Feature Maps

When a Convolutional Neural Network processes an image, it does not see the image the way humans do.

Instead, it creates a series of internal representations called feature maps. These feature maps are the true “vision” of a CNN.

In this lesson, you will understand what feature maps are, how they are created, and why they are the most important internal component of CNNs.

---

What Is a Feature Map?

A feature map is the output of a convolution operation.

Each feature map highlights a specific pattern present in the input image.

Instead of raw pixels, feature maps represent learned visual features.

---

From Image to Feature Map

Think of an input image as raw information.

When a convolution filter slides over the image:

• It reacts strongly to certain patterns
• It reacts weakly to others

The result of this reaction is a feature map.

Every filter produces exactly one feature map.

---

One Filter → One Feature Map

This relationship is critical:

• 1 filter = 1 feature map
• 10 filters = 10 feature maps
• 64 filters = 64 feature maps

More filters mean the network can detect more types of visual patterns.

---

What Do Feature Maps Actually Represent?

Early feature maps usually detect:

• Edges
• Lines
• Simple textures

As depth increases, feature maps represent:

• Corners
• Shapes
• Object parts

In deep layers, feature maps represent high-level concepts.

---

Why Feature Maps Are Smaller Than the Input

Feature maps are usually smaller than the input image because:

• Pooling layers reduce size
• Strides skip positions
• Padding choices affect dimensions

Smaller maps mean faster computation without losing important information.

---

Depth of Feature Maps (Channels)

Feature maps are stacked together.

This stack forms the depth of the CNN layer.

For example:

• Input image: 3 channels (RGB)
• First conv layer: 32 feature maps
• Next layer: 64 feature maps

Depth grows as the network learns richer features.

---

Feature Maps Are Not Images

A common misunderstanding:

Feature maps are not meant to look like images.

They are mathematical representations, not visual pictures for humans.

However, we can visualize them to understand what the network is focusing on.

---

Activation Strength Matters

Brighter or higher values in a feature map mean:

• Strong presence of a feature
• High confidence in that pattern

Dark or low values mean the feature is absent.

CNNs make decisions based on these activations.

---

Feature Maps vs Filters (Important Distinction)

---

How Feature Maps Evolve Through the Network

CNN layers follow a pattern:

• Early layers → simple features
• Middle layers → combined features
• Deep layers → semantic meaning

This gradual abstraction is why CNNs work so well.

---

Is This Theory or Coding?

This lesson is conceptual.

You are building intuition for:

• What CNNs learn internally
• Why depth matters
• How vision emerges from math

In upcoming lessons, you will visualize and generate feature maps using code.

---

Practice Questions

---

Observation Exercise

• Imagine detecting edges in an image
• Now imagine detecting corners using edges
• Finally imagine detecting faces using corners

Each step represents deeper feature maps.

---

Quick Recap

• Feature maps are outputs of convolution
• Each filter creates one feature map
• Feature maps show where patterns exist
• Depth increases feature complexity
• CNNs build meaning layer by layer

Next lesson: CNN Architectures Overview — how layers are organized into real models.