Image processing method and apparatus with neural network adjustment
AssigneeSamsung Electronics Co., Ltd.
InventorsCheolhun Jang, Dokwan Oh, Dae Hyun Ji
Life of this patent
Estimating the term…
Abstract
A method and apparatus for adjusting a neural network that classifies a scene of an input image into at least one class is provided. The method generates a feature image having a size that is less than a size of an input image by applying a convolutional network to the input image, determines at least one class corresponding to the feature image, generates a class image having a size corresponding to the size of the input image by applying a deconvolutional network to the feature image, calculates a loss of the class image based on a verification class image preset with respect to the input image, and adjusts the neural network based on the loss.
Claims17
A processor-implemented neural network adjustment method comprising:
- generating a feature image by applying a input image to a convolutional network;
- determining at least one class corresponding to the feature image;
- generating a class image by applying the feature image to a deconvolutional network;
- calculating a loss of the class image based on a verification class image preset with respect to the input image;
- and
- adjusting a neural network based on the loss,
- wherein the adjusting of the neural network comprises:
- calculating a main loss between the verification class image and the class image;
- calculating an auxiliary loss based on a reduced verification class image of verification class image and auxiliary class image of the feature image, wherein the reduced verification class image is generated based on a mode class of a target block of a plurality of blocks of verification class image;
- and
- adjusting the neural network based on the calculated main loss and the calculated auxiliary loss.
A neural network apparatus comprising:
- a processor;
- and
- a memory storing instructions that when implemented by the processor, the processor is configured to:
- generate a feature image by applying a convolutional network to a input image;
- determine at least one class corresponding to the feature image;
- generate a class image by applying a deconvolutional network to the feature image;
- calculate a loss of the class image based on a verification class image preset with respect to the input image;
- and
- adjust a neural network based on the loss,
- wherein the processor is configured to:
- calculate a main loss between the verification class image and the class image;
- calculate an auxiliary loss based on a reduced verification class image of verification class image and auxiliary class image of the feature image, wherein the reduced verification class image is generated based on a mode class of a target block of a plurality of blocks of verification class image;
- and
- adjust the neural network based on the calculated main loss and the calculated auxiliary loss.
Dependencies are inferred from the claim text.