Food safety is a great public concern, and outbreaks of food-borne illnesses can lead to disturbance to the society. Whenever foreign objects are found, food producers recall all of the contaminated food. The brand image, which takes a long time to build, becomes undermined. The Ministry of Health, