Back to results
US 12,591,616 B2Granted
Method, system and non-transitory computer-readable recording medium for searching similar products using a multi task learning model
AssigneeMUSINSA Co., Ltd.
InventorsJae Young Jun, Jun Cheol Park, Yun Hun Jang, Hyung Won Choi, Essalim Chahrazad
Life of this patent
Estimating the term…
Abstract
A method of searching for similar products using a multi-task learning (MTL) model is provided. The method includes converting, by using a multi-task learning model utilizing a unified backbone network, each of a plurality of original images including a fashion item into a single vector including at least two multi-task attributes, and generating a visual search database by storing the original image and the single vector of the original image together.
Claims7
A computer-implemented method for searching for similar products using a multi-task learning (MTL) model, comprising the steps of:
- converting, by using a multi-task learning model utilizing a unified backbone network, each of a plurality of original images including a fashion item into a respective single vector including at least two multi-task attributes;
- generating a visual search database by storing, for each original image, the original image in association with the respective single vector;
- and
- performing a similarity search using the visual search database by comparing a query image single vector corresponding to a query image including a fashion item with the respective single vectors stored in the visual search database to obtain at least one similar product,
- wherein each respective single vector is a vector in which information of the fashion item is positioned in a style space, and represents each attribute of the fashion item,
- wherein the step of converting to a single vector including the at least two multi-task attributes comprises the steps of:
- detecting, for each original image, a location of a representative item in the original image;
- generating, for each original image, intermediate embedding feature vectors by embedding vectorizing the representative item in the original image in accordance with the at least two multi-task attributes;
- and
- generating one single vector positioned in the style space by fusing the intermediate embedding feature vectors, and
- wherein the generating intermediate embedding feature vectors comprises the step of:
- generating, for the at least two multi-task attributes, intermediate embedding feature vectors having different sizes determined based on a task complexity associated with the respective multi-task attributes.
9 more elements
A system for searching for similar products using a multi-task learning (MTL) model, comprising:
- one or more processors;
- and
- a memory storing instructions that, when executed by the one or more processors, cause the one or more processors to:
- convert, by using a multi-task learning model utilizing a unified backbone network, each of a plurality of original images including a fashion item into a respective single vector including at least two multi-task attributes;
- generate a visual search database by storing, for each original image, the original image in association with the respective single vector;
- and
- perform a similarity search using the visual search database by comparing a query image single vector corresponding to a query image including a fashion item with the respective single vectors stored in the visual search database to obtain at least one similar product,
- wherein each respective single vector is a vector in which information of the fashion item is positioned in a style space, and represents each attribute of the fashion item,
- wherein causing the one or more processors to convert each of the plurality of original images into the respective single vector comprises causing the one or more processors to:
- detect, for each original image, a location of a representative item in the original image;
- generate, for each original image, intermediate embedding feature vectors by embedding-vectorizing the representative item in the original image in accordance with the at least two multi-task attributes;
- and
- generate the respective single vector positioned in the style space by fusing the intermediate embedding feature vectors, and
- wherein generating the intermediate embedding feature vectors comprises generating, for the at least two multi-task attributes, intermediate embedding feature vectors having different sizes determined based on a task complexity associated with the respective multi-task attributes.
11 more elements
Dependencies are inferred from the claim text.