Automatic CAPTCHA solver using YOLOv5. Created for the NOTS project at HAN.
Made by:
- Mike van Egmond
- Alex Cheng
- Steven Velderman
- Auke Onvlee
See the code working in this Google Colab Notebook.
Weights are located in the weights folder. The version number is the version of the dataset used. Best performance was seen using v37-yolov5m6-hyper-300. This model was trained for 300 epochs over a period of 6 hours and 45 minutes using an Nvidia Tesla P100 GPU.
Using Roboflow over 400 images were labelled by hand.
To increase the dataset size data augmentation was used. Using Roboflow we went from 400 images to 1000+ images, greatly increasing the accuracy of the model. Several augmentations were used.
The performance of the models was tracked using Weights and Biases. This allowed us to see how well the model was doing and compare it to earlier runs.
