jubiiz / suguru Goto Github PK

From a single case image, extract which sides are bold and which are thin.

The code was not built to support non-square problems. Some parts (like in the processor.py automatic grid selection process) assume that the number of rows = the number of columns.

The code has a number of user-unfriendly features that would benefit to be altered.
These features need to be corrected before issue #13 can be tackled.
Here is a brief list of what they are:

• multiple changes of filename required per source-file
• inconsistent filename input (sometimes it's "filename.txt", others just "filename")
• unclear and inconsistent data size input (a few files still rely on the fact that the problem will be 20x20, and do not have a centralized way of changing those numbers)

Manual contour selection to find the ROI (the four corners of the grid) takes time and can be inefficient if done hastily. Consider adding an automatic feature so the program can do it for you (like with cv2.contour).

At line 72 of pentomino.py, the slice to extract the individual case image is difficult to read.
The function of each term is not easily understood.
The issue could be resolved by relaying the cropping (or just slice production) task to a function.

Using the technology as used in CS50 AI to extract what number lies in each cell.
The cell images might have to be cropped again to remove any sidelines: they may interfere with the digit recognition.

The CNN used to extract numbers from the grid is not optimal: it cannot recognize empty cases:
Here's how it works:

1. reads the number in the case
2. if its a 1, check if the average of all the pixel is very low
3. if so, it's not a 1, it's a blank cell

A CLI could be used to navigate the different source-codes during the suguru resolution process (not only solving: processing, extraction, correction, solving)

more resiliency of the neural network is needed. either different training data or different neural net
diff training data seems best rn imo