preprocessing-decisions

Preprocessing Decisions

When to Use This Skill

• Choosing a blur/smoothing filter for noise reduction
• Selecting an edge detection algorithm
• Identifying noise type in an image
• Setting kernel sizes and filter parameters
• Building a preprocessing pipeline before segmentation or detection

Decision Framework

Filter Selection Decision Tree

What type of noise?
├── Salt & Pepper (random black/white dots)
│   └── ✅ Median Filter (cv2.medianBlur) — BEST IN THE WORLD for this
│
├── Gaussian noise (camera sensor heat, general grain)
│   └── ✅ Gaussian Blur (cv2.GaussianBlur)
│
├── Unknown noise + must preserve edges
│   └── ✅ Bilateral Filter (cv2.bilateralFilter) — kills noise, keeps edges
│
├── General smoothing (no specific noise type)
│   └── ✅ Mean Filter (cv2.blur) — simplest, fastest
│
└── Medical image with bias field / Rician noise
    └── ✅ Non-Local Means (cv2.fastNlMeansDenoising)

Filter Comparison Matrix

Filter	Speed	Edge Preservation	Noise Removal	Best For
Mean (cv2.blur)	⚡⚡⚡	❌ Poor	⭐⭐	General smoothing
Gaussian (cv2.GaussianBlur)	⚡⚡⚡	⭐ Fair	⭐⭐⭐	Gaussian noise, pre-Canny
Median (cv2.medianBlur)	⚡⚡	⭐⭐ Good	⭐⭐⭐⭐⭐ (S&P)	Salt & Pepper noise
Bilateral (cv2.bilateralFilter)	⚡	⭐⭐⭐⭐⭐	⭐⭐⭐⭐	Edge-aware denoising

Rule of thumb: If you don't know the noise type, start with Gaussian. If edges matter, use Bilateral. If you see random black/white dots, use Median — nothing else comes close.

Edge Detection Priority

What do you need to detect?
├── General edges (most use cases)
│   └── ✅ Canny (cv2.Canny) — gold standard, 5-step pipeline
│
├── Directional edges (horizontal OR vertical)
│   └── ✅ Sobel (cv2.Sobel) — first derivative, specify dx/dy
│
├── Fine detail + corners + all boundaries
│   └── ✅ Laplacian (cv2.Laplacian) — second derivative, zero-crossing
│
└── Text/document character edges (OCR preprocessing)
    └── ✅ Prewitt — better than Sobel for text sharpness