👀 양재 허브 인공지능 오픈소스 경진대회 👀

맑은 눈의 광인이 되어 세상을 바라보자!

Team

Name : 맑은 눈의 광인

Members : KUBIG 15기 염윤석, 장수혁, 최민경 / 16기 박민규

Project Descriptions

Link : AI 양재 허브 인공지능 오픈소스 경진대회 (DACON)

[주제]
이미지 초해상화(Image Sper-Resolution)를 위한 AI 알고리즘 개발

[배경]
오픈 소스 이미지 데이터를 활용하여 인공지능 컴퓨터 비전의 '이미지 초해상화' 분야 연구개발

[설명]
품질이 저하된 저해상도 촬영 이미지(512x512)를 고품질의 고해상도 촬영 이미지(2048x2048)로 생성

[평가 산식]
PSNR(Peak Signal-to-Noise Ratio) = $10log_{10}(R^2/MSE)$

• 생성 혹은 압축된 영상의 화질에 대한 “손실 정보”를 평가
• 손실이 적을수록(=화질이 좋을 수록) 높은 값

👣 Score(Public)

RRDB+ : 23.40812(38th)

🌐 Environment

Colab Pro+
GPU: A100-SXM4-40GB * 1(Main) , Tesla T4*1(Sub)

🔥 Competition Strategies

1. Patches(for data augmentation)

• Train patches : original 1640 images → 26240(1640*16) patches (X4 downsampling, non-overlapping)
• Test patches: original 18 images → 882(18*49) patches (X4 downsampling, overlapping(to remove border artifacts))

2. Data Transform
Non-destructive transformations (not to add or lose the information)

• Flip
• Transpose
• RandomRotate
• ShiftScaleRotate

3. Training Methods

• EarlyStopping

  To prevent overfitting
  If validation loss does not improve after given patience(=2), training is earlystopped

• Fine-tuning with pre-trained model

  pretrained model : RRDB_PSNR_x4.pth(the PSNR-oriented model with high PSNR performance)
  Retraining entire model : Judging that the similarity between DF2K dataset(pretrained model) and our training datset is small

4. Loss Function

• L1 loss + L2 loss (2:1)

  L2 loss : PSNR is based on MSE
  L1 loss: For better convergence https://arxiv.org/pdf/1707.02921v1.pdf

5. Learning Scheduler, Optimizer

• StepLR

  step_size = 3, gamma = 0.5
  Decays the learning rate of each parameter in half once per 3 epochs

• Adam

6. Post Processing

• Geometric Self-Ensemble https://arxiv.org/pdf/1707.02921v1.pdf

  

Main configuration & Hyperparameters

'''

1. Manuel_seed : 42

2. Model :

   num_feat : 64 , Channel number of intermediate features.
   growth_channel: 32 , Channels for each growth(dense connection).
   num_block: 23 , number of RRDB blocks.

3. Dataloader :

   train_batch_size : 4
   test_batch_size: 1
   num_workers: 4

4. Train :

   epochs: 7
   optim_g: {type: Adam, lr: 1e-4, betas: [0.9, 0.99]}

'''

Code Descriptions

1. DACON_AISR_TRIAL

• EDSR, SRGAN, SWINIR

2. DACON_AISR_BEST

• RRDB, RRDB+(Self-ensemble)