It is really hard to separate this image into four different letters.
Even if we want to try separating all test images into five different letters, I think we'd better discarding those harder to separate into different letters among the training images

Since the CAPTCHAs are not too complicated, maybe we can achieve a good recognition rate without even training the parameters.

When run project this show error like:

Traceback (most recent call last):
File "main.py", line 6, in
import dataset_manager
File "/media/io/HocTap/Lapsoft - Deeep Learning/bilibili-captcha/dataset_manager.py", line 14, in
from captcha_recognizer import CaptchaRecognizer
File "/media/io/HocTap/Lapsoft - Deeep Learning/bilibili-captcha/captcha_recognizer.py", line 124
nonlocal seq
^
SyntaxError: invalid syntax

Can you help me fix it?

Thanks for that.

What that means?
Does the tool only handle image from bilibili or other captchar?

@kenathe You may want to add yourself into AUTHORS and LICENSE.

• Introduction
• Usage
• Outcome
• Suggestion of further study & development

The current procedure of remove_noise_with_hsv is as follows.

1. Find the 2nd most frequent color (the standard color), (std_h, std_s, std_v)
2. For each pixel in the original CAPTCHA image, calculate its deviation from the standard color, (delta_h, delta_s, delta_v)
3. If delta_h <= h_tol && delta_s <= s_tol && delta_v <= v_tol, set the new grayscale value to be 1 - delta_v, or 0 otherwise.

Problem

This method may produce very different thicknesses for some CAPTCHA images.

Case A

Case B

The original chars appear to have almost identical thicknesses; after this procedure, the char in Case A is ~2x as thick as in Case B.

If this is fixed...

• Better performance of char partitioning (less wrong partitions)
• More hint on char partitioning
• Better performance of the MLP