[Python] MNIST multi digit prediction

I tried to train a model on mnist dataset. It's working fine for single digit prediction and I got 96.92% accuracy on test data too but I have everything but it's not working for multidigit. even though I took an image with spacing in between and it was able to split and preprocess it properly. I have tried multiple threshold values, tried to increase and decrease hidden layers.

the input image had 1234 but model is predicting 3356.enter image description here

import numpy as np
import pandas as pd
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.metrics import accuracy_score
import matplotlib.pyplot as plt
import tensorflow
from tensorflow import keras
from tensorflow.keras import Sequential
from tensorflow.keras.layers import Dense
from PIL import Image
def apply_threshold(img, threshold=125):
img = np.where(img > threshold, 1, 0).astype(np.uint8)
return img

def image_background(img):
if img.mean() > 127:
img = 255 - img
return img

def padding_and_resize(img):
h, w = img.shape
if h < w:
new_img = np.zeros((int((w - h) / 2), w))
new_img = np.append(new_img, img, axis=0)
zeros = np.zeros(((w - h) - int((w - h) / 2), w))
new_img = np.append(new_img, zeros, axis=0)
img = new_img
elif h > w:
new_img = np.zeros((h, int((h - w) / 2)))
new_img = np.append(new_img, img, axis=1)
zeros = np.zeros((h, (h - w) - int((h - w) / 2)))
new_img = np.append(new_img, zeros, axis=1)
img = new_img

img = Image.fromarray(img.astype(np.uint8))
img = img.resize((28, 28))
img = np.array(img)
return img

def reshaping_and_standardizing(img):
return img.reshape(-1) / 255.0

def split_digits(img):
img_mean = img.mean()
new_img = []

for i in img:
if np.sum(i) != 0:
new_img.append(i)

new_img = np.array(new_img)
output_digits = []

moving_on_digit = False
start = 0
end = 0
for i in range(len(new_img[0])):
if np.sum(new_img[:, i]) != 0 and (not moving_on_digit):
moving_on_digit = True
start = i
elif np.sum(new_img[:, i]) == 0 and moving_on_digit:
moving_on_digit = False
end = i
output_digits.append(new_img[:, start:end + 1])

if moving_on_digit:
output_digits.append(new_img[:, start:len(new_img[0])])

return output_digits

def img_processor(img, threshold=None):
if threshold is None:
threshold = img.mean()
img = image_background(img)
img = apply_threshold(img, threshold=threshold)
img = padding_and_resize(img)
img = reshaping_and_standardizing(img)
return img

(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()

x_train = pd.DataFrame(img_processor(img) for img in x_train)
x_test = pd.DataFrame(img_processor(img) for img in x_test)

y_train = pd.DataFrame(np.array([[0] * i + [1] + [0] * (9 - i)]).reshape(10) for i in y_train)

model = Sequential()
model.add(Dense(784, activation='relu', input_dim=784))
model.add(Dense(128, activation='relu'))
model.add(Dense(64, activation='relu'))
model.add(Dense(10, activation='softmax'))

model.summary()

model.compile(loss='categorical_crossentropy', optimizer='Adam')
model.fit(x_train, y_train, epochs=11)

y_predict = model.predict(x_test)

y_predict = np.argmax(y_predict, axis=1)

str(round(accuracy_score(y_test, y_predict) * 100, 2)) + "%"

# testing

from PIL import Image

# single digit testing
img = Image.open('/content/drive/MyDrive/datasets/image_set/Screenshot 2025-07-30 052716.png')
img = img.convert('L')
img = img.resize((28, 28))
img_array = np.array(img)
img_array = img_array.reshape(1, 784)
if img_array.mean() > 127:
img_array = 255 - img_array
img_array = img_array / 255.0
y_predict = model.predict(img_array)
np.argmax(y_predict, axis=1)

img = Image.open('/content/Screenshot 2025-07-31 175136.png').convert('L')
img = np.array(img)
img = image_background(img)
img = apply_threshold(img, threshold=img.mean())
images = split_digits(img)
digit = 0
for img in images:
img = padding_and_resize(img)
plt.imshow(img, cmap='gray')
plt.axis('off')
plt.show()
img = img.reshape(1, 784) / 255.0
prediction = model.predict(img).argmax(axis=1)[0]
digit *= 10
digit += prediction
print(digit)enter image description here

Continue reading...