2nd/10_Wiki/Topics/AI_and_ML/Pattern-Recognition.md

---
id: wiki-2026-0508-pattern-recognition
title: Pattern Recognition
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Pattern Classification, Statistical Pattern Recognition]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [pattern-recognition, classification, ml, signal-processing]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: scikit-learn, pytorch
---

# Pattern Recognition

## 매 한 줄
> **"매 raw signal → category label"**. 매 1960s statistical pattern recognition (Bayes, kNN, LDA) → 1980s neural pattern recognition → 매 2020s deep learning 의 absorbed sub-field. 매 modern frame: 매 supervised classification + representation learning. Bishop's PRML (2006) 의 canonical reference.

## 매 핵심

### 매 history & framing
- 매 1960s: Bayesian decision theory + linear classifiers.
- 매 1970-80s: HMM (speech), template matching (OCR).
- 매 1990s: SVM, kernel methods 의 dominance.
- 매 2010s: DL absorbed it — "pattern recognition" 의 vintage term.
- 매 modern: 매 ML/DL classification + representation learning.

### 매 classical approaches
- **Statistical**: Bayesian, MAP, ML, GMM, LDA, QDA.
- **Neural**: perceptron → MLP → CNN → transformer.
- **Structural / syntactic**: grammar-based, less common today.
- **Template matching**: cross-correlation, used in OCR, fingerprint.
- **Kernel methods**: SVM with RBF/poly kernels.

### 매 pipeline (classic)
1. Sensor / data acquisition.
2. Preprocessing (denoise, normalize).
3. Feature extraction (HOG, SIFT, MFCC) — 매 hand-crafted.
4. Classifier (SVM, RF, NN).
5. Post-processing (smoothing, thresholding).

### 매 modern deep pipeline
- 매 raw input → end-to-end DNN → label.
- Feature extraction = learned (no HOG/SIFT).
- Backbone (ResNet, ViT, CLIP) → head (linear / MLP).

### 매 응용
1. Computer vision (face, OCR, object detection).
2. Speech recognition (Whisper, ASR).
3. Biometrics (fingerprint, iris).
4. Medical imaging (radiology AI).
5. Anomaly detection (fraud, network intrusion).

## 💻 패턴

### Classic: Bayesian classifier
```python
from sklearn.naive_bayes import GaussianNB
from sklearn.datasets import load_iris
X, y = load_iris(return_X_y=True)
clf = GaussianNB().fit(X, y)
print(clf.score(X, y))
```

### Classic: SVM with RBF
```python
from sklearn.svm import SVC
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import make_pipeline

clf = make_pipeline(StandardScaler(), SVC(kernel="rbf", C=1.0, gamma="scale"))
clf.fit(X_train, y_train)
```

### Modern: CNN classification
```python
import torch.nn as nn
import torchvision.models as tvm

model = tvm.resnet50(weights=tvm.ResNet50_Weights.IMAGENET1K_V2)
model.fc = nn.Linear(2048, num_classes)  # transfer learn
```

### Modern: CLIP zero-shot
```python
import clip, torch
model, preprocess = clip.load("ViT-B/32")
classes = ["cat", "dog", "bird"]
text = clip.tokenize([f"a photo of a {c}" for c in classes])
with torch.no_grad():
    img_feat = model.encode_image(preprocess(image).unsqueeze(0))
    txt_feat = model.encode_text(text)
    logits = (img_feat @ txt_feat.T).softmax(-1)
print(classes[logits.argmax()])
```

### HOG + SVM (classic CV pipeline)
```python
from skimage.feature import hog
from sklearn.svm import LinearSVC

features = [hog(img, pixels_per_cell=(8,8)) for img in images]
clf = LinearSVC().fit(features, labels)
```

### Anomaly detection
```python
from sklearn.ensemble import IsolationForest
detector = IsolationForest(contamination=0.01).fit(X_train)
anomalies = detector.predict(X_test) == -1
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Tabular small data | RF / GBDT / SVM |
| Image | Pretrained CNN/ViT (transfer) |
| Speech / audio | Whisper / wav2vec finetune |
| Few-shot / zero-shot | CLIP / SigLIP / VLM |
| Anomaly (no labels) | IsolationForest / autoencoder |
| Real-time embedded | Quantized CNN (MobileNet) |

**기본값**: 매 image/speech 의 pretrained foundation model 의 fine-tune; tabular 의 GBDT.

## 🔗 Graph
- 부모: [[Machine-Learning]] · [[Statistical-Inference]]
- 변형: [[Recognition]]
- 응용: [[Computer Vision|Computer-Vision]] · [[OCR]] · [[Biometrics]]
- Adjacent: [[Feature Engineering|Feature-Engineering]] · [[CLIP]]

## 🤖 LLM 활용
**언제**: 매 classification problem framing, choosing classical vs DL approach, transfer learning decision.
**언제 X**: 매 generative tasks (use diffusion / LLM instead).

## ❌ 안티패턴
- **Hand-crafted features in 2026**: 매 HOG/SIFT 의 99% case 에서 pretrained CNN feature 가 우수.
- **No baseline**: 매 jumping to DL without trying logistic / RF baseline.
- **Class imbalance ignored**: 매 99% accuracy on 99/1 split = trivial. 매 use F1, ROC-AUC, balanced metrics.
- **Test contamination**: 매 train/test split 의 leakage (especially time series).
- **Calibration ignored**: 매 raw softmax ≠ probability. 매 use Platt scaling / temperature.

## 🧪 검증 / 중복
- Verified (Bishop "PRML" 2006, Duda & Hart "Pattern Classification" 2001).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — classical-to-modern framing, pipeline patterns |