hlin117/mahjongcv

The objective of this project is to be able to segment mahjong tiles from photos, and identify them. To find regions of interest that contain a mahjong tile, we use a series of bounding boxes, and to identify the tile in a regions of interests, we use a method involving homographies.

mahjongcv's Issues

Implement color thresholding

This should increase both performance and accuracy

Output should be an image of the input image, with labels on each tile

I think this would look better than the current output.

Removing shadows

Because of the way our algorithm is implemented, the shadows of the tiles in the pictures augment the size of the bounding boxes around the tiles. This will cause some tiles to be accidentally merged together.

Implement scaling of tiles

Pictures of larger tiles are not differentiated by pictures of smaller tiles. Hence, our segmentation algorithm only works for tiles of a fixed size in the picture.

Box distance algorithm

It's still buggy. Distances are not symmetric; we might have the case when dist(box1, box2) > dist(box2, box1)

That being said, the smaller distance is usually correct anyways, and since we go about calculating every distance combination anyways, it doesn't affect the output of the program.

Template matching

We haven't experimented how good template matching would be for our algorithm.

For more information about template matching, refer to this link:
http://docs.opencv.org/doc/tutorials/imgproc/histograms/template_matching/template_matching.html

Removing shadows

Because of the way our algorithm is implemented, the shadows of the tiles in the pictures augment the size of the bounding boxes around the tiles. This will cause some tiles to be accidentally merged together.

Implement PEP 8 style guidelines

Refer to the following link:
http://legacy.python.org/dev/peps/pep-0008/#introduction

Command line inputs

The program should receive a command line argument with the image whose tiles should be identified.

Use neural networks instead of KNN for classification

Neural networks have been the rave of computer vision recently. Michael Nielsen has a tutorial on how to use NNs to recognize handwritten digits.

I'm pretty sure we could come up with faster and more accurate results if we train a NN.

Parallelize Nearest Neighbor Algorithm