Spot the Difference between two images using AI

Question

I'm looking for a method to Spot the difference between two images using AI.

It's my university project, that my professor asked me to create a program to detect and spot the differences in two pairs of images using AI.

I deployed it using the Siamese Network, to calculate the difference, and if the difference was greater than the threshold then, I used the following code to show the differences :

input_images = np.array([[img1, img2]])
difference_image = np.abs(input_images[0, 0] - input_images[0, 1])
plt.imshow(difference_image)

But my prof didn't accept it he gave me a hint to split images into smaller shapes using Conv2D and then compare those shapes and if there is a difference, highlight that using the bounding box.

Can anyone help to deploy this code?

My previous code is :

import numpy as np
import matplotlib.pyplot as plt
from tensorflow import keras
from tensorflow.keras import layers

img1 = plt.imread('1-1.jpg')
img2 = plt.imread('1-2.jpg')

input_shape = img1.shape  # Assuming images are of the same shape


# Function to create    
# def create_siamese_model(input_shape):
input_image_1 = layers.Input(shape=input_shape, name='input_image_1')
input_image_2 = layers.Input(shape=input_shape, name='input_image_2')

# Base network
base_network = keras.Sequential([
    layers.Conv2D(40, (3, 3), activation='relu', input_shape=input_shape),
    layers.MaxPooling2D(pool_size=(2, 2)),
    layers.Flatten(),
    layers.Dense(256, activation='relu')
])
# Encoded representations of input images
encoded_image_1 = base_network(input_image_1)
encoded_image_2 = base_network(input_image_2)

# L1 distance layer
l1_distance = layers.Lambda(lambda tensors: keras.backend.abs(tensors[0] - tensors[1]))([encoded_image_1, encoded_image_2])

# Output layer
output_layer = layers.Dense(15, activation='sigmoid')(l1_distance)

model = keras.Model(inputs=[input_image_1, input_image_2], outputs=output_layer)

input_images = np.array([[img1, img2]])
predictions = model.predict([input_images[:, 0], input_images[:, 1]])


threshold=0.5

if predictions[0, 0] > threshold:
    # Highlight differences if the prediction is above the threshold
    difference_image = np.abs(input_images[0, 0] - input_images[0, 1])
    difference_image
    plt.imshow(difference_image)
    plt.show()

Answer 1

I found a way that uses CNN network to find differences between two images code:

# Importing necessary libraries
import tensorflow as tf
import matplotlib.pyplot as plt

# Specify the file paths for the two images
image_path1 = '1.jpg'
image_path2 = '2    .jpg'

# Read and decode images, then normalize pixel values to the range [0, 1]
img1 = tf.io.read_file(image_path1)
img1 = tf.image.decode_image(img1, channels=1)
img1 = tf.cast(img1, tf.float32) / 255.0

img2 = tf.io.read_file(image_path2)
img2 = tf.image.decode_image(img2, channels=1)
img2 = tf.cast(img2, tf.float32) / 255.0

# Add a batch dimension to the images
img1 = tf.expand_dims(img1, axis=0)
img2 = tf.expand_dims(img2, axis=0)

# Create a Conv2D layer with specified parameters
conv2d_layer = tf.keras.layers.Conv2D(filters=1, kernel_size=(3, 3), activation='relu', padding='same')

# Apply the Conv2D layer to both images
output1 = conv2d_layer(img1)
output2 = conv2d_layer(img2)

# Calculate the absolute difference between the Conv2D outputs
diff = tf.abs(output1 - output2)

# Plotting the images and Conv2D outputs for visualization
plt.figure(figsize=(10, 5))

plt.subplot(1, 4, 1)
plt.imshow(tf.squeeze(img1), cmap='gray')
plt.title('Image 1')
plt.axis('off')

plt.subplot(1, 4, 2)
plt.imshow(tf.squeeze(img2), cmap='gray')
plt.title('Image 2')
plt.axis('off')

plt.subplot(1, 4, 3)
plt.imshow(tf.squeeze(output1), cmap='gray')
plt.title('Conv2D Image 1')
plt.axis('off')

plt.subplot(1, 4, 4)
plt.imshow(tf.squeeze(diff), cmap='gray')
plt.title('Absolute Difference')
plt.axis('off')

# Display the plot
plt.show()

this code use CNN network to Calculate the distance between two image array