Image Classification · Study

Topics covered

Inspect misclassified samples behind image classification accuracy to examine the effects of preprocessing and data splitting.

The structure of image samples

The small grayscale images in load_digits also have pixel arrays, ground truth, and separate samples. First define normalization and the training/evaluation split, then inspect actual misclassified images.

What to Examine after Average Accuracy

Examine which classes or blurred inputs account for errors. Evaluate results with small added noise as a separate condition, and keep transformed samples and their originals from being mixed across training and evaluation sets.

Worked example

Use Classes 0–4 from the 8×8 images in sklearn load_digits, which requires no external download. Use a stratified 70/30 split with seed=7.

from sklearn.datasets import load_digits
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression
D=load_digits();mask=D.target<5;X=D.data[mask];y=D.target[mask]
Xtr,Xte,ytr,yte=train_test_split(X,y,test_size=.3,random_state=7,stratify=y)
scaler=StandardScaler().fit(Xtr);clf=LogisticRegression(max_iter=1000).fit(scaler.transform(Xtr),ytr)
print(clf.score(scaler.transform(Xte),yte))

1.0

Reference example

View more examples

Fashion-MNIST Example · Image Arrays and Flatten

Refer to the image display and array shape in the introduction. This activity uses 8×8 images from load_digits.