Image Recognition in Python using Machine Learning

By Shagun Bidawatka

Image Recognition is the process of providing a category to the image. We have to train our machine and build a model that can recognize the image for this purpose we use Keras and Tensorflow.

Image Recognition using Keras and TensorFlow

The objective of image recognition is to get the label or category of the input image by the neural network.

Creating the Basic Image Recognition Model:

• Importing Necessary Modules

import keras
import tensorflow as tf
import matplotlib.pyplot as plt
import random
import numpy as np
import pydot

• Loading Dataset

Loads the Fashion-MNIST dataset.

Dataset : 60,000 28×28 grayscale images

Categories: 10

Test set images: 10,000

Label Description :

0- T-shirt/top

1- Trouser

2 -Pullover

3 -Dress

4 -Coat

5 -Sandal

6- Shirt

7 -Sneaker

8 -Bag

9- Ankle boot

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

• Providing Labels to the values(0,1,2….)

class_names=['T-shirt/Top','Trouser','Pullover','Dress','Coat',
             'Sandal','Shirt','Sneaker','Bag','Ankle boot']

• Normalization of the data: Performing normalization to get the values in a confined range.

x_train_n=x_train/255.         
x_test_n=x_test/255.

• Splitting dataset into validation/train/test: We have taken 5000 rows for validation and the remaining for training.

x_valid,x_training=x_train_n[:5000],x_train_n[5000:]
y_valid,y_training=y_train[:5000],y_train[5000:]
x_test=x_test_n

np.random.seed(42)
tf.random.set_random_seed(42)

Output:

1.28*28 pixel (convert in 1-d)

2.Input layer

3.Hidden layer 1

4.Hidden layer 2

5.Output layer

6.10 categories

• Training Model: We are training the model using Keras and we are building a sequential model having a dense layer with 300 neurons and relu activation function and an output layer with 10 categories..

model=keras.models.Sequential()  #model object
model.add(keras.layers.Flatten(input_shape=[28,28]))  #input layer

#dense layer with 300 neurons and relu activation function
model.add(keras.layers.Dense(300,activation="relu"))  
model.add(keras.layers.Dense(100,activation="relu"))

#output layer with 10 categories 
model.add(keras.layers.Dense(10,activation="softmax"))
model.summary()

Output:

Model: "sequential_2" _________________________________________________________________
Layer (type) Output Shape Param # ================================================================= 
flatten_2 (Flatten) (None, 784) 0 _________________________________________________________________ 
dense_4 (Dense) (None, 300) 235500 _________________________________________________________________ 
dense_5 (Dense) (None, 100) 30100 _________________________________________________________________ 
dense_6 (Dense) (None, 10) 1010 ================================================================= 
Total params: 266,620 Trainable params: 266,610 Non-trainable params: 0 
________________________________________________________________

• Compiling the model

model.compile(loss="sparse_categorical_crossentropy", optimizer="sgd", metrics=["accuracy"]) 
#sochastic gradient design 
model_history=model.fit(x_training,y_training,epochs=30, validation_data=(x_valid,y_valid))

Output:

Train on 55000 samples, validate on 5000 samples
Epoch 1/30 55000/55000 [==============================] - 8s 145us/step - loss: 0.7326 - accuracy: 0.7609 - val_loss: 0.4999 - val_accuracy: 0.8366 
Epoch 2/30 55000/55000 [==============================] - 6s 109us/step - loss: 0.4890 - accuracy: 0.8294 - val_loss: 0.4330 - val_accuracy: 0.8526 
Epoch 3/30 55000/55000 [==============================] - 7s 128us/step - loss: 0.4419 - accuracy: 0.8457 - val_loss: 0.4077 - val_accuracy: 0.8602
Epoch 4/30 55000/55000 [==============================] - 7s 136us/step - loss: 0.4158 - accuracy: 0.8534 - val_loss: 0.4049 - val_accuracy: 0.8612 
Epoch 5/30 55000/55000 [==============================] - 8s 145us/step - loss: 0.3949 - accuracy: 0.8621 - val_loss: 0.3932 - val_accuracy: 0.8646 
Epoch 6/30 55000/55000 [==============================] - 11s 192us/step - loss: 0.3802 - accuracy: 0.8658 - val_loss: 0.3882 - val_accuracy: 0.8670 
Epoch 7/30 55000/55000 [==============================] - 13s 233us/step - loss: 0.3664 - accuracy: 0.8695 - val_loss: 0.3616 - val_accuracy: 0.8726 
Epoch 8/30 55000/55000 [==============================] - 11s 206us/step - loss: 0.3550 - accuracy: 0.8742 - val_loss: 0.3754 - val_accuracy: 0.8622 
Epoch 9/30 55000/55000 [==============================] - 11s 197us/step - loss: 0.3452 - accuracy: 0.8776 - val_loss: 0.3569 - val_accuracy: 0.8770 
Epoch 10/30 55000/55000 [==============================] - 13s 244us/step - loss: 0.3364 - accuracy: 0.8804 - val_loss: 0.3498 - val_accuracy: 0.8740 
Epoch 11/30 55000/55000 [==============================] - 9s 162us/step - loss: 0.3275 - accuracy: 0.8826 - val_loss: 0.3582 - val_accuracy: 0.8746 
Epoch 12/30 55000/55000 [==============================] - 11s 195us/step - loss: 0.3198 - accuracy: 0.8867 - val_loss: 0.3473 - val_accuracy: 0.8756 
Epoch 13/30 27104/55000 [=============>................] - ETA: 4s - loss: 0.3112 - accuracy: 0.8878

model_history.params

Output:

{'batch_size': 32, 
'epochs': 30, 
'steps': None, 
'samples': 55000, 
'verbose': 1, 
'do_validation': True, 
'metrics': ['loss', 'accuracy', 'val_loss', 'val_accuracy']}

• Evaluation: Evaluating the accuracy of the model.

model.evaluate(x_test,y_test)

Output:

10000/10000 [==============================] - 1s 94us/step

[0.3345632088780403, 0.878600001335144]

• Testing

y_proba=model.predict(x_new)
y_proba.round(2)

Output:

array([[0. , 0. , 0. , 0. , 0. , 0. , 0. , 0.01, 0. , 0.99], 
[0. , 0. , 0.97, 0. , 0.03, 0. , 0. , 0. , 0. , 0. ],
 [0. , 1. , 0. , 0. , 0. , 0. , 0. , 0. , 0. , 0. ]], dtype=float32)

y_pred=model.predict_classes(x_new)
y_pred
np.array(class_names)[y_pred]

Output:

array([9, 2, 1], dtype=int64)

array(['Ankle boot', 'Pullover', 'Trouser'], dtype='<U11')

To check the testing image print image.

print(plt.imshow(x_new[0]))

We are done with our basic training and testing part of the model, you can add GUI to this. We can change the number of units in the hidden layer, the optimizer, the number of epochs of training, the size of batches, and analyze the change in the accuracy of the model.