AREA EXTRACTION MODEL LEARNING DEVICE, AREA EXTRACTION MODEL LEARNING METHOD, AND PROGRAM

To provide a technique for extracting a person area from images that can reduce the preparation cost of learning data.SOLUTION: An area extraction model learning device includes: a synthetic learning data generation unit for generating synthetic learning data which is a set of a synthetic image and...

Full description

Saved in:
Bibliographic Details
Main Authors MIKAMI DAN, MATSUMOTO AYUMI, KIMATA HIDEAKI
Format Patent
LanguageEnglish
Japanese
Published 12.12.2019
Subjects
Online AccessGet full text

Cover

Loading…
More Information
Summary:To provide a technique for extracting a person area from images that can reduce the preparation cost of learning data.SOLUTION: An area extraction model learning device includes: a synthetic learning data generation unit for generating synthetic learning data which is a set of a synthetic image and a synthetic mask indicating a person area of the synthetic image from a background image serving as a background of the synthetic image and existing learning data which is a set of an image including the person area and a mask indicating the person area; and a learning unit for learning a model parameter by using the synthetic learning data. The synthetic learning data generation unit includes: a synthetic parameter generation unit for generating a synthetic parameter which is a set of a magnification rate, a translation degree, and a rotational degree by using the mask of the learning data; and a synthetic image/synthetic mask generation unit for extracting a synthetic person area from the image of the learning data by using the mask of the learning data, generating a synthetic image from the background image and the synthetic person area by using the synthetic parameter, generating a synthetic mask from a mask generation image having the same size as the synthetic image and the synthetic person area by using the synthetic parameter, and generating synthetic learning data.SELECTED DRAWING: Figure 1 【課題】学習データの準備コストを抑えることができる、画像から人物領域を抽出する技術を提供する。【解決手段】人物領域を含む画像と人物領域を示すマスクの組である既存の学習データと合成画像の背景となる背景画像から合成画像と合成画像の人物領域を示す合成マスクの組である合成学習データを生成する合成学習データ生成部と、合成学習データを用いてモデルパラメータを学習する学習部とを含み、合成学習データ生成部は、学習データのマスクを用いて、拡大率、並進度、回転度の組である合成パラメータを生成する合成パラメータ生成部と、学習データのマスクを用いて学習データの画像から合成人物領域を抽出し、合成パラメータを用いて背景画像と合成人物領域から合成画像を生成し、合成パラメータを用いて合成画像と同一サイズのマスク生成用画像と合成人物領域から合成マスクを生成し、合成学習データを生成する合成画像・合成マスク生成部とを含む。【選択図】図1
Bibliography:Application Number: JP20180108662