Latest version: 1.0.0 BETA
CaptchAmI - Stars and Numbers Captcha Solver
What is CaptchAmI
I created this project to show how it is possible to use AI + Computer Vision to solve Captcha. In particular, I had to deal with two captcha types: stars and numbers. For the captcha containing the stars, I had to count them; for the one containing numbers, I had to do a mathematical operation (addition or subtraction).
The expected output of both the elaboration is a result number, which is: the total of stars printed on the image, or the result of the operation described.
Computer Vision elaboration
Stars elaboration
To recognize how many stars are present, the image is initially preprocessed: it is converted to gray, then colors are inverted, finally a threshold is found. Everything below this threshold is deleted (usually the background) while the rest of the image is kept intact. After that, I find how many stars are in the image by calculating the number of objects.
Number elaboration
When there are numbers on the image, the process is different because I have to perform a calculation (addition or subtraction). For this reason, I decided to split the images into three parts (first operand, operator, second operand) and process them accordingly using a neural network. The image elaboration starts in the same way as the previous step by removing the background and cleaning the image of any possible disturbance. After that, I look for three different regions of pixels (ideally it should be two numbers and one operator). Occasionally the operator is connected to the same region of one of the numbers, so if it happens I have decided to "manually" split the biggest area to 5px starting from the rightmost limit. This ensures to get a "usable" operator without cutting too much of the operand.
The Neural Network Part
I created two different neural networks to classify the two different types of input. As I wanted the program to be able to separate at the same time both the stars from the numbers and elaborate those to get the operation result, there are two neural networks:
-
A NN separates the stars from the numbers
-
Another NN recognizes the numbers and the operators
All the NNs expect to work on 136x47 8-bit colour png images. They have the same structure with two convolutional layers and three linear layers. The only difference between the nets is the number of linear units for the classification of the stars and the numbers/operator. The details of the NN are in the neural_net.py file.
Dataset
You can find the dataset I used in /dataset.
Please refer to the README.md in that folder for further information about the structure of the dataset.
WebService
This version of CaptchAmI is based on a microservice that is used for both the classification and the training. There are a bunch of endpoints that can be reached by launching the server:
/classify [POST]
It accepts a JSON with the image encoded with a field called "base64_img", in which there is the image encoded with BASE64. It will return the number of stars or the result of the operation.
/retrain/binary [GET]
Once called, the neural network that is used to discriminate the image with numbers from the one with the stars is re-trained. It will return the accuracy of the network on the test set.
/retrain/numbers [GET]
Once called, the neural network specialized to recognize the numbers and the operational sign is re-trained. It will return the accuracy of the network on the test set.
Configuration
To be able to correctly use the CaptchAmI server, you have to modify (if needed) the config.yaml file in the root folder of this project.
This file contains all the paths used in the project and eventually needs to be replaced according to your needs.
Please refer to the description of the YAML directly in the file.
Conclusions
I just put down some notes.
-
The program is still under development and will be tested in the future
-
I will create soon a docker image as well
-
The accuracy on the test set is around 1. It works pretty well :)