issues
search
DeepResearchTeam
/
Deep_Research
To summarize the information of machine learning
3
stars
0
forks
source link
Image-to-Image Translation with Conditional Adversarial Networks
#22
Open
KenichiSasaki
opened
4 years ago
KenichiSasaki
commented
4 years ago
概要
アルゴリズム系論文(
2018
)
GANを利用し,2つのペア画像から画像間の関係を学習し,1枚の画像から補完しペア画像を生成する技術
アルゴリズム
Conditional GANを使用,通常のDCGANと異なりGeneratorに入力したinput とoutput画像を両方学習に利用,イメージ的には正解ペアなのか不正解ペアなのかを学習
GeneratorにはUnet等のEncoder/ Decoderモデルを使用
またcGANモデルにL1正則化を組み込み精度を向上
実装
建物,物,景色の写真や衛星画像とその線画(またはセグメンテーション,地図)の変換を実施
夏冬や昼夜画像の変換等も実施
Unetの方がEncoder/ Decoderモデルより基本構造を維持
Skip結合による特徴量の保持
L1項を追加するとより鮮明な画像生成
所感
GANを一躍有名にした潜在変数を補完することで別の要素を加えた画像を生成した研究の代表例
衛星画像と地図の変換もしていたのは知らなかった
異常検知をGANで行う研究を応用すれば車の検知にも使える?
概要
アルゴリズム
実装
所感