Feature Extraction from Images

Introduction

• Feature extraction is crucial in image data processing.
• It involves understanding how a computer interprets image data, often stored as a matrix.

Simple Approach to Feature Extraction

Black and White Images

• Images are interpreted as matrices.
• Example: A 22x16 pixel image is stored as a 352 pixel matrix.
• Pixel values range from 0 (black) to 255 (white).
• Features can be represented by flattening the matrix into a 1x352 vector.

Color Images

• Stored as three matrices (channels) for blue, green, and red.
• Similar approach by calculating mean pixel values across channels.

Limitations

• Simple approach is prone to noise (e.g., background, color differences, expressions).
• Computationally expensive and retains unnecessary information.

Advanced Feature Extraction: HOG (Histogram of Oriented Gradients)

Concept

• HOG computes pixel-wise gradients and orientations.
• Represents images in a noise-reduced form.

Steps to Build HOG Representation

1. Image Pre-processing

• Standardize image size, e.g., resizing to 32x64 pixels.
• Divide resized image into smaller cells, e.g., 4x4 pixels.

2. Calculating Gradients and Orientations

• Compute gradients in x (gx) and y (gy) directions.
• Calculate gradient magnitude and orientation.
• Use Pythagoras theorem for magnitude and inverse tangent for orientation.

3. Plotting Histograms

• Orientations are binned (e.g., 20-degree bins, 9 bins total).
• Distribute gradient magnitudes into bins based on orientation.
• Results in a 1x9 feature matrix per cell.

4. Normalization

• Form blocks of cells for normalization to reduce sensitivity to lighting variations.
• Calculate block features and normalize using root sum of squares.
• Shift blocks across the image to compute normalized vectors for all blocks.

Feature Matrix

• HOG results in a 1x3780 dimensional image descriptor.

Dimensionality Reduction

• The high dimensionality of features can lead to computational expense and overfitting.
• PCA (Principal Component Analysis) is used to reduce dimensions.

Conclusion

• HOG is an effective feature descriptor for minimizing noise and focusing on crucial information in images.