Solving jigsaw puzzles with deep CNN's .

A jigsaw puzzle is a tiling puzzle that requires the assembly of often oddly shaped interlocking and tessellating pieces. Each piece usually has a small part of a picture on it; when complete, a jigsaw puzzle produces a complete picture.

Problem statement: Given an source image of size NxN and shuffled image with block size of BxB, our aim is to unscramble the shuffled jigsaw image to produce the original image.Here we limit ourselves to square shaped images where the tessellating pieces have equal height and width.The blocks should be square shaped and can have four possible orientations i.e rotations in degrees - {90,180,270,360}.

A direct permutation based approach to solve the puzzle will require an algorithm with exponential time complexity.Here we use a simple encoder-decoder based CNN with two heads for block shuffling and block rotations respectively, for solving the jigsaw puzzle in real-time.

You may run the entire training, testing and data preparartion in Google Colaboratory with the IPython notebook.

N.B: Install the correct version of keras to train the network. Also ensure sufficient ram is available, since we will be loading our full data-set into memory(currently) at the start of training.

• Python3
• PIL, matplotlib
• Scipy, skimage
• Tensorflow
• Keras 2.2.2

• Keras installation (TF backend)
• GPU with CUDA support

pip install keras==2.2.2

python data.py # Prepare the dataset
python train.py # Train the network
python infer.py # Perform inference

Version 1.0

Anil Sathyan

• "https://keras.io/"
• "https://www.tensorflow.org/"
• "https://stackoverflow.com/"