It's a system that allows your educational organization to check offline tests faster with an ability of sharing students' mistakes faster and in more elegant way.

TuckleJudge intro movie

Initially you have to register student account with _admin name.

Now you have a capability of creating teacher account. Verification code for it is written on the main page of _admin account. It's being changed every 2 days. student account can be created without this code.

As you created a teacher account you get the option of creating/editing/checking tests and viewing checked tests' summary. students can only view their own tests.

Initially I thought about the idea of multiplying greyscale pixels' values on a large numbers like 250. As you can see on the image below, picture after such operation can be split on multiple near lying components as though on puzzles from which the picture is combined, and size of each component depends on the number of various details on the fragment. It's clear that the fields of the test list are full of details and that's how we can distinguish it from the other parts of the image.

But it used to work only on a specific type of images, only with specific noise or something. I've spent too much time on the process of exploration, but it hadn't paid off, so I put it off.

Initially program receives a photo of the test and converts it to greyscale. Then threshold applies and various components are detected using plain bfs. After that algorithm detects components which areas shape is similar to a square. That's how it detects these 3 black squares. Then it's applying various transformation depending on these 3 squares positions: rotation, shear, flipping, perspective transformation.

It's not ideal yet, I'd say it's pretty far from being called that way. It's disgusting filthy stinky muddy bastard shit creation. It's just full of horrorshow arrogance, nothing nobel in it. Just a shitty thing. Can't stand it. Ok, ok... calm down. Yes, I've done it on my oddy knocky. Bold queer poo - just having no healthy thoughts about it. Eh, that's right so what is it about. And why am I harshly bewaring it.

So, initially I was told that this problem could be solved easily by single PERCEPTRON layer. That's true, single perceptron layer can be trained on MNIST database having pretty nice results - 86% detected photos. But using the same model on my own dataset (created by fields' detector) is just an absurd, it had literally zero accuracy. So I had to experiment with a lot of image alterations to make MNIST images look mostly identical to mine. Then I used gathered all my ML knowledge and created perceptron models for all the digits. But it crushed, just refused to work properly. 60% accuracy, by God, it's enormously tiny number.

BTW, the algorithm was quite interesting and amusing, pretty technological at some point, so if you're interested in the way I've implemented it, check out my Telegram channel.

Then I brought CNN from this repository, wrote all the functions for image transformation and for saving layers to file. And what??? It's just unfair. This CNN can obtain only one convolutional layer because it just crushed if I used some different model from what the author presented. And I found out that the bug was in the divisions by zero, cringe exponential functions, unsafe operations with numbers. So the whole thing was a mess. It's my subjective thought so don't blame me if it's not fully accurate, but it won't change my mind(, coz the CNN presented in that repository was something weird. So I wanted to create mine. But at the time summer had approached the finish line, and I had a lot of other things to do, so I called it a day at that point.