Feature Map Size in Convolutional Neural Networks

Understanding Feature Map Size in Convolutional Neural Networks

Convolutional Neural Networks (CNNs) have become a cornerstone in the field of deep learning, particularly for tasks involving image and video recognition. A critical aspect of CNN architecture is the feature map size, which undergoes transformation at each convolutional layer. Knowing how to compute it is fundamental for building effective models.

The Formula

The feature map size after a convolutional layer in a CNN is determined using the following formula:

outputSize = (inputSize   kernelSize + 2 * padding) / stride + 1

Here’s a breakdown of each parameter involved:

• inputSize: The size of the input feature map (measured in pixels).
• kernelSize: The size of the convolutional kernel (measured in pixels).
• padding: The number of zero pixels added to the border of the input (measured in pixels).
• stride: The number of pixels by which the kernel moves across the input feature map (measured in pixels).

Inputs and Outputs

Inputs

• inputSize: Integer, number of pixels (px).
• kernelSize: Integer, number of pixels (px).
• padding: Integer, number of pixels (px).
• stride: Integer, number of pixels (px).

Output

• outputSize: Integer, number of pixels (px).

Real life Example

Consider a popular use case where you have an input image of size 224x224 pixels. You apply a convolutional layer with a kernel size of 3x3, padding of 1, and a stride of 1. Here’s how you compute the feature map size:

inputSize: 224, kernelSize: 3, padding: 1, stride: 1

Plugging these values into our formula:

outputSize = (224   3 + 2 * 1) / 1 + 1 = 224

The resulting feature map will still be 224x224 pixels.

Data Validation

For this calculation to work, all input parameters must be greater than zero. Moreover, ensure that the stride is an integer that divides the modified input size (inputSize kernelSize + 2 * padding) evenly, else the feature map size will not be an integer and the formula will break.

Example Values:

• inputSize= 32
• kernelSize= 5
• padding= 2
• stride= 1
• outputSize= resulting feature map size

Output:

• outputSize= 32

Summary

Calculating the feature map size in convolutional neural networks is crucial for model architecture and optimization. By understanding and correctly employing the formula (inputSize kernelSize + 2 * padding) / stride + 1, data scientists and engineers can design more efficient networks, improving performance and efficacy.

Frequently Asked Questions (FAQs)

By following these guidelines and understanding each parameter, you can harness the full potential of Convolutional Neural Networks and optimize your deep learning models efficiently.