The standard practice of padding images with zeros offers significant computational efficiency because adding zeros is mathematically trivial. Furthermore, deep learning operators can be specifically engineered to take advantage of zero-padding implicitly, executing the operation without needing to allocate any additional memory for the expanded tensor.

Computational Benefits of Zero-Padding

Beyond preserving spatial dimensions, zero-padding provides an unexpected representational advantage for convolutional neural networks. By extending the boundaries of an image with zero-valued 'whitespace', the network can learn to recognize these artificial borders, thereby allowing the CNN to implicitly encode and utilize spatial position information from within the input.

Encoding Implicit Position Information with Zero-Padding

While extending inputs with zeros is the predominant method due to its efficiency, there are numerous alternative, nonzero padding techniques available for convolutional layers. However, these alternatives are generally only utilized in specific cases where standard zero-padding produces unwanted artifacts in the network's output.

Alternatives to Zero-Padding

To resolve the issues of shrinking output dimensions and underutilized border pixels, we can apply padding to our input. Padding involves adding extra rows and columns of filler pixels—typically with values set to zero—around the boundary of the input image, effectively increasing its spatial dimensions before the cross-correlation operation is computed.

University of Michigan - Ann Arbor

Claude

When applying standard cross-correlation operations without padding, we encounter two main limitations. First, the spatial dimensions of the output shrink after each layer, which restricts the total number of successive convolutional layers that can be applied before the feature map is reduced to a $$1 	imes 1$$ size. Second, because the convolution kernel must fit entirely within the image boundaries, the pixels on the perimeter and corners participate in far fewer window calculations than the central pixels, leading to a significant loss of boundary information.

Limitations of Unpadded Convolution

Machine Learning Yearning, a free ebook from Andrew Ng, teaches you how to structure Machine Learning projects.
https://www.deeplearning.ai/machine-learning-yearning/

Note: The content of the book is aligned with the Coursera Deeplearning.ai specialization.  https://www.deeplearning.ai/deep-learning-specialization/ 

Learn Before

Related

Learn After