Low-Light Image Enhancement and Deep Learning with Python

Set up your working directory in Google Drive for the image enhancement project, download the image enhancement folder, and include training set, validation set, evaluation set folders with the code.ipynb.

Explore the load data set of 500 image pairs for low-light photography enhancement, with training, evaluation, and validation subsets containing high and low subfolders, all sized 400 by 600.

Open the code.ipynb in the image enhancement folder to develop, train, and evaluate image enhancement models using the full dataset.

Open the code.ipynb file in Google Colab by right-clicking and selecting open with Google Colab, then run the notebook to execute the code, leveraging Google Drive for access to resources.

Enable the gpu in Google Colab to accelerate computation by selecting a hardware accelerator, speeding up deep learning tasks such as training large models and processing complex data.

Mount your Google Drive in a Google Colab notebook by importing the drive module and granting permission, making Drive contents accessible at a mounted path for file access.

Import essential libraries for image processing and deep learning, including os, glob, PIL, OpenCV, NumPy, Matplotlib, TensorFlow, Keras, and the layers module.

Set the random seed to ten to ensure reproducible results across runs, then set the image size to 128 and the batch size to four to standardize processing and training.

Read and preprocess an image with a helper that uses tf.io.read_file and tf.image.decode_png, enforces RGB, casts to float32, and normalizes to 0–1 for TensorFlow models.

Implement a random crop function that crops both low light and enhanced images with shared coordinates, enabling data augmentation for training with tf.shape-derived height and width.

Define a load_data function that reads and preprocesses images from low_light_image_path and enhanced_image_path, applies random crops for augmentation, and returns a tuple of preprocessed images for training or evaluation.

Construct a TensorFlow dataset pipeline that pairs low light and enhanced image paths, uses from_tensor_slices and map to load and preprocess images, batches with drop_remainder, ready for training or evaluation.

Define file paths for training, validation, and test data using the globe function, creating sorted lists of train lowlight images and train enhanced images.

initialize train and validation datasets for lowlight image enhancement using TensorFlow tf.data.dataset API, loading and pre-processing lowlight and enhanced images, and printing dataset metadata.

Selectively fuse multiscale features using a kernel-based approach, compute channel statistics via global average pooling, reduce channels with 1x1 convolution and ReLU, and generate per-scale descriptors via softmax to aggregate.

Learn a dynamically learned spatial attention mechanism that concatenates average and max pooled tensors, uses a 1x1 convolution and sigmoid activation to generate an attention map that modulates input tensor.

Learn to build a channel-wise attention block using global average pooling and 1x1 convolutions with ReLU and sigmoid to create per-channel weights, then multiply the input by these weights.

Presents a dual attention using 3x3 convolutions with ReLU and a 3x3 conv without activation, applying channel-wise and spatial-wise attention, concatenating features along channels, 1x1 conv adjustment, and residual connection.

Explain a down sampling module with skip connections that preserves information while extracting features via a two-branch design: main conv and max pooling, plus a skip path, combined by addition.

Explore an upsampling module with skip connections that increases feature map spatial dimensions through convolutional and upsampling branches, preserving earlier information to reconstruct finer details.

Extract multi-scale features by downsampling, applying dual attention units and selective kernel feature fusion, then fuse with a residual connection to enhance low-light images.

Builds a recursive residual group by stacking multiscale residual blocks with three-by-three convolutions and a residual connection to refine features across scales and preserve original information.

Define the multiple iterative residual network architecture for image enhancement, using recursive residual groups and multiscale residual blocks with residual connections to produce high-quality enhanced images.

Implement charbonnier loss and peak signal to noise ratio as a custom loss and evaluation metric for deep learning image restoration tasks, using TensorFlow or Keras.

Compile a deep learning image enhancement model in TensorFlow or Keras, using Adam optimizer with a learning rate of 1e-4, Charbonier loss, and PSNR as the monitoring metric during training.

Train a deep learning model with TensorFlow or Keras on train and validation datasets for 50 epochs, using a callback that reduces learning rate when val PSNR stops improving.

Save the trained model with model.save to serialize architecture, weights, and optimizer state into a single .h5 file named Enhancement Model.h5, enabling inference, retraining, or sharing on Google Drive.

Plot training and validation loss over epochs with matplotlib on the same graph to visualize learning curves, compare train and val loss, and assess convergence, overfitting, or underfitting.

Plot the training and validation psnr over epochs with matplotlib to compare how psnr evolves across approaches, using a legend to distinguish curves and assess reconstructed image quality.

Visualize multiple images with titles in a plot using a function that creates subplots, displays images with imshow, hides axes, and shows figure to compare original, enhanced, and PIL-adjusted images.

Apply a pre-trained model to enhance images, preprocess inputs to numpy, normalize to 0-1, add a batch dimension, run the model, and convert results back to a PIL image.

Visual inspection of six low-light images enhanced by the MirNet model, using infer and PIL auto-contrast, displayed side by side with original and auto-contrast results in a grid.