Discuss the importance of 'Normalization' in image preprocessing.

Example Answer

I would match preprocessing to the model and its training recipe. Dividing an eight-bit image by 255 changes its numerical scale; channel normalization then subtracts a mean and divides by a standard deviation. Neither operation makes images invariant to illumination or automatically reduces inference computation. For pretrained weights, use their documented transforms. If statistics are estimated from our data, estimate them from the training split only and apply the same fixed transformation during validation and serving.

Make it your own

Illustrative contract: color order, numeric range, resize/crop policy, mean, standard deviation and weight version. Compare a known image through training and inference preprocessing.

Why this works

An RGB/BGR or range mismatch can damage predictions without changing the model. Follow-up: what checks would detect accidental double normalization?

Sources: TorchVision Normalize, TorchVision model preprocessing.