patchdistances 1.0

Invariant patch distances and clustering

This package includes methods for clustering patches and reconstructing images using invariant patch distances. The package is build upon the CUDA C++ project: invariant_patch_distances. The package has the following dependencies: an NVIDIA graphics card of compute capability 3.x or higher, CUDA 10.0 and CMake 3.10.

Example

The following example shows how to cluster patches and reconstruct an image using these clusters.

import patchdistances as pd

# Load image as Float32 numpy array. It can either be gray-scale or rgb.
img = load_image()

// Initialize the distance class.
// Alternatively: similarity_inv_dist.
aid = pd.affine_inv_dist()

// Greedy-k-center clustering of the image patches using the distance.
labels, indices_of_labels = aid.greedy_k_center(
    image=img,
    patch_shape=(7,7),
    clusters=20,
    first_patch_idx=0,
    solver=pd.solver_t.procrustes,
    func_family=pd.func_t.superlevelsets,
    use_imed=False,
    interpolation=pd.inter_t.bicubic,
)

// Reconstruct the image using the previously clustered patches.
// Alternatively: reconstruct_w_translation.
reconstructed_image, nearest_neighbor_labeling = aid.reconstruct(
    image=img,
    labels=labels,
    rec_type=pd.rec_t.Median,
    solver=pd.solver_t.procrustes,
    func_family=pd.func_t.superlevelsets,
    use_imed=False,
    interpolation=pd.inter_t.bicubic,
)