Image resizing is a computer vision preprocessing step used to give dataset images a common target width and height when an algorithm requires fixed-size inputs. Resizing uses interpolation to determine pixel values at the target dimensions; reducing the spatial dimensions is downsampling.

University of Michigan - Ann Arbor

Google

Acquired data is usually messy and can come from different sources, angles, color scales, etc. In order to use machine and deep learning techniques they need to be standardized. Usually the goal is to increase the accuracy while reducing the complexity of the model used. There is no single answer to preprocessing but there are a group of techniques commonly chosen from.

Computer Vision Pre-Processing

Color images are stored as size 3 by width by height where the 3 refers to the RGB (red., green, blue) dimensions. For a lot computer vision problems color isn't important and the size of the image can be reduced to 1 dimension that is grayscale. This reduces the space and computational complexity of the model in training.

Converting to grayscale

Involves augmenting the dataset with scaled, rotated, decolorized, etc versions of the original image. This makes it so your model is invariant to a wide variety of variations of an image. Makes it more likely that it'll recognize objects when they appear in any form.

Data augmentation

A common computer vision preprocessing step is to standardize the pixel intensities of each RGB (red, green, and blue) color channel independently. For every pixel in a given channel, the channel's mean value is subtracted and the result is divided by the channel's standard deviation. When fine-tuning a model pretrained on ImageNet, the standardization must use the source dataset's statistics rather than those of the target dataset. The canonical ImageNet channel means are $$[0.485, 0.456, 0.406]$$ and the corresponding standard deviations are $$[0.229, 0.224, 0.225]$$.

Learn Before

Related