Main feedback, summed up:
A close to optimal bubble detection can be obtained by retraining the Mask RCNN network. Here is the link to the official github repository: https://github.com/matterport/Mask_RCNN
Traditional pre-processing
Our data set is affected by noise (the cells) and non-contant enivronment parameters such as illumination, shadows, various disturbtions. When this is the case, traditional preprocessing shows very bad performance.
Hence, our alternatives are the following:
- Ensure constant parameters: contants lighting, uniform background, decrease probabilities of shadows, dust etc.
- Model the noise and the change in parameters in order to compensate for these in a more structured and controlled way (rather than relying on built in adaptive thresholds that may or may not work).
- Training of more advanced neural networks
Darknet YOLO (You Only Look Once)
Darknet YOLO is a veryo common neural network, but the reason why it is so broadly used is because people adapt from others and this one is known as the one "everybody uses " (according to Leonardi).
In our small period of trying this out, we experienced that training this network with the raw images resulted in relatively bad performance. Lots of false positives and few actual detections. However our results improved drastically when the dataset was binarized with edge detection and dilation (hence a fairly heavy and slow preprocessing procedure in advance). According to Leonardi this makes sense as these networks are may easier to train on binary images. However, he was not convinced that the results were optimal.
Mask RCNN vs. Darknet YOLO
There are two main differences between these to neural networks that in Leonardi's opinion would lead to way better performance using RCNN.
- Mask RCNN labels each pixel in the image. YOLO on the other side only tries to find a bounding box around the desired contours.