The aim is to detect outlier seeds in a sample of a supposedly homogeneous population. I imagine my approach is going to be quite overkill but that's fine as I want to learn my computer-vision and deep learning concepts.
My initial approach is going to fine tune a pre-trained DNN (i.e. EfficientNet) on different kinds of seeds. Once fine-tuned I'll use the DNN to create embeddings which can then be used to compute distance from every other seed in a given photo. My thinking is that by fine-tuning on enough types of seeds with different shapes/sizes, the DNN will learn good representations that can then be used for outlier detection. This most definetely would be do-able by just segmenting seeds and computing known variables to run outlier detection on but eh, this seems fun :)
This repo is heavily a work in progress with not much structure to it. I've done very litte image processing before, so the resulting code is quite messy and lacks domain expertise (both computer vision and biology wise).
- Get bounding box code finalised to extract decent bounding boxes with little oversight
- Take way more seed photos (with more variation)
- Initial embedding creation
- Create image augmentation pipeline
- Demonstrate performance
Initial exploration notebook for looking at bounding box extraction. This was my first notebook I think? I tried masking and then running DBSCAN over the mask to get my groups. From there I ran some area filters on the DBSCAN clusters. It worked OK but was for sure convoluted and not a very good approach (i.e. cv2.findConnectedComponents on a mask does just that without DBSCAN).
Could not deal with shadows.
Explorating into the removing of shadows as the first approach would suffer from shadows being present. This notebook ended in me finding me a good filter for dealing with shadows.
Probably should rename this but this notebook was to finalise the image processing pipeline for bounding box extraction. Stil a work in progress but I've got a good baseline for bounding box etraction, just need to make it configurable for different seeds. (and then onto training !!)
Current pipeline:
