Image & Video Compression: Techniques, Trade-offs & Real-world Use

TAKEAWAYS:

• Lossy vs. lossless: Lossy compression (JPEG, MPEG) sacrifices minor data for huge size gains (e.g., 90% reduction), while lossless (RLE, Huffman) preserves 100% data but offers modest savings (e.g., 50%).
• Perceptual models: Humans ignore high-frequency details (e.g., fine textures in grass) and redundant colors—JPEG exploits this via DCT and quantization.
• Video compression: MPEG uses inter-frame prediction (I/P/B frames) to store only changes between frames, cutting bandwidth by 90%+ for 4K streams.
• Trade-offs: Higher compression → worse quality → more artifacts (blockiness, blur). Real-world examples: eSewa’s thumbnails use JPEG (lossy) for fast loading; Ncell’s video calls use MPEG (lossy) for smooth streaming.
• Hardware matters: GPUs/TPUs accelerate compression (e.g., YouTube’s VP9 codec runs on Google’s TPUs).
• Standards matter: JPEG for photos, MPEG-4 for videos, WebP for web—each balances speed, quality, and compatibility.

1. Why Compress Multimedia?

Multimedia files (images, videos) are huge:

• A 1-minute 4K video at 60 fps = ~1.5 GB uncompressed.
• A single 8MP photo = ~20 MB raw (16-bit color depth).
• Problem: Networks (e.g., NTC’s 10 Mbps), storage (e.g., Daraz’s servers), and devices (e.g., Pathao’s phones) can’t handle this.

Solution: Compression reduces size while keeping perceived quality intact. Key metrics:

• Compression ratio = Uncompressed size / Compressed size.
• PSNR (Peak Signal-to-Noise Ratio): Measures quality loss (higher = better).
• Bitrate: Bits per second (e.g., 10 Mbps for 4K Netflix).

2. Lossless Compression: No Data Lost

Definition: Reversibly reduces redundancy without losing any data. Used for medical images, CAD files, or archival photos.

A. Run-Length Encoding (RLE)

How it works:

1. Replace consecutive identical values with a count + value pair.
2. Example: "aaaaabbbbcccdde" → (5,a)(4,b)(3,c)(2,d)(1,e).

Worked Example: Traffic Light Sequence Assume a 10-second traffic light cycle recorded as RGB values every second: Red(255,0,0), Red(255,0,0), Green(0,255,0), ..., RLE compresses repeated Red frames into (2,Red) → 80% smaller.

Limitations:

• Only works for simple patterns (e.g., fax machines, BMP files).
• Not effective for complex images (e.g., photos).

graph LR
    A["Uncompressed: R R R G G B B"] -->|"RLE"| B["(3,R)(2,G)(2,B)"]
    B -->|"Decompress"| A

B. Huffman Coding

How it works:

1. Assign shorter codes to frequent symbols, longer to rare ones.
2. Example: For "aaaaabbbbcccdde":
   • a (5x) → 0, b (4x) → 10, c (3x) → 110, d (2x) → 1110, e (1x) → 1111.

Advantages:

• Works for any data (text, audio, images).
• Optimal for lossless compression (theoretical limit: entropy).

Disadvantage:

• Slow for real-time apps (e.g., video calls).

Real-world use:

• PNG images (lossless web format) use Huffman + LZ77.
• ZIP files combine Huffman with LZ77.

3. Lossy Compression: Sacrifice Quality for Size

Definition: Permanently discards perceptually irrelevant data. Used for photos (JPEG), videos (MPEG), music (MP3).

A. JPEG for Images

How it works (8-step process):

1. Color space conversion: RGB → YCbCr (luma + chroma).
   • Humans see luma (Y) better than color (CbCr), so chroma is subsampled (e.g., 4:2:0).
2. Downsampling: Reduce resolution (e.g., 8MP → 2MP).
3. DCT (Discrete Cosine Transform): Split image into 8×8 blocks and convert to frequency domain.
   • High frequencies = textures/details (e.g., grass blades). . Quantization: Discard high-frequency coefficients (e.g., keep only top 10%).
4. Zig-zag scan: Reorder coefficients for Huffman coding.
5. Entropy coding: Huffman/RLE on quantized data.
6. Huffman coding: Assign short codes to frequent coefficients.

