value index corresponds to one of the one or more entry values in the vector and determining a layer response.

System and method for hashed compressed weighting matrix in neural networks

A method includes identifying, using at least one processor, input words associated with a user query. The method also includes, for each of one or more of the input words that are contained in a high-frequency word set, retrieving pre-computed element-wise products associated with the input word from a cache. The method further includes performing, using the at least one processor, a convolution operation using the pre-computed element-wise products. In addition, the method includes generating, using the at least one processor, a response to the user query based on results of the convolution operation. The method may also include, for each of one or more of the input words that are not contained in the high-frequency word set, calculating additional element-wise products associated with the input word, and the convolution operation may be performed using the pre-computed element-wise products and the additional element-wise products.

System and method for cached convolution calculation

A method for segmenting an image containing handwritten text into line segments and word segments. The image is horizontally down sampled at a first ratio. Connected regions in the down-sampled image are detected; horizontal neighboring ones are merged to form lines, to segment the original image into line images. Each line image is horizontally down sampled at a second ratio which is smaller than the first ratio. Connected regions in the down-sampled line image are detected to obtain potential word segmentation positions. A path is a way of dividing the line at some or all of the potential word segmentation positions into multiple path segments; for each of all possible paths, word recognition is applied to each path segment to calculate a word recognition score, and an average word recognition score for the path is calculated; the path with the highest score gives the final word segmentation.

Method for line and word segmentation for handwritten text images

Disclosed herein is a method of optimizing data normalization by selecting the best height normalization setting from training RNN (Recurrent Neural Network) with one or more datasets comprising multiple sample images of handwriting data, which comprises estimating a few top place ratios for normalization by minimizing a cost function for any given sample image in the training dataset, and further, determining the best ratio from the top place ratios by validating the recognition results of sample images with each top place ratio.

Data normalization for handwriting recognition

Disclosed herein is a method that improves the performance of handwriting recognition by calculating path scores so as to identify the path with the highest score as the basis for interpreting handwritten characters. Specifically, the method comprises the following steps: detecting connected regions in an input image comprising handwritten characters; determining a plurality of segmentation positions of the input image; obtaining a plurality of recognition results for each segment of each path in the input image, wherein each recognition result represents a character candidate for the segment and each path comprises one or more segments; obtaining a plurality of scores corresponding to the recognition results; calculating scores for each path in the input image based on segment lengths and the scores corresponding to the recognition results; and using the path with the highest score to interpret the handwritten characters in the input image.

Growing community of inventors

San Jose, CA, United States of America

Duanduan Yang