Step-by-Step CNN Architecture: How Images Flow Through a Neural Network

Here we implement a Convolutional Neural Network illustrating how each layer processes and transforms the input image.

Step 1: Import Required Libraries

Here we import TensorFlow for CNN operations and Matplotlib for visualization.

python

import tensorflow as tf
import matplotlib.pyplot as plt

plt.rc('image', cmap='gray')
plt.rc('figure', autolayout=True)

Step 2: Load and Preprocess the Image

Load the image convert it to grayscale, resize it to 300×300 and normalize pixel values.

python

image_path = "Image Path"

image = tf.io.read_file(image_path)
image = tf.io.decode_jpeg(image, channels=1)  
image = tf.image.resize(image, [300, 300])
image = tf.image.convert_image_dtype(image, tf.float32)

print("Original Image Shape:", image.shape)

plt.figure(figsize=(5,5))
plt.imshow(tf.squeeze(image))
plt.title("Original Image")
plt.axis('off')
plt.show()

# Add batch dimension
image = tf.expand_dims(image, axis=0)

Output:

Step 3: Define Convolution Kernel

We define an edge detection filter (Laplacian kernel) to extract important image features.

python

kernel = tf.constant([
    [-1, -1, -1],
    [-1,  8, -1],
    [-1, -1, -1]
], dtype=tf.float32)

kernel = tf.reshape(kernel, [3, 3, 1, 1])

Step 4: Apply Convolution Layer

The convolution layer applies the filter to the image to detect edges and features.

python

conv_output = tf.nn.conv2d(
    input=image,
    filters=kernel,
    strides=[1, 1, 1, 1],
    padding='SAME'
)

print("After Convolution Shape:", conv_output.shape)

plt.figure(figsize=(5,5))
plt.imshow(tf.squeeze(conv_output))
plt.title("After Convolution")
plt.axis('off')
plt.show()

Output:

Step 5: Apply ReLU Activation Function

ReLU removes negative values and introduces non-linearity into the network.

python

relu_output = tf.nn.relu(conv_output)

print("After ReLU Shape:", relu_output.shape)

plt.figure(figsize=(5,5))
plt.imshow(tf.squeeze(relu_output))
plt.title("After ReLU Activation")
plt.axis('off')
plt.show()

Output:

Figure 6. After apply ReLU activation function for sample image

Step 6: Apply Max Pooling Layer

Max pooling reduces spatial dimensions while keeping important features.

Output:

Step 7: Apply Flatten Layer

The flatten layer converts 2D feature maps into a 1D feature vector for fully connected layers

python

flatten_layer = tf.keras.layers.Flatten()
flatten_output = flatten_layer(pool_output)

print("After Flatten Shape:", flatten_output.shape)

print("First 20 Flattened Values:")
print(flatten_output.numpy()[0][:20])

Output:

javascript

After Flatten Shape: (1, 22500)

First 20 values of Flattened Vector:

[135.  81.  81.  81.  81.  81.  81.  81.  81.  81.  81.  81.  81.  81.

  81.  81.  81.  81.  81.  81.]

Step 8: Add Fully Connected (Dense) Layer

The fully connected layer learns high-level patterns from the flattened feature vector and produces output predictions.

python

dense_layer = tf.keras.layers.Dense(
    units=64,         
    activation='relu' 
)

dense_output = dense_layer(flatten_output)

print("After Fully Connected Layer Shape:", dense_output.shape)

Output:

javascript

After Fully Connected Layer Shape: (1, 64)

Step-by-Step CNN Architecture: How Images Flow Through a Neural Network

Step 1: Import Required Libraries

Step 2: Load and Preprocess the Image

Step 3: Define Convolution Kernel

Step 4: Apply Convolution Layer

Step 5: Apply ReLU Activation Function

Step 6: Apply Max Pooling Layer

Step 7: Apply Flatten Layer

Step 8: Add Fully Connected (Dense) Layer

Nguyen Bao Huy

Discussion

No comments yet

Leave a Reply