Image File Formats

🖼️ Why Image Formats Matter in Computer Vision

Image file formats define:

• How pixels are stored
• Whether data is compressed
• Whether information is lost
• How images are read and written

To process images correctly, you must understand what happens before pixels become a matrix.

---

📦 Common Image Formats Overview

Format	Compression	Lossy	Alpha	Typical Use
BMP	None	❌	❌	Raw storage, debugging
PNG	Lossless	❌	✅	Masks, UI, screenshots
JPEG	Lossy	✅	❌	Photos, datasets

---

🟦 BMP (Bitmap Image)

Concept

BMP stores raw pixel values with almost no processing.

• Very simple structure
• Large file size
• Easy to parse manually

---

BMP File Structure

[ Bitmap File Header ]
[ DIB Header ]
[ Color Table ] (optional)
[ Pixel Array ]

Bitmap File Header (14 bytes)

• Signature: ‘BM’
• File size
• Pixel data offset

DIB Header (40 bytes typical)

• Image width / height
• Bits per pixel (8 / 24 / 32)
• Compression (usually none)

---

Pixel Storage

• Stored bottom-up by default
• Each row padded to 4-byte alignment
• Common format: BGR (not RGB)

---

When to Use BMP

• Debugging
• Teaching
• Raw inspection

Not suitable for large-scale CV pipelines.

---

🟩 PNG (Portable Network Graphics)

Concept

PNG uses lossless compression.

• Pixel-perfect reconstruction
• Supports alpha channel
• More complex structure

---

PNG File Structure

[ Signature ]
[ IHDR ]
[ PLTE ] (optional)
[ IDAT ] (compressed data)
[ IEND ]

Each block is called a chunk.

---

Key Chunks

IHDR

• Width, Height
• Bit depth (8 / 16)
• Color type (Gray, RGB, RGBA)

IDAT

• zlib-compressed pixel data

---

PNG Filtering

Before compression, each row is filtered:

• Sub
• Up
• Average
• Paeth

Purpose:

• Improve compression efficiency

---

When to Use PNG

• Binary masks
• Labels
• Screenshots
• Lossless datasets

---

🟥 JPEG (JPG)

Concept

JPEG uses lossy compression based on human perception.

• Much smaller files
• Information is lost
• No alpha channel

---

JPEG Compression Pipeline

RGB
 ↓
YCbCr conversion
 ↓
Block splitting (8×8)
 ↓
DCT
 ↓
Quantization (lossy)
 ↓
Entropy coding

---

Key Idea: Quantization

High-frequency components are discarded.

Result:

• Compression
• Blocking artifacts

---

When to Use JPEG

• Natural images
• Large datasets
• When small size matters

Avoid JPEG for:

• Masks
• Depth images
• Scientific measurements

---

🧠 Reading Images (Conceptual)

Reading means:

1. Parse header
2. Decode compression (if any)
3. Reconstruct pixel array
4. Convert to matrix

Result: \(I \in \mathbb{R}^{H \times W \times C}\)

---

✍️ Writing Images (Conceptual)

Writing means:

1. Prepare pixel array
2. Apply format-specific encoding
3. Write headers + data

Key decision:

• Lossy vs lossless
• Bit depth
• Channel order

---

⚠️ Practical CV Pitfalls

• JPEG introduces non-linear noise
• PNG preserves exact values
• BMP row padding causes bugs
• Channel order differs (RGB vs BGR)

---

🧩 Summary: How to Choose

Need	Best Choice
Exact pixel values	PNG / BMP
Small file size	JPEG
Binary mask	PNG
Debugging	BMP
Training images	JPEG / PNG

---

🎯 Takeaway

Image formats are not just containers.

They define:

• Pixel integrity
• Numerical accuracy
• Pipeline correctness

Understanding BMP / PNG / JPEG is essential to build reliable computer vision systems 🚀