Stop 10 · Computer VisionTopics · 1/9

Image Preprocessing

In short

A model trained on neat 224x224 photos will quietly fail on a raw 4000x3000 phone picture. Preprocessing makes every image look the way the model expects, every single time.

Resizing, normalizing, and formatting images before feeding them into a model — small mistakes here (wrong normalization, wrong channel order) silently break model accuracy.

The blob trims photos of different sizes with a paper cutter so they fit identical square frames.

Think of it like this

The passport photo studio

You walk into a photo studio with a selfie, a group photo and a blurry scan. The studio crops each one, fixes the size, evens out the brightness and prints them all in the same frame. A model is just as fussy as the passport office.

Watch it happen

One photo on its way into a model

Load

Resize

To RGB

Normalise

Tensor

1/5A 3000x4000 JPEG is read as a grid of pixel values from 0 to 255.

The big ideas

What you need to know

Fixed size in

Most CNNs expect one input size. Resize and crop so nothing important gets squashed or cut off.

Channel order matters

BGR vs RGB, and channels-first vs channels-last. Get it wrong and the model sees blue skies as orange.

Normalise like training

Use the exact mean and std the pretrained model used. Many ImageNet models use mean [0.485, 0.456, 0.406].

Same steps at train and test

If training resized one way and your app resizes another, accuracy drops silently. Share one preprocessing function.

See it in code

Walk through the code

python

1from PIL import Image2from torchvision import transforms3 4preprocess = transforms.Compose([5    transforms.Resize(256),6    transforms.CenterCrop(224),7    transforms.ToTensor(),  # HWC 0-255 -> CHW 0.0-1.08    transforms.Normalize(mean=[0.485, 0.456, 0.406],9                         std=[0.229, 0.224, 0.225]),10])11 12img = Image.open("chai.jpg").convert("RGB")13x = preprocess(img).unsqueeze(0)14print(x.shape)  # torch.Size([1, 3, 224, 224])

1/3Resize the short side to 256, then cut a 224x224 square from the centre.

The classic ImageNet preprocessing recipe with torchvision. Pretrained models expect exactly this.

Save yourself the pain

Common mistakes

Oops

Feeding OpenCV images straight into a model trained on RGB.

Do this instead

Convert with cv2.cvtColor(img, cv2.COLOR_BGR2RGB) first. The model will not complain, it will just be wrong.

Oops

Stretching every image to a square, whatever its shape.

Do this instead

Resize the short side and centre-crop, or pad to a square, so objects keep their real proportions.

Your mission

Break it on purpose

~25 min

Quick check

Your model works in the notebook but gives odd results in the app. The app uses OpenCV to read images. What is the most likely cause?

Go deeper

1Practice resizing, cropping and normalizing images with PIL/OpenCV.
2Learn why normalization statistics must match what a pretrained model expects.
3Build a reusable preprocessing pipeline for a vision project.