Artifacts:

• Blocking: Visible 8×8 grid at high compression.
• Blurring: Loss of high-frequency details.

Worked Example: eSewa Thumbnail

• Original: 5MB PNG (lossless).
• JPEG at 80% quality: 500 KB (90% smaller).
• Trade-off: Slight blur but loads instantly on mobile.

flowchart TD
    A["RGB Image"] --> B["YCbCr\n(Subsample CbCr)"]
    B --> C["8×8 DCT Blocks"]
    C --> D["Quantize\n(Discard high freq)"]
    D --> E["Zig-zag Scan"]
    E --> F["Huffman\nEncode"]
    F --> G["JPEG File"]

B. MPEG for Video

How it works:

1. Temporal compression: Store only changes between frames (unlike JPEG, which treats each frame independently).
2. Frame types:
   • I-frame: Full frame (like JPEG), used as reference.
   • P-frame: Stores differences from previous I/P frame (predictive).
   • B-frame: Stores differences from past/future frames (bidirectional).
3. Motion compensation: Track object movement (e.g., a car in traffic) and store only the delta.
4. DCT + Quantization: Same as JPEG for each frame.

Example: Daraz Streaming

• Uncompressed: 1080p60 = 1.5 Gbps.
• MPEG-4 (H.264): 5 Mbps (300× smaller).
• Why? Uses I/P/B frames + motion compensation.

sequenceDiagram
    participant I as I-Frame
    participant P as P-Frame
    participant B as B-Frame
    I->>P: "Store changes from me"
    P->>B: "Store changes from I/P"
    B->>P: "Bidirectional\n(uses future frame)"

4. Perceptual Models: Why Lossy Works

Humans can’t perceive all details:

Example: Kathmandu Traffic Video

• Uncompressed: 4K @ 30 fps = 10 Gbps.
• MPEG-4 (H.265): 2 Mbps (5000× smaller).
• Why? Motion compensation ignores tiny car movements; DCT blurs distant buildings.

5. Real-World Applications in Nepal

Worked Example: Pathao Driver’s Video Feed

• Problem: Driver’s phone must stream 360p @ 15 fps to the app in real-time.
• Solution: H.264 with GOP (Group of Pictures) every 1 second (I-frame + 2 P-frames).
• Result: 500 Kbps instead of 10 Mbps → works on 2G networks.

6. Trade-offs: Lossy vs. Lossless

Example: Bank Cheque Scanning

• Lossy (JPEG): Blurs text → invalid for processing.
• Lossless (PNG): Preserves all details → used by Nabil Bank’s cheque readers.

7. Hardware Acceleration

Modern devices use GPUs/TPUs to speed up compression:

• NVIDIA NVENC: Encodes 4K60 H.264 in real-time (used in YouTube Live).
• Google TPUs: Run VP9 for YouTube videos (faster than CPUs).
• Apple A-series chips: Hardware-accelerate HEVC (used in iPhone videos).

Exam Tip

1. Compare JPEG vs. MPEG:

   • JPEG: Single-frame, spatial compression (DCT).
   • MPEG: Multi-frame, temporal + spatial (I/P/B frames).
   • Exam trick: Always mention perceptual models (e.g., "MPEG exploits temporal redundancy between frames").

2. Trade-offs are key:

   • Why lossy for 4K video? → "Bandwidth constraints (e.g., NTC’s 10 Mbps) require 90%+ compression; lossless would need 100 Mbps."
   • Why lossless for medical images? → "No data loss allowed (e.g., tumor detection)."

3. Worked examples:

   • For RLE, use a traffic light sequence or fax text.
   • For JPEG, trace an 8×8 DCT block with quantization.
   • For MPEG, draw an I/P/B frame timeline.

4. Real-world ties:

   • eSewa/Khalti: "JPEG thumbnails for fast loading."
   • Ncell/Pathao: "MPEG-4 for low-bandwidth streaming."
   • Daraz: "HEVC for high-quality product videos."

5. Avoid common mistakes:

   • ❌ "JPEG is lossless." → It’s lossy!
   • ❌ "MPEG stores every frame." → It uses I/P/B frames.
   • ❌ Ignoring perceptual models → Always mention "human vision limits."

Quick Revision Table