How to Build and Crack Image Captchas with Python and Tesserocr

Writing crawlers inevitably encounters captchas, which can be roughly divided into four types:

Image

Slide

Click

Voice

We first examine image captchas, usually composed of digits and letters (sometimes Chinese characters) with added noise, interference lines, distortion, overlapping, and varied colors to increase difficulty.

The recognition process generally includes the following steps:

Grayscale conversion

Contrast enhancement (optional)

Binarization

Denoising

Skew correction and character segmentation

Training set creation

Recognition

In this experimental setup, captchas are generated programmatically using the Claptcha library (the Captcha library is also a good alternative), allowing large labeled datasets.

To generate a simple numeric captcha without interference, modify claptcha.py line 285 ( _drawLine) to return None, then generate captchas:

The generated captcha shows slight deformation. For such simple captchas, the open‑source Tesserocr OCR engine can be used directly.

First install Tesserocr:

Then start recognition:

Even with this simple captcha, the recognition rate is already high without additional processing.

Next, we add noise to the captcha background and observe the effect:

Recognition after adding noise:

We then generate an alphanumeric captcha (letters and digits):

Resulting image shows characters that are visually confusing (e.g., lowercase “o”, uppercase “O”, digit “0”, etc.).

Adding an interference line (by restoring the original _drawLine) makes the captcha unreadable:

To remove interference lines, we first convert the image to grayscale (otherwise we get RGB tuples):

After grayscale conversion, the image is sharper:

We then apply a 4‑neighbour or 8‑neighbour algorithm to remove thin lines. Pixels whose surrounding neighbours have a high count of white (255) are considered noise and removed:

The processed image appears sharper, but the removal is ineffective when the interference line width matches the character strokes.

For captchas where noise pixels differ in color from the characters (e.g., those generated by the Captcha library), denoising works better. Multiple denoising passes can further improve results:

After the final denoising, recognition yields:

This first article records how to perform grayscale conversion, binarization, and denoising on captcha images and use Tesserocr for simple captcha recognition; further techniques will be covered in the next article.