Analysis of Military Aircraft Detection Dataset

Link: https://www.kaggle.com/datasets/a2015003713/ (it was single +/- big dataset of aircrafts that I found)

The goal was:

• 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

Implementations steps (Research):

1. Step 1: analyse of dataset . Conclusion: dataset is not really good
2. 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
3. Step 3: attempt to use MobileNetV3Small

Plans TODO:

1. Improve accuracy
2. Attempt to clean up data set
3. Try to use trained models on video
4. Try to use the same model on single channel images and regular 3 channels, to compare difference (not finished in the step 3)
5. Try to use TinyML approaches: