CODEWITHSUNDEEP
tensorflowopencvmnistdigitspython
fhs14647
fhs14647

Reputation: 163

Why is every prediction for my MNSIT digit project wrong?

I have done this basic project with Anaconda notebook. Everything runs fine, but every prediction with my own digit picture is wrong. I am using the MNIST Set for digital numbers and I am trying to paint my own digit with black background and white painting. But every prediction is wrong. Could one see what´s missing in the code?

enter image description here

enter code here
# Install TensorFlow
import tensorflow as tf

# Import matplotlib library
import matplotlib.pyplot as plt 

#Import numpy
import numpy as np

#import cv
import cv2

#Dataset
mnist = tf.keras.datasets.mnist

(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0
model = tf.keras.models.Sequential([

tf.keras.layers.Flatten(input_shape=(28, 28)),
tf.keras.layers.Dense(128, activation='relu'),
tf.keras.layers.Dropout(0.2),
tf.keras.layers.Dense(10, activation='softmax')])

model.compile(optimizer='adam',
          loss='sparse_categorical_crossentropy',
          metrics=['accuracy'])
model.fit(x_train, y_train, epochs=5)
print("Evaluierung");
model.evaluate(x_test,  y_test)

plt.imshow(x_train[1], cmap="gray") # Import the image
plt.show() # Plot the image
predictions = model.predict([x_train]) # Make prediction, works perfect
print(np.argmax(predictions[1])) # Print out the number, works perfect

# my own picture, black background, white color, 28 px * 28 px in size
img = cv2.imread("bild1.png", cv2.IMREAD_GRAYSCALE)
plt.imshow(img) # Import the image
plt.show() # Plot the image

cv2.imshow("image",img)
cv2.waitKey(2000)
img = img/255.0
img = img.reshape(1,28,28)
pred = model.predict_classes(img)
print("Prediction: ", pred)

Every prediction from the training and test data are correct, my own pictures are ALL wrong with NO error code! It would be great if you could help me

Upvotes: 1

Views: 292

Answers (2)

fhs14647
fhs14647

Reputation: 163

Thanks very much for your answer

I changed the code for normalizing like you told me, but it didn´t change anything.

Then I looked in Google for some handwritten digits (outside the mnist database) with black background, I resized it and took them for predict -> to my surprise every pic is recognized correctly!! So the problem I think is not my code but the way I paint my own handwritten digits. I make it with photoshop, new gray-scale template, black background, white painting und resizing it to 28 * 28 pixels. But my own painting is not recognized correcty - no error code, but simply a wrong prediction. Do you have an idea what is wrong with my painting ... ?? Thanks again!!

Upvotes: 0

thushv89
thushv89

Reputation: 11333

There's nothing particularly wrong with the way you are doing this. However I can give few pointers to why you are getting these results.

Loading images from JPEG/PNG

Sometimes when you load the images using image loading libraries, they will have values clipped as opposed to being between exactly 0-255. Therefore, when normalizing, instead use something like,

img = img - np.min(img)
img = img/np.max(img)

More generalization

Training model on the raw 60000 samples and expecting it to work on real-world data ... well that's not going to work. For example, you will have the following differences compared to the data your model was trained on.

  • Brush stroke width
  • Angle the digit drawn in
  • The way the letters written (for example if you look at 1 in the data set (at least the most) it's just a single stroke, where your image has two strokes)

So, if you want the model to perform better, use data augmentation

  • Introduce noise
  • Random rotation
  • Random scaling
  • Random brightness
  • Random cropping/scaling
  • etc.

Upvotes: 1

Related Questions