1. Subcutaneous Fat Depth Regression Using Hyperspectral and Depth Imaging
- Author
-
Pena, F, Mehami, J, Guenot-Falque, R, Patten, T, Alempijevic, A, Vidal Calleja, T, Pena, F, Mehami, J, Guenot-Falque, R, Patten, T, Alempijevic, A, and Vidal Calleja, T
- Abstract
Robotic perception is becoming an important component for automation in the meat processing industry. Whether for contaminant detection or automatic cutting, multimodal perception systems, in particular, based on hyperspectral imaging have the ability to provide information that goes beyond the texture and colour of a surface. In this paper, we present a learning-based method to estimate subcutaneous fat depth in meat cuts by leveraging hyperspectral data models that rely on the knowledge of modelled light sources and surface shape information. Data from a fully calibrated hyperspectral and colour depth (RGB-D} camera system is used as input. Fat depth ground truth is recovered via a novel systematic approach that ray casts a computed tomography (CT) mesh of the meat cuts, which is non-rigidly aligned with a depth reconstruction captured by the {RGB-D} camera. We thus evaluate machine learning methods that can handle small datasets, by employing dimensionality reduction and data augmentation to address the limited amount of imbalanced data that is acquired. Our results show that leveraging shape and light models, coupled with machine learning methods that capture nonlinearities and spatial correlations produces the most accurate results.
- Published
- 2022