---
id: wiki-2026-0508-image-classification-mastery
title: Image Classification
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [image classification, ResNet, ViT, EfficientNet, ImageNet, CLIP]
duplicate_of: none
source_trust_level: A
confidence_score: 0.96
verification_status: applied
tags: [computer-vision, classification, resnet, vit, efficientnet, clip, imagenet]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Python
  framework: PyTorch / timm / Transformers
---

# Image Classification

## 매 한 줄
> **"매 image → class label"**. 매 ImageNet benchmark. 매 evolution: AlexNet 2012 → VGG → ResNet 2015 → EfficientNet → ViT 2020 → CLIP / DINOv2. 매 modern: 매 foundation model 의 zero-shot.

## 매 핵심

### 매 architecture evolution
- **AlexNet** (2012): 매 deep learning revival.
- **VGG** (2014): 매 deeper.
- **ResNet** (2015): 매 skip connection.
- **EfficientNet** (2019): 매 compound scaling.
- **ViT** (2020): 매 transformer.
- **ConvNeXt** (2022): 매 modern CNN.
- **DINOv2** (2023): 매 self-supervised.
- **CLIP** (2021): 매 zero-shot.

### 매 응용
1. **Medical** (X-ray, pathology).
2. **Industrial** (defect detection).
3. **Retail** (visual search).
4. **Wildlife** (camera trap).
5. **Content moderation**.

## 💻 패턴

### timm (modern model zoo)
```python
import timm
model = timm.create_model('vit_base_patch16_224', pretrained=True)
data_config = timm.data.resolve_data_config({}, model=model)
transforms = timm.data.create_transform(**data_config)
```

### Fine-tune (PyTorch)
```python
import torch
from torchvision.models import resnet50, ResNet50_Weights
model = resnet50(weights=ResNet50_Weights.IMAGENET1K_V2)
model.fc = torch.nn.Linear(2048, n_classes)

# 매 freeze backbone (transfer learning baseline)
for p in model.parameters(): p.requires_grad = False
for p in model.fc.parameters(): p.requires_grad = True

optim = torch.optim.AdamW(model.fc.parameters(), lr=1e-3)
```

### CLIP zero-shot
```python
from transformers import CLIPProcessor, CLIPModel
model = CLIPModel.from_pretrained('openai/clip-vit-large-patch14')
processor = CLIPProcessor.from_pretrained('openai/clip-vit-large-patch14')

texts = ['a photo of a dog', 'a photo of a cat', 'a photo of a bird']
inputs = processor(text=texts, images=image, return_tensors='pt')
out = model(**inputs)
probs = out.logits_per_image.softmax(dim=-1)
```

### DINOv2 (self-supervised features)
```python
import torch
dinov2 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitl14')
features = dinov2(image)  # 매 frozen embedding for downstream
```

### Mixup
```python
def mixup(x, y, alpha=0.4):
    lam = np.random.beta(alpha, alpha)
    idx = torch.randperm(x.size(0))
    return lam * x + (1 - lam) * x[idx], y, y[idx], lam
```

### CutMix
```python
def cutmix(x, y, alpha=1.0):
    lam = np.random.beta(alpha, alpha)
    H, W = x.shape[-2:]
    cut_w = int(W * (1 - lam) ** 0.5)
    cut_h = int(H * (1 - lam) ** 0.5)
    cx, cy = np.random.randint(W), np.random.randint(H)
    x1, y1 = max(0, cx - cut_w//2), max(0, cy - cut_h//2)
    x2, y2 = min(W, cx + cut_w//2), min(H, cy + cut_h//2)
    idx = torch.randperm(x.size(0))
    x[:, :, y1:y2, x1:x2] = x[idx, :, y1:y2, x1:x2]
    return x, y, y[idx], 1 - (x2-x1)*(y2-y1)/(W*H)
```

### Augmentation (albumentations)
```python
import albumentations as A
augment = A.Compose([
    A.RandomResizedCrop(224, 224),
    A.HorizontalFlip(),
    A.ColorJitter(0.2, 0.2, 0.2, 0.1),
    A.RandomErasing(p=0.25),
    A.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
```

### Test-Time Augmentation
```python
def tta_predict(model, image, n_aug=5):
    predictions = []
    for _ in range(n_aug):
        aug_img = random_augment(image)
        predictions.append(model(aug_img).softmax(-1))
    return torch.stack(predictions).mean(0)
```

### Top-K accuracy
```python
def topk_accuracy(logits, labels, k=5):
    topk = logits.topk(k, dim=-1).indices
    correct = (topk == labels.unsqueeze(-1)).any(dim=-1).float().mean()
    return correct
```

### Model card (best practice)
```yaml
model: my-classifier-v2
backbone: vit_base_patch16_224
training_data: ImageNet-1k + custom 100k
classes: 1000
augmentation: RandAugment + Mixup + CutMix
ttest_top1: 84.5
test_top5: 97.2
calibration_ece: 0.034
inference_ms_a100: 8
```

### Modern recipe (DeiT, ViT)
```python
def modern_train_recipe():
    return {
        'optimizer': 'AdamW',
        'lr': 1e-3, 'wd': 0.05,
        'scheduler': 'cosine + warmup 5 epochs',
        'epochs': 300,
        'augmentation': 'RandAugment + Mixup 0.8 + CutMix 1.0',
        'label_smoothing': 0.1,
        'stochastic_depth': 0.1,
        'ema': True,
    }
```

## 매 결정 기준
| 상황 | Model |
|---|---|
| Need pretrained | timm |
| Best ImageNet | DeiT III / ViT-L |
| Mobile | MobileNetV3 / EfficientNet-Lite |
| Zero-shot | CLIP |
| Self-supervised | DINOv2 |
| Tiny | ResNet18 / EfficientNet-B0 |

**기본값**: 매 timm + 매 ViT-B/L pretrained + 매 modern recipe (RandAug + Mixup + CutMix + label smooth) + 매 TTA 의 critical eval.

## 🔗 Graph
- 부모: [[Computer Vision|Computer-Vision]]
- 변형: [[ResNet]] · [[ViT]] · [[EfficientNet]]
- 응용: [[CLIP]] · [[Image-Segmentation]]
- Adjacent: [[Foundation-Models]]

## 🤖 LLM 활용
**언제**: 매 image task. 매 visual search. 매 medical.
**언제 X**: 매 segmentation / detection (다른 task).

## ❌ 안티패턴
- **Train from scratch**: 매 timm pretrained 의 use.
- **No augment**: 매 overfit.
- **Top-1 only**: 매 also top-5 / calibration.
- **No TTA at eval**: 매 lose 1-2%.

## 🧪 검증 / 중복
- Verified (timm, He ResNet 2015, Dosovitskiy ViT 2020, Oquab DINOv2 2023).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — 매 evolution + timm / CLIP / DINOv2 / Mixup / TTA code |