/autoencoder_colorization

Colorizes grayscale images using a loaded model and displays original and predicted colorized versions.

Primary LanguagePythonGNU General Public License v3.0GPL-3.0

autoencoder_colorization

Colorizes grayscale images using a loaded model and displays original and predicted colorized versions.

This code block is for training a convolutional autoencoder model to colorize grayscale images. Let's break down the code and analyze it:

  1. Importing Libraries: The necessary libraries such as NumPy, Keras layers, optimizers, and callbacks are imported.

  2. Loading Data: Grayscale and corresponding color images are loaded from numpy files. The grayscale images are stored in x_gray and the color images in x_color.

  3. Data Preparation: The loaded data is split into training and testing sets.

  4. Model Definition:

    • The model architecture is defined using the Sequential API.
    • The encoder part consists of several Conv2D layers with increasing numbers of filters and varying strides. These layers are responsible for extracting features from the input grayscale images.
    • The decoder part mirrors the encoder architecture in reverse, aiming to reconstruct the colorized image from the extracted features. Dropout layers are included to prevent overfitting.
    • The final layer uses Conv2D with 3 filters (representing RGB channels) to produce the colorized output image.

  5. Model Compilation: The model is compiled using Mean Squared Error (MSE) as the loss function and Adam optimizer with a learning rate of 0.0001. The accuracy metric is also specified.

  6. Model Summary: The summary of the model architecture is printed, displaying the layers, output shapes, and number of parameters.

  7. Model Training:

    • The model is trained using the fit method on the training data (x_train_gray and x_train_color).
    • It is trained for 50 epochs with a batch size of 32.
    • Validation data (x_test_gray and x_test_color) is provided for evaluating the model's performance during training.
    • ModelCheckpoint callback is used to save the best model based on validation accuracy.

Overall, this code block defines, compiles, trains, and evaluates a convolutional autoencoder model for colorizing grayscale images using Keras. The model architecture consists of encoder and decoder parts, and it is trained to minimize the reconstruction error between the colorized images and the ground truth color images.

This code block defines a function process that preprocesses image data for a colorization task. Let's analyze it step by step:

  1. Importing Libraries: The necessary libraries such as load_img and img_to_array from Keras, NumPy, and glob are imported.

  2. Function Definition:

    • process function is defined, which takes two arguments: path (path to the directory containing the images) and target_size (desired size of the images after resizing).
    • Inside the function, empty lists x and y are created to store grayscale and corresponding color images, respectively.

  3. Processing Images:

    • The function iterates through each image file in the specified directory using glob.glob(path).
    • For each image file, it loads the color image (img_color) and grayscale image (img_gray) using the load_img function from Keras. Both images are resized to the specified target_size and converted to NumPy arrays using img_to_array.
    • The pixel values of both grayscale and color images are normalized to the range [0, 1] by dividing by 255.
    • The normalized images are appended to the x and y lists, respectively.

  4. Converting to NumPy Arrays:

    • After processing all images, the lists x and y are converted to NumPy arrays using np.array.

  5. Saving Processed Data:

    • Finally, the processed grayscale images (x) and corresponding color images (y) are saved as NumPy arrays using np.save.

  6. Function Call:

    • The process function is called with the specified path and target_size.

Overall, this function preprocesses color images by converting them to grayscale, resizing them, and normalizing their pixel values. It then saves the processed grayscale and color images as NumPy arrays, which can be used for training a colorization model.

This code defines a function named colorize that takes a list of image paths (test_set), the path to a trained model (model_path), and the target_size of the images. It then loads each grayscale image from the test set, passes it through the loaded model to predict the colorized version, and displays the original grayscale image along with the predicted colorized image using matplotlib.

Let's break down the code:

  1. Importing Libraries: The necessary libraries such as load_model, img_to_array, array_to_img, load_img, numpy, matplotlib.pyplot, and glob are imported.

  2. Function Definition:

    • The colorize function is defined with parameters test_set, model_path, and target_size.
    • Inside the function, the model is loaded using load_model from the specified model_path.

  3. Colorization Loop:

    • The function iterates through each image path in the test_set.
    • For each image, it loads the grayscale image using load_img, converts it to a NumPy array, expands its dimensions to match the input format expected by the model, and normalizes the pixel values.
    • The normalized grayscale image is passed through the model to predict the colorized version.
    • The predicted colorized image is converted back to an image format using array_to_img.

  4. Displaying Results:

    • For each grayscale image, the function plots two subplots using matplotlib.
    • The first subplot displays the original grayscale image, and the second subplot displays the predicted colorized image.
    • Finally, plt.show() is called to display the plot.

  5. Function Call:

    • The colorize function is called with the specified parameters: test set of images (obtained using glob.glob), target_size, and the path to the trained model.

This function can be used to visualize the colorized versions of grayscale images using the trained model. However, it seems like there might be an issue with the input images' paths as glob.glob("../custom/*.*") may not be pointing to the correct directory. You might need to adjust the path according to your directory structure.