Back to results
US 11,636,580 B2Granted
Human portrait segmentation based image processing system
AssigneeBlack Sesame Technologies Inc.
InventorsFangwen Tu, Bo Li, Jin Xu, Jizhang Shan
Life of this patent
Estimating the term…
Abstract
The present invention discloses system and method for processing an image. The invention processes the image by segmenting a human portrait region of the image. The invention uses ahierarchical hybrid loss module for masking the portrait region generating masked portrait region. The invention also uses data learning the masked portrait region.
Claims14
An image processing system for processing an image, wherein the image processing system comprising:
- a classification module, wherein the classification module classifies a portrait region of the image, further wherein the classification module comprising:
- an encoder-decoder unit, wherein the encoder-decoder unit comprising:
- an encoder, wherein the encoder extracts and classifies a plurality of discriminative features and a plurality of similar features from the portrait region, further wherein the plurality of similar features are classified based on semantic classification at one or more predefined levels;
- and
- a decoder, wherein the decoder refines the one or more similar features to generate a plurality of refined features;
- a skip fusion connection logic, wherein the skip fusion connection logic concatenates the plurality of refined features to generate a refined portrait region of the image;
- a hierarchical hybrid loss module, wherein the hierarchical hybrid loss module comprising:
- a binarycross entropy (BCE) loss unit, wherein the binarycross entropy loss unit regulates a first masking applied on a first portion of the refined portrait region;
- a gradient loss unit, wherein the gradient loss unit applies a second masking the refined portrait region;
- and
- a local binarycross entropy (BCE) loss unit, wherein the local binarycross entropy loss unit regulates a third masking applied on a second portion of the refined portrait region to form a masked portrait region;
- a human centric data augmentation module, wherein the human centric data augmentation module performs a pre-alignment of the masked portrait region to center area and generates an output image.
9 more elements
A method for processing an image, wherein the method comprising:
- extracting a plurality of discriminative features and a plurality of similar features from the image at one or more predefined levels,
- classifying the plurality of similar features based on semantic classification;
- refining the one or more similar features to generate a plurality of refined features;
- concatenating the one or more refined features with identical sizes to generate a refined portrait region:
- applying a first mask on a first portion of the refined portrait region based on a binarycross entropy (BCE) loss;
- applying a second masking on the refined portrait region based on a gradient loss;
- applying a third masking on a second portion of the refined portrait region based on a local binarycross entropy (BCE) loss;
- tuning a hybrid loss based by associating weights to each of the binarycross entropy (BCE) loss, the gradient loss and the local binarycross entropy (BCE) loss to generate a masked portrait region;
- and
- pre-aligning the masked portrait region to center area to generate an output image.
7 more elements
A non-transitory computer program product comprising a computer useable medium having computer program logic for enabling at least one processor in a computer system for processing an image via a software platform, said computer program logic comprising:
- extracting a plurality of discriminative features and a plurality of similar features from the image at one or more predefined levels,
- classifying the plurality of similar features based on semantic classification;
- refining the one or more similar features to generate a plurality of refined features;
- concatenating the one or more refined features with identical sizes to generate a refined portrait region;
- applying a first mask on a first portion of the refined portrait region based on a binarycross entropy (BCE) loss;
- applying a second masking on the refined portrait region based on a gradient loss;
- applying a third masking on a second portion of the refined portrait region based on a local binarycross entropy (BCE) loss;
- tuning a hybrid loss based by associating weights to each of the binarycross entropy (BCE) loss, the gradient loss and the local binarycross entropy (BCE) loss to generate a masked portrait region;
- and
- pre-aligning the masked portrait region to center area to generate an output image.
7 more elements
Dependencies are inferred from the claim text.