Character Classification

Using images of written characters

Sam Hopkins

#!/usr/bin/python

# by Sam Hopkins

import time

import numpy as np
import matplotlib.pyplot as plt
import pandas as pd
import seaborn as sn

from sklearn.datasets import fetch_openml
from sklearn.cluster import KMeans
from sklearn.metrics import confusion_matrix

mnist = fetch_openml('mnist_784', version=1)

# Randomly generate 60000 training/10000 testing indices
indices = np.random.permutation(len(mnist['target']))
trainingIndices = indices[:60000]
testingIndices = indices[60000:70000]

# Split data into training and testing using random indices
Training = [mnist['data'].iloc[trainingIndices], mnist['target'].iloc[trainingIndices]]
Testing = [mnist['data'].iloc[testingIndices], mnist['target'].iloc[testingIndices]]

# Split data into training and testing using known indices
# Training = [mnist['data'].iloc[:60000], mnist['target'].iloc[:60000]]
# Testing = [mnist['data'].iloc[60000:70000], mnist['target'].iloc[60000:70000]]

# Training histogram
plt.hist(np.sort(np.array(Training[1])), edgecolor='black')
plt.xlabel('digit 0-9')
plt.ylabel('number of images')
plt.title('Training digit frequency')
plt.show()

# Testing histogram
plt.hist(np.sort(np.array(Testing[1])), edgecolor='black')
plt.xlabel('digit 0-9')
plt.ylabel('number of images')
plt.title('Testing digit frequency')
plt.show()

# Split training images into categories
categories = [[] for i in range(10)]
for i in range(len(Training[1])):
  c = int(Training[1].iloc[i])
  categories[c].append(Training[0].iloc[i])

# Identify center images for 9 clusters for each of 10 categories (90 total center images)
centerImages = [[] for i in range(10)]
for c in range(len(categories)):
    kmeans = KMeans(n_clusters=9, random_state=42)
    kmeans.fit(categories[c])

    clusterCenters = kmeans.cluster_centers_.reshape(-1, 28, 28)
    centerImages[c] = clusterCenters
    
    # Display center images
    fig, axes = plt.subplots(3, 3, figsize=(8, 8))
    for i, a in enumerate(axes.flat):
        a.imshow(clusterCenters[i], cmap='gray')
        a.axis('off')
        a.set_title(f'digit {c} cluster {i}')

    plt.tight_layout()
    plt.show()

trainingImages = np.concatenate(centerImages)
trainingImages = [image.reshape(-1) for image in trainingImages]
testImages = []
for i in range(len(Testing[0])):
    testImages.append(Testing[0].iloc[i].values.reshape(-1))

predictions = []

predictedImages = []
clusters = []

time0 = time.time()
for i in range(len(testImages)):
    distances = np.linalg.norm(trainingImages - testImages[i], axis=1)

    min = np.argmin(distances)
    predictions.append(min // 9)
        
    predictedImages.append(trainingImages[min])
    clusters.append(min % 9)
    
time1 = time.time()
predTime = time1 - time0

actual = np.array(Testing[1]).astype(int)

accuracy = 0.0;
for i in range(len(predictions)):
    accuracy += 1.0 if actual[i] == predictions[i] else 0.0
    
accuracy /= len(predictions)/100

confMatrix = confusion_matrix(actual, predictions)

plt.figure(figsize=(8, 8))
sn.heatmap(confMatrix, annot=True, fmt=".4g")

plt.xlabel('true')
plt.ylabel('predicted')
plt.title(f"test accuracy {accuracy} %  prediction time {predTime:.2f} seconds")
plt.show()

First 20 images tested vs predicted cluster center

for i in range(20):
    testImage = testImages[i].reshape(28, 28)
    closeCluster = predictedImages[i].reshape(28, 28)
    fig, axes = plt.subplots(1, 2, figsize=(8, 8))
        
    axes[0].imshow(testImage, cmap='gray')
    axes[0].axis('off')
    axes[0].set_title(f'test {i} label {actual[i]} ')
        
    axes[1].imshow(closeCluster, cmap='gray')
    axes[1].axis('off')
    axes[1].set_title(f'closest digit {predictions[i]} cluster {clusters[i]}')

    plt.tight_layout()
    plt.show()