Big data and Graphic Processing Unit (GPU) accelerated data science for high-throughput carcass and pork quality phenotyping

Project 21-06 - Dr. Stephanie Lam

Objectives:

1. Develop high-throughput computer vision system to efficiently measure pork meat quality and carcass primal cut traits.
2. Provide a platform using the resulting database and data analytics to develop precision farming technology.

The project resulted in satisfying both objectives by augmenting the size of the existing AAFC phenomics database with more than 3,000 new images and developing a series of analytic and automated segmentation approaches for marbling and identification + separation of hams, shoulders and loins from background in digital images. The project explored the impact of various environment-, experimental- and sensor-based factors on the detection and scoring quality of intact pork products such as hams, shoulders, loins and bellies. The segmentation procedure includes deep learning modelling based on Mask R-CNN and two neural network architectures (ResNet 50 and 101), which proved to reach close to real-time FPS processing speeds.

Through this project, the existing pork phenomics database was expanded, establishing correlations between desirable meat quality traits, and creating a foundation for precision farming. This initiative produced an excellent technological starting point for innovative solutions that enhance research and production efficiency via the application of deep learning modelling for automatic instance segmentation of important meat part regions of interest and computer vision analytic approaches for uniform and automatic meat scoring, positioning the Canadian Pork industry at the forefront of advancements in research and sustainability.