The Failure of RGB Systems in Food Quality Control

However, in spite of all their obvious advantages, vision systems are still not widely exploited in the food industry. Indeed, the problems related to food products, that is, biological products that have been submitted to various processing operations, stem from the wide variety of states they can be found (liquid, solid, fragmented, etc.), shape, color, chemical composition and so on. The same commodity can vary drastically depending on origin, or between the beginning and the end of the season. 

When technically - and economically - feasible, appropriate artificial vision systems have been developed for the food industry based on classical RGB video cameras. However, far too often, RGB vision systems fail to detect defects or contaminants. There are two main reasons for this poor performance. 

The first reason is that the colors of the objects to be discriminated are too similar the three color components (RGB or hue intensity saturation coordinates) are not discriminative enough to detect the differences between the object of interest and the surrounding objects. This problem is amplified by the high color variability associated with biological products the variance of the colorimetric components inside an object may be greater than the variance between this object and the objects to be removed. In this case, the solution is to look for other pieces of information, either in the visible spectrum or in the NIR spectrum. The NIR spectrum provides information about the chemical composition and the internal physical structure, which can be used to distinguish same-color objects of different compositions. 

The second reason for the failure of RGB systems is the complexity ofthe scenes to be analyzed. Except for flat products (meat slices, fish, biscuits, etc.), the third dimension of food products can be relatively unusual, with concave areas leading to shadowing and color changes. This is typically the case of the calyx cavity in fruits, which is generally seen as a defect by classical RGB vision systems. The same misclassification problems due to concavities occur for chicken carcasses. However, NIR imaging can be used to reduce the effects of shadowing and of depth of field. 

Although technically and economically feasible artificial vision systems have been developed for the food industry, based on classic RGB video cameras, the 

Consequently, within an industrial environment, hyperspectral NIR imaging has been used primarily to improve object detection previously carried out using RGB cameras and optical systems. Yet, the many other advantages of these systems make them highly suited to not only on-line analysis but also high-speed laboratory analyses for product grading. 

---