Feature Maps and Filters

In the previous lesson, we learned how padding and stride control the size and movement of convolution operations.

Now we focus on the true learning components inside a Convolutional Neural Network — filters and feature maps.

This lesson explains what CNNs actually learn and how raw pixels slowly turn into meaningful patterns.

What Is a Filter?

A filter (also called a kernel) is a small matrix of numbers that slides across the input image.

Each filter is responsible for detecting a specific pattern.

These patterns are not hand-written. They are learned automatically during training.

How Filters Learn

At the beginning of training, filters contain random values.

As training progresses, backpropagation adjusts filter values to reduce prediction error.

Over time, filters specialize in detecting:

• Edges • Corners • Textures • Shapes

What Is a Feature Map?

A feature map is the output produced when a filter is applied to the input.

Each filter produces one feature map.

If a convolution layer has 32 filters, it produces 32 feature maps.

Why Feature Maps Matter

Feature maps show where a specific pattern exists in the image.

High values mean the pattern strongly appears at that location.

Low values mean the pattern is absent.

Hierarchical Learning

CNNs learn in layers.

Early layers detect simple features:

• Horizontal edges • Vertical edges • Color contrasts

Middle layers detect combinations:

• Corners • Curves • Object parts

Deep layers detect:

• Faces • Wheels • Text • Complex objects

Filters in Code

Let us define a convolution layer with multiple filters.

from tensorflow.keras.layers import Conv2D

conv_layer = Conv2D(
    filters=64,
    kernel_size=(3,3),
    padding="same",
    activation="relu"
)

This layer creates:

• 64 learnable filters • 64 feature maps

Why Small Filters Are Preferred

Modern CNNs mostly use 3 × 3 filters.

Smaller filters:

• Capture fine details • Reduce parameters • Allow deeper networks

Stacking small filters creates a larger receptive field without heavy computation.

Receptive Field Intuition

The receptive field is the portion of the input image that affects a neuron.

As layers go deeper, the receptive field increases.

This allows neurons to understand larger context.

Real-World Example

In face recognition:

Early filters detect edges of eyes.

Mid-level filters detect eyes and nose.

Deep filters detect entire faces.

This hierarchy makes CNNs powerful.

Mini Practice

Think about this:

Why would using many filters in early layers increase learning capacity?

Exercises

Quick Quiz

In the next lesson, we will study classic CNN architectures and see how feature maps are stacked to build deep models.