VKCaptchaBreaker_v1.5
v1.5 Improvements:
- Recognition logic has migrated to the background script, causing faster recognition, lightweight content script and better extension experience in general;
- ONNX tensor's re-usage, so memory will be utilized efficiently, and V8's Garbage Collector will less interrupt recognition process;
- ONNX's trick, when we use float32 tensor from the first model as int32 tensor, simply changing its type, meaning no actual types casting is required;
- Fixed bug in Vkontakte's login form setup, when the extension has been pressing a login button in an infinite loop;
- The extension now works at backend script on Android devices (tested in Kiwi browser);
- Updated tensor preprocessing (before sending array to the backend script filter alpha-channel);
- Updated model weights (previous versions actually used 850k trained images (~82% accuracy). Now latest available model, trained on 1500k images with knowledge distillation is used (~91% accuracy).