만나보세요
Prezi AI.
새로운 프레젠테이션 도우미가 기다리고 있어요.
뎌욱 빠르게 컨텐츠를 다듬고, 보강하고, 편집하고, 원하는 이미지를 찾고, 시각자료를 편집하세요.
트렌드 검색
콘텐츠 로딩 중...
Machine_Learning_Pre
5
스크립트
Content Representation
Representation of Content and Style
Style Representation
- Pre-trained deep CNN outputs high-level feature of the image.
- Higher level features approximated the objects that human can recognize.
Training Flowchart
Images Style Transfer Using CNN
Group Member:
Introduction
Transfer Photo
Content
& Style
Training
Data Source
Output
Demo
- Style transfer is the technique of recomposing original images in the style of target images.
- CNN can learn the style,content and generate new image.
Style Photo:
22 oil painting images
Content Photo:
80 thousands Photos of figures and still-lifes
Transfer Video
Original Video：
Content Image + Style Image =
Style Transfer Image
Content Video = Streaming Images + Stlye Image=
Style Transfer Video
Transfered Video：
Jiaming Nie
Ruojun Li
Yu Li
Yang Tao
Guangda Li
Loss Definition
Training Loss Plot
Minimize: Total Loss = α×Content Loss + β×Style Loss
Conclusion
- VGG 16 was used to extract the feature and reconstruct images.
- Minimize the total loss
- Use Adam optimizer, learning rate = 0.001
- Training the network takes very long time, a light CNN may be used for time reduction.
- The tune of parameter α and β to modify the output
Thank you
Questions & Answer
템플릿을 선택하세요
프레지로 더욱 인상깊고 역동적인 프레젠테이션을 만들어 보세요