Identifying invasive hyacinth flowers in images of forest scenes using computer vision (competition link)
First attempt: convolutional neural net (Keras with Theano backend) using images resized to 150 x 200 pixels: 68% accuracy in 2 epochs. Very slow with the larger images. Definitely room for improvement.
Second attempt (code is in /src): CNN using images centered and resized to square (256x256), then cropped to 224x224. Used keras image generator to generate images that are rotated, random height/width shifts, flipped on horizontal axis). Trained neural net on top of VGG16. 98% accuracy in 40 epochs. This code was adapted from (work by kaggle user fujisan).
Things to try next:
- Train on larger images (Using images resized to 512x512 and cropped to 488x488 so far is not working very well.)
- Try multiscale model
- Identify regions/partitions of the image where flowers are located