Computer Visions System for Damaged Yolk Optical Recognition

The system utilizes a Machine Learning technique, namely Mask R-CNN (Mask Regional Convolutional Neural Network) to detect egg yolks and classify them into two categories – proper yolks (ready for separation) and damaged yolks.
Machine learning means that our engineers didn’t manually specify the rules for yolk shape – instead, we have labeled images extracted out of production videos and made the algorithm use try-and-error with backpropagation algorithm to adjust it’s internal parameters.
We have chosen this approach (after delivering initial solution based on manual feature engineering few years ago), because it achieved best accuracy and robustness to changing conditions.
This particular neural network type (Mask R-CNN) is highly resilient to changes in orientation, scale (zoom), lighting conditions, minor dirt elements, even to major disturbances such as different background.
It is also a ‘deeply supervised’ model, which means that it’s relatively easy to explain why the model has made particular decision (compared to so called end-to-end models, such as image classifiers or regression models) and improve it if needed.

 

Watch on YouTube: