Solving "Where's Waldo" with deep learning in the form of feedforward convolutional neural networks + clever preprocessing to increase our data set
Previous attempts at automatically solving Where's Waldo puzzles have failed to adhere to the standard rules of the puzzle and have not yet solved anything. The previously best known solution uses a naive pattern-matching procedure to find waldo in an image given a cropped picture of his exact pose in that image. Implementations like this are trivial and equate to a simple linear search.
The way a human is expected to solve a Where's Waldo puzzle is different; humans must use their memory of what waldo looks like, do mental transformations to account for differences in pose, and recognize him freely by scanning the image.
We implement a feedforward convolutional neural network to solve this. In optimization we use: ** Backpropogation** Propogating the layers back to all layers. ** Convolution** Nonlinear transformation on the error. ** Hidden layers** Latent ** Hinton's Dropout** Randomized dropout for feedforward, back in when feedbackward. ** Alpha Optimization** iterate to find best learning parameter. ** Gradient Descent** Optimization In preprocessing we use: Image Reflection We perform a matrix operation to transform the image along the y axis. This doubles our data set and accounts for different directions Waldo may be facing. Pose shifting We perform manual shifts to the head to simulate human pose shifting.