What is feature extraction?
Feature extraction derives a smaller set of informative values from raw data. A single photo holds millions of pixel values, most of them redundant. Feature extraction reduces the photo to values that describe what matters, such as edges, textures, shapes, or a learned summary of the whole image. The model then works from those values.
Older computer vision systems used hand-designed extractors that measured edges, corners, and color histograms. Deep learning systems instead pass the image through a pre-trained neural network, such as a convolutional neural network, and keep the output of a late layer. That output is an embedding, a vector of a few hundred to a few thousand numbers. Feature selection keeps a subset of existing columns, while feature extraction creates new values, which makes extraction a form of dimensionality reduction.
Feature extraction in practice
A county assessor’s office reviews aerial imagery to find unpermitted construction. A pre-trained network converts each parcel’s image from the latest flight and from the previous one into feature vectors. Parcels whose vectors changed the most go to an appraiser, who checks for new pools, decks, or additions. Flights in different seasons can shift the vectors on their own, so the office compares images taken under similar conditions. Reusing a pre-trained extractor, a form of transfer learning, lets a small team build classification, visual search, and change detection without training a vision model from scratch.