ColourPulse

The code performs automatic colorization of black-and-white images using a pre-trained deep learning model. It loads the model and cluster centers, processes the input image, predicts color channels, combines them with the luminance channel, and converts the result to a colorized image.

Execution Guide:

1. Run the following command line in the terminal:

   pip install opencv-python opencv-contrib-python numpy
   

2. Enter the path of the black and white images in the code, also enter the output directory.

3. Download the following models and paste their path in the code:

   a. colorization_deploy_v2.prototxt

   b. colorization_release_v2.caffemodel

   c. pts_in_hull.npy

4. Run the code and it will output a colour image

Model Prediction:

Image Input:

Image Output:

Overview:

This script implements automatic colorization of black-and-white images using a pre-trained deep learning model in OpenCV. Below is a step-by-step breakdown:

1. Library Imports

• The script uses essential libraries:
  • numpy for numerical computations.
  • cv2 (OpenCV) for image processing.
  • os for file path management.

2. Model Paths

• Specifies the paths to the required model files:
  • prototxt: Defines the network architecture.
  • caffemodel: Contains the pre-trained weights.
  • npy: Stores cluster centers for color distribution.
• Ensures paths are dynamically adjusted and checks the existence of critical files.

3. Model Initialization

• Loads the pre-trained model (.prototxt and .caffemodel) using OpenCV's cv2.dnn.readNetFromCaffe.
• Loads the color cluster centers from the .npy file.
• Modifies the network by adding cluster centers as 1x1 convolutions.

4. Colorization Function (colorize_image)

• Input: Path to a black-and-white image.
• Steps:
  1. Reads the input image and verifies its existence.
  2. Converts the image to the LAB color space, where:
     • L: Lightness (input channel for colorization).
     • a and b: Color channels (predicted by the model).
  3. Preprocesses the image:
     • Normalizes pixel values.
     • Resizes the image to match the model input dimensions.
     • Extracts and adjusts the L channel.
  4. Feeds the processed L channel to the network.
  5. Predicts the a and b channels and resizes them to the original image dimensions.
  6. Combines the original L channel with the predicted a and b channels.
  7. Converts the LAB image back to the BGR color space for display.
  8. Clamps pixel values to the valid range and converts the image to 8-bit format.

Key Features:

• Utilizes OpenCV's DNN module to load and process pre-trained deep learning models.
• Automatically converts and colorizes black-and-white images using the LAB color space.
• Provides robust error handling for missing files or invalid inputs.