This is a side-project I was working on a few years ago. I created an algorithm that can take pictures of honeycombs and find out how many cells of each type are on it: small/medium/big larvae, capped/empty cells, cells with eggs/nectar/pollen.
I splitted the task into 2 problems. First, segment the whole honeycomb picture into small pictures of each cell. Then classify each cell based on its small picture.
- segmentation a segmentation algorithm using classic computer vision
- classification a convolutional neural net for classification
My results are so so: >80% accuracy in both detecting cells and classifying them. However, if you are reading this now I would recommend DeepBee. They essentially did the same, but their results are much better. If their code would have been public earlier, I probably would have just used DeepBee.
