Link: https://www.kaggle.com/datasets/a2015003713/ (it was single +/- big dataset of aircrafts that I found)
- inspect some public Aircraft datasets, does them suitable for learning
- attemt to lear network on greyscale version of images (to use just 1 channel instead of 3)
- attemt to lear tiny neural network, to not consume a lot of resources
- attemt to use out-of-box neural network, that suitable for small devices (with low resources) - MobileNetV3Small
- Step 1: analyse of dataset . Conclusion: dataset is not really good
- Step 2: work with grayscale images and custom NN based on SeparableConv2D and GlobalAveragePooling2D Conclusion:
- Accuracy is bad - near 25%, and according to graphics model overfit a little.
- Looks like 128x128 is to small size to make good prediction, 256x256 is quite better
- Single channel make computation faster but in the same time accuracy lower. But in this case it make sense coz we do not want rely on colors
- Step 3: attempt to use MobileNetV3Small