Skip to content

Latest commit

 

History

History
44 lines (24 loc) · 2 KB

CNN_Version2.md

File metadata and controls

44 lines (24 loc) · 2 KB
Raw

CNN with periodic loss functin - Training the CNN neural network

Details on setup and training of the CNN neural network

Network Structure

The network consists of several Conv2D, MaxPooling and Flatten Layers:

Training Data

The training data consists of images of the analog pointer with the size of 32x32 pixels and RGB color code. The expected readout value is encoded in the filename itself. Details can be found Training_Data.md

For reshaping the images the following code can be used: [Image_Preparation.ipynb]

Training strategy

The details can be found in the Jupyter file itself: Train_CNN_Analog-Readout_Version2.ipynb

The following aspects are implemented:

Scattering input images

Scattering the input images by brightness as well as a pixel shift for training variations improved the stability of the network a lot.

Brightness was scattered with +/-30%

The position was scattered with +/-1 pixel in each direction. As the original picture is more than 4 times bigger (142x142 pixel) this ensures enough uncertainty to the upstream image alignment and cut out procedure.

Results

The results show the expected behaviour:

sinus / cosinus Orginal / Prediction

Deviation from expected value

Checking the deviation by a comparison to the label shows, that the error is very small. The statistic shows an average value of 0.0%+/-1.1% with a maximum deviation of 3.8%. With respect to the pointer, that means that the error is in the subdigit range in the maximum +/-0.38 over the full range. Part of this error is also caused by a uncetainty in the manual labeling. Taking this into account, the result is fully sufficient to be robust for continious readout of analog meters.