Chess Board Computer Vision

This is a topic that I keep coming back to. This time for Chess. The goal is to extract chess notation for a game played in view of a webcam. This time I made it harder for myself by viewing the board from an angle. I wrote it in Groovy using platform-independent JAI (Java Advanced Imaging). I've cheated on the webcam capture, and used Photo Booth stills to begin with. This time, though, I'm sharing the code on GitHub.

For each frame I do the following:
1. Convert to grayscale using "bandcombine":

2. Correct for perspective using PerspectiveTransform.getQuadToQuad and "warp": (I noted the pixel coordinates for the corners of the board by hand)

3. Binarize upper and lower bands of the image histogram and OR them together (so that a white pixel ideally identifies the presence of a piece on the board) -- using "histogram", "binarize", OrDescriptor, and SubtractDescriptor:


Then I take each processed frame and its following processed frame, and do the following:
1. XOR the two frames using XorDescriptor:

2. Erode the result using "erode":

3. Open it using "dilate" and then "erode":


The result would ideally show clusters of pixels in the board squares that have changed (e2 to e4 in this example). Unfortunately, a naive solution here isn't enough as pieces span multiple board locations. Also the piece detection algorithm in step 3 of frame processing is too weak.

Feel free to ask questions if you're interested.