Digital Images: Pixels, Systems & Real-World Applications

TAKEAWAYS:

• Pixels are the atomic units of digital images, defined by their spatial coordinates $(x,y)$, intensity values (grayscale or RGB), and relationships (adjacency, neighborhood) that enable operations like filtering and segmentation.
• Image processing systems follow a pipeline: Acquisition → Preprocessing → Analysis → Compression → Display, with each block (e.g., noise reduction, Fourier transforms) serving a distinct purpose.
• Enhancement vs. Restoration: Enhancement improves visual quality (e.g., histogram equalization), while restoration corrects degradations (e.g., Wiener filtering for blur).
• Color quantization trades off fidelity for storage (e.g., 512 colors = 9 bits/pixel), and sampling determines spatial resolution (e.g., 1024×1024 pixels = 1.048M pixels).
• Real-world ties: WhatsApp uses JPEG compression (DCT + quantization) to reduce file sizes, while Ncell’s satellite imagery relies on Fourier transforms for noise filtering.
• Exam focus: Draw block diagrams, calculate image sizes, and differentiate between operations (e.g., edge detection vs. segmentation) with concrete examples.

1. Digital Images: The Pixel Foundation

Digital images are 2D arrays of pixels, where each pixel stores:

• Spatial coordinates $(x,y)$: Defines its position in the image grid.
• Intensity value: For grayscale, this is a single value (0–255). For color, it’s an RGB triplet (e.g., (255, 0, 0) for red).
• Neighborhood relationships: Pixels interact with their 4-connected (up/down/left/right) or 8-connected (diagonal included) neighbors for operations like smoothing or edge detection.

Pixel Relationships Visualized

graph LR
    A["Pixel (x,y)"] -->|"4-connected"| B["(x+1,y)"]
    A -->|"4-connected"| C["(x,y+1)"]
    A -->|"8-connected"| D["(x+1,y+1)"]
    A -->|"8-connected"| E["(x-1,y+1)"]

Key Idea: The choice of connectivity affects operations like morphological dilation (used in license plate detection) or edge thinning.

2. The Image Processing Pipeline

All digital image systems follow this functional block diagram (critical for exams—draw this!):

flowchart LR
    A["Image Acquisition"] --> B["Preprocessing"]
    B --> C["Image Enhancement/Restoration"]
    C --> D["Image Analysis"]
    D --> E["Image Compression"]
    E --> F["Image Display/Transmission"]

Purpose of Each Block:

3. Image Sampling and Quantization: The Trade-off

Sampling

• Definition: Converting a continuous scene into a discrete grid of pixels.
• Resolution: Determined by the number of pixels (e.g., 1024×1024 = 1.048M pixels).
• Worked Example: A 1024×1024 image with 512 colors requires:
  • Color depth: ${\log }_2(512)=9$ bits/pixel.
  • Total bits: $1024\times 1024\times 9=9.436\times 10^6$ bits ≈ 1.18 MB.

Quantization

• Definition: Mapping continuous intensity values to a finite set (e.g., 8-bit grayscale = 256 levels).
• Impact: Lower quantization → more compression but loss of detail (e.g., 2-bit color = 4 shades).
• Real-World Tie: WhatsApp’s photo compression uses JPEG’s DCT + quantization to reduce file sizes by ~80% while preserving perceived quality.

4. Image Enhancement vs. Restoration: Key Differences

Worked Example: Histogram Equalization

Given: Gray-level histogram:

Steps:

1. Compute cumulative distribution function (CDF): $CDF(k)={\sum }_{i=0}^k\text{pixels}(i)$. Example: $CDF(2)=300+500+1000=1800$.
2. Normalize CDF to new intensity range (0–7): $s_k=\lfloor \frac{CDF(k)}{N}\times L\rfloor$, where $N=3900$ (total pixels), $L=7$. For $k=2$: $s_2=\lfloor \frac{1800}{3900}\times 7\rfloor =3$.

Result: The enhanced image will have a flatter histogram, improving contrast.

5. Image Compression Standards

Real-World Tie: NEPSE’s stock chart images use PNG compression to preserve sharp lines in candlestick graphs, while Daraz’s product photos use JPEG for faster loading.

6. Common Image Operations

In the Real World

1. WhatsApp Photo Compression:

   • Idea Used: JPEG compression (DCT + quantization).
   • How: Converts RGB images to YCbCr color space, applies DCT to 8×8 blocks, and quantizes high-frequency coefficients to reduce file size by ~80%.
   • Result: Faster uploads and lower data usage for users in Nepal.

2. Ncell’s Satellite Imagery:

   • Idea Used: Fourier transforms for noise filtering.
   • How: Captured images are corrupted by atmospheric noise. A low-pass filter in the frequency domain removes high-frequency noise while preserving edges (e.g., roads, buildings).
   • Output: Clearer images for disaster monitoring.

3. Daraz’s Product Tagging:

   • Idea Used: Image segmentation + morphological operations.
   • How: After capturing product photos, thresholding separates the product from the background. Dilation then fills gaps in the product’s outline for accurate tagging.
   • Example: A red T-shirt is segmented from a white background using Otsu’s thresholding.

Worked Example: Calculating Image Size

Problem: A grayscale image has dimensions 1024×1024 and uses 512 colors. Calculate its size in MB. Solution:

1. Colors to bits: ${\log }_2(512)=9$ bits/pixel.
2. Total pixels: $1024\times 1024=1,048,576$ pixels.
3. Total bits: $1,048,576\times 9=9,436,928$ bits.
4. Convert to MB: $\frac{9,436,928}{8\times 1024\times 1024}\approx 1.13$ MB.

Exam Tip

1. Block Diagrams: Always draw the image processing pipeline (acquisition → display) and label each block’s purpose. This is a high-mark question in TU/PU exams.
2. Differentiate Enhancement vs. Restoration:
   • Enhancement = "making it look better" (e.g., histogram equalization).
   • Restoration = "undoing damage" (e.g., Wiener filter for blur).
3. Quantization Calculation:
   • Memorize: Number of colors = $2^{\text{bits per pixel}}$.
   • Example: 512 colors = 9 bits/pixel.
4. Real-World Links:
   • Tie JPEG to social media (WhatsApp, Instagram), PNG to logos, and Fourier transforms to satellite imagery (Ncell).
5. Pixel Math:
   • For image size questions, always:
     1. Calculate bits per pixel (color depth).
     2. Multiply by total pixels.
     3. Convert to MB (divide by $8\times 1024\times 1024$).
6. Avoid Common Mistakes:
   • Don’t confuse sampling (spatial resolution) with quantization (intensity levels).
   • In histogram equalization, normalize the CDF to the new range (e.g., 0–7 for 8 levels).

Final Visual Summary:

mindmap
  root((Digital Image Processing))
    Pixel Basics
      Coordinates (x,y)
      Intensity (Grayscale/RGB)
      Neighborhood (4/8-connected)
    Processing Pipeline
      Acquisition
      Preprocessing
      Enhancement/Restoration
      Analysis
      Compression
      Display
    Key Operations
      Point Operations
      Spatial Filtering
      Frequency Filtering
      Morphological
    Real-World Examples
      WhatsApp (JPEG)
      Ncell (Fourier)
      Daraz (Segmentation)