Feature Maps and Multiple Filters
Understand how filters create spatial maps of learned visual responses.
Feature Maps and Multiple Filters explained visually
See how several filters produce complementary maps for one image.
What you will learn
Follow filters, pooling, flattening, and classification through complete CNN examples.
ONE FILTER
One filter creates a grid of local responses called a feature map.
LOCAL RESPONSES
Strong activations show where the image matches the filter's learned pattern.
FEATURE MAP
A convolutional layer uses several filters to produce complementary feature maps.
MULTIPLE FILTERS
Deeper layers combine these maps while retaining useful spatial information.
Quick check: test your understanding
What is a feature map?
A spatial grid of responses produced by one filter.
Why use multiple filters?
Each filter can detect a different pattern.
What does a strong activation mean?
The local image region closely matches the learned pattern.
Accessible referenceLesson 47 transcript305 words · Matches the published narration and captionsOpen transcript
When a filter moves across an image, it produces an output value at every valid position. The complete grid of output values is called a feature map, or activation map. A strong positive value indicates that the image region matches the pattern represented by the filter.
A small value indicates a weak match. Imagine applying an edge-detection filter to a street image. Regions containing strong edges, such as building boundaries or vehicle outlines, may produce large activations. Smooth areas, such as a clear sky, may produce smaller responses. One filter produces one feature map.
A convolutional layer normally contains many filters because one pattern cannot describe everything in an image. One filter may respond to vertical edges. Another may respond to horizontal edges. Additional filters may detect curves, textures, colors, or other learned patterns. If a layer contains thirty-two filters, it produces thirty-two feature maps.
These maps become the input to the next layer. A deeper filter does not necessarily process the original photograph directly. It may examine the feature maps created by the previous layer. This allows the deeper filter to combine several simpler detections. For example, one feature map may indicate curved edges while another indicates a particular texture.
A later layer can combine those activations into a more meaningful structure. Feature maps also preserve spatial location. If a filter detects a pattern near the upper-right corner, the strong activation appears in the corresponding area of the feature map. The feature map may be smaller than the original image depending on the filter size, stride, and padding.
The important idea is that convolution transforms an image into a collection of learned visual responses. Each feature map records where one pattern appears and how strongly the image matches it. In the next lesson, we will see how stride controls filter movement and output size.