- Solve captcha using TensorFlow.
- Learn CNN and TensorFlow by a practical project.
Follow the steps, run the code, and it works!
the accuracy of 4 digits version can be as high as 99.8%!
There are several more steps to put this prototype on production.
Ping me for paid technical supports.
-
Solve Captcha Using CNN Model
- Training: 4-digits Captcha
- Training: 4-letters Captcha
- Inference: load trained model and predict given images
-
Generate DataSet for Training
- Usage
- Example 1: 4 chars per captcha, use digits only
- Example 2: sampling random images
old code that using tensorflow 1.x is moved to tensorflow_v1.
this is a perfect project for beginers.
we will train a model of ~90% accuracy in 1 minute using one single GPU card (GTX 1080 or above).
if we increase the dataset by 10x, the accuracy increases to 98.8%. we can further increase the accuracy to 99.8% using 1M traning images.
here is the source code and running logs: captcha-solver-tf2-4digits-AlexNet-98.8.ipynb
Images, Ground Truth and Predicted Values:
there is 1 predicton error out of the 20 examples below. 9871 -> 9821
Accuracy and Loss History:
Model Structure:
- 3 convolutional layers, followed by 2x2 max pooling layer each.
- 1 flatten layer
- 2 dense layer
this is a more practical project.
the code is the same as the 4-digits version, but the training dataset is much bigger.
it costs 2-3 hours to generate training dataset and costs 30 min to train a 95% accuracy model.
here is the source code and running logs: captcha-solver-tf2-4letters-AlexNet.ipynb
example: captcha-solver-model-restore.ipynb
$ python datasets/gen_captcha.py -h
usage: gen_captcha.py [-h] [-n N] [-c C] [-t T] [-d] [-l] [-u] [--npi NPI] [--data_dir DATA_DIR]
optional arguments:
-h, --help show this help message and exit
-n N epoch number of character permutations.
-c C max count of images to generate. default unlimited
-t T ratio of test dataset.
-d, --digit use digits in dataset.
-l, --lower use lowercase in dataset.
-u, --upper use uppercase in dataset.
--npi NPI number of characters per image.
--data_dir DATA_DIR where data will be saved.
examples:
1 epoch has 10*9*8*7=5040
images, generate 6 epoches for training.
generating the dataset:
$ python datasets/gen_captcha.py -d --npi=4 -n 6
10 choices: 0123456789
generating 6 epoches of captchas in ./images/char-4-epoch-6/train
generating 1 epoches of captchas in ./images/char-4-epoch-6/test
write meta info in ./images/char-4-epoch-6/meta.json
preview the dataset:
$ python datasets/base.py images/char-4-epoch-6/
========== Meta Info ==========
num_per_image: 4
label_choices: 0123456789
height: 100
width: 120
n_epoch: 6
label_size: 10
==============================
train images: (30240, 100, 120), labels: (30240, 40)
test images: (5040, 100, 120), labels: (5040, 40)
scenario: use digits/upper cases, 4 chars per captcha image.
1 epoch will have 36*35*34*33=1.4M
images. the dataset is too big to debug.
using -c 10000
param, sampling 10k random images.
generating the dataset:
$ python3 datasets/gen_captcha.py -du --npi 4 -n 1 -c 10000
36 choices: 0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ
generating 1 epoches of captchas in ./images/char-4-epoch-1/train.
only 10000 records used in epoche 1. epoche_count: 1413720
tensorflow image: https://hub.docker.com/r/jackon/tensorflow-2.1-gpu
docker pull jackon/tensorflow-2.1-gpu
# check if gpu works in docker container
docker run --rm --gpus all -t jackon/tensorflow-2.1-gpu /usr/bin/nvidia-smi
# start jupyter server in docker container
docker run --rm --gpus all -p 8899:8899 -v $(realpath .):/tf/notebooks -t jackon/tensorflow-2.1-gpu