Image Classification

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

Run import numpy as np and import matplotlib.pyplot as plt in a code cell before the practice code. Use np for array operations and plt for plots.

Starting conditions

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.

Procedure

1. Inspect one image's 8×8 array and its Flattened 64 features.
2. Perform baseline classification using StandardScaler fitted on train data and LogisticRegression(max_iter=1000).

Results to retain

• A data structure table, example image panel, and Test Accuracy.

Starting conditions

Select up to 6 misclassified Test samples and display each image, true label, prediction, and maximum probability together. If there are none, select low-confidence samples.

Procedure

1. Find Indices where Test predictions differ from true labels and select the first 6 at most. If there are no misclassifications, select up to 6 in ascending order of maximum predicted probability.
2. Plot the original 64 values for each Index as a $8\times8$ image, with Test Index, true label, prediction, and maximum probability in the title.
3. Distinguish features observed in the plots from model predictions in your explanation. Do not retrain the model or Scaler.

Results to retain

• An image panel showing Test Index, true label, prediction, and probability.

Starting conditions

Add small noise using seed=7 to one selected test image, clip to the 0~16 range, and compare predictions.

Procedure

1. Choose one test image from Activity 2 and record its index. Add small noise with Seed 7 and limit the pixel range using np.clip(...,0,16). For example, you can use normal noise with mean 0 and standard deviation 0.5.
2. Display original and noisy images using the same brightness range. Feed each image into the same Scaler and model to obtain the predicted Class and per-Class probabilities.
3. Record the noise distribution and magnitude used, both predictions, and their probabilities. Explain whether visual changes and prediction changes mean the same thing.

Results to retain

• Compare original and noisy images, along with the same model's predictions and probabilities.