Dataset of the "Unwrapping Non-Locality in the Image Transmission Through Turbid Media"

This post provides the dataset associated with the study titled "Unwrapping Non-Locality in Image Transmission Through Turbid Media." The "X_random.npy" file contains speckle image data with dimensions (150,000, 128, 128). The first axis spans the 150,000 images used in this study. The first 120,000 images are speckles corresponding to randomly generated images. The next 10,000 images (from 120,000 to 130,000) correspond to the first 10,000 images of the CelebA dataset. The remaining 20,000 (from 130,000 to 150,000) images consist of the first 10,000 images of the CIFAR and MNIST datasets, respectively. The ground truth images (SLM commanded images) are provided in "gt_random.npy", "gt_cifar.npy", and "gt_mnist.npy" for randomly generated images, CIFAR, and MNIST datasets, respectively. Each image in these datasets is scaled to a resolution of 128 by 128 pixels. The faces dataset are the 128 by 128 resized images of the CelebA dataset. A code example of the presented neural network model and its architecture implementation is given in "GAM on ImageNet data.py". This code reads the training dataset (speckle images) from the Ref [1] study ("x_train.npy") and rescales them to 128 by 128 pixels. The rescaled data, along with their corresponding resized SLM commanded images ("y_train.npy"), are used for training the GAM model (the model presented in the manuscript). This code then saves three files: "Trained GAM.h5," "samp_final_16_0.npy," and "samp_final_16_1.npy," which contain the trained model weights and biases and the sampling index of the model nodes, as described in the manuscript and commented on in the "GAM on ImageNet data.py" code. Code for testing this model is also provided in "test GAM.py." This file reads the testing dataset from the [1] study, specifically "x_test_cat.npy," "x_test_horse.npy," "x_test_punch.npy," and "x_test_parrot.npy" speckle images. It applies the trained model ("Trained GAM.h5," "samp_final_16_0.npy," and "samp_final_16_1.npy") to them. The code then calculates the structural similarity index (SSIM) to the ground truth files, namely "y_test_cat.npy," "y_test_horse.npy," "y_test_punch.npy," and "y_test_parrot.npy" for the "x_test_cat.npy," "x_test_horse.npy," "x_test_punch.npy," and "x_test_parrot.npy" data, respectively. Reference [1] Caramazza, P., Moran, O., Murray-Smith, R., & Faccio, D. (2019). Transmission of natural scene images through a multimode fibre. Nature Communications, 10(1), 2029.