Introduction
Food process control necessitates real-time monitoring at critical processing points. Fast and precise analytical methods are essential to ensure product quality, safety, authenticity and compliance with labelling. Traditional methods of food monitoring involving analytical techniques
such as high performance liquid chromatography (HPLC) and mass spectrometry (MS) are time consuming, expensive and require sample destruction. Near infrared spectroscopy (NIRS) is well established as a non-destructive toolfor multi-constituent quality analysis of food materials
(Scotter, 1990). However, the inability of NIR spectrometers to capture internal constituent gradients within food products may lead to discrepancies between predicted and measured composition. Furthermore, spectroscopic assessments with relatively small point-source measurements do not contain spatial information, which is important to
many food inspection applications (Ariana, Lu, & Guyer,2006). Recent advances in computer technology have led to the development of imaging systems capable of identifying quality problems rapidly on the processing line, with the minimum of human intervention (Brosnan & Sun,2004; Du & Sun, 2004). RedeGreeneBlue (RGB) colour vision systems find widespread use in food quality control for the detection of surface defects and grading
operations (Chao, Chen, Early, & Park, 1999; Daley,Carey, & Thompson, 1993; Throop, Aneshansley, & Upchurch, 1993). However, conventional colour cameras are poor identifiers of surface features sensitive to wavebands other than RGB, such as low but potentially harmful con-
centrations of animal faeces on foods (Liu, Chen, Kim, Chan, & Lefcourt, 2007; Park, Lawrence, Windham, & Smith, 2006). To overcome this, multispectral imaging systems have been developed to combine images acquired at a number (usually 3e4) of narrow wavebands, sensitive to
features of interest on the object. Compared with conventional analyticalmethods such as HPLC,multispectral imaging systems can perform non-destructive analyses in a fraction of the time required (Malik, Poonacha, Moses, & Lodder, 2001).