- Retraining of CNN with new analog counter type
6.1.0 Added new analog counter type (thanks to muerzi@iobroker.net) (2020-01-05)
- Retraining of CNN with new analog counter type
- introduction of new tlite-format --> will be preferred in future versions
- Retraining of CNN with slightly improved image classification
- Updated to Tensorflow 2.1
- additional export to TF-Lite Version (.tflite)
- Training with new picture from iobroker users
- Training with new picture from iobroker users
- Removal of standalone server - (included in main project)Training of additional digital number (provided from iobroker users)
- Training of the network with a second type of analog counters (different pointer)
- Image processing changed to Pillow (remove OpenCV)
- Usage of Tensorflow 2.0 for training
- Image processing changed to Pillow (remove OpenCV)
- Increased precision by handwise relabeling of the input
- Addtional training images from different illuminatin (ESP32-CAM)
- Adaption of neural network output to sinus and cosinus of pointer angle. Calculation of angle by arctan as unique function of full roation --> removal incontinuity at output neuron.
- Handling of periodic nature with differnt non continious strategies (switching of output neurons, introduction of periodic loss instead of mean square, ...)
- Improve learning by stepwise training strategie and adaption of neutral network structure / size neuron