The intensive use of bone band saws in the meat processing industry results in frequent cuts and lacerations on the part of the operators of this equipment. These injuries are serious, both in terms of personal suffering and in economic losses. The aim of this research project is to develop a method based on reflection spectroscopy and an accompanying sensor system for non-contact detection of human skin and its distinction from dead animal tissue in the danger zone of saws. Previous research projects have used image-evaluating techniques or methods based on infrared reflection. However, neither method has provided a reliable distinction between animal tissue and human skin.
In the present research project, an approach based on visible and near ultraviolet reflection spectroscopy was examined. In a first step, extensive measurements of the reflectance spectra were carried out on human skin (fingertip, back of the hand, palms, forearm) as well as on various types of meat (pork top, pork neck, pork fat, pork rind, lamb chop, lamb knuckle, beef, veal, chicken breast).
Meat and human skin contain a number of chromophores (light absorbing molecules), which lead to different reflection spectra. The most important chromophore in the context of our research is melanin, because it occurs in human skin but not in meat. This manifests itself in a significant difference in reflection spectra in the near UV between 300 and 400 nm. Classification of the spectra was done using classical machine learning techniques: Support Vector Machines (SVM), Boosted Decision Trees, Artificial Neural Networks and k-Nearest Neighbor Algorithms. With supervised learning with data sets (in our case spectra) whose belonging to different classes (in our case the two classes "human skin" and "dead animal tissue") is known, and the parameters of the classifier are fixed. The classifier is a mathematical decision function that is able to assign new data sets to the respective classes. The best detection rates were achieved with Support Vector Machines (SVM) and with Artificial Neural Networks.
If a sufficiently large number of spectra are used as training data of the machine learning method, the classifier achieves the required low error rates (less than 10-6). In addition, it could be shown that one can achieve a reliable separation with as few as 9 or more wavelengths.
In addition to the theoretical work, a prototype was also developed. The detection unit used was a Hamamatsu micro spectrometer. The data processing (including classification) was undertaken with an Arduino 101 microcontroller. The special feature of this microcontroller is that it contains an additional chip with special features: this chip (Intel Curie) is a so-called neurochip, which implements a neural network in the form of dedicated hardware.
The objective of the research project (reliably and contactless distinction of human skin from dead animal meat by means of reflection spectra) has been achieved. Extensive studies on suitable classifiers and algorithms have shown that the refection spectra of human skin and animal meat can be differentiated safely.
The use of a low-cost micro spectrometer together with an implemented neural network make the process developed in this research project so universal and so adaptable that it can be used in many areas to prevent saw accidents.
foodType of hazard:
prevention, machine safetyDescription, key words:
Hand detection, distinguish hand - dead animal meat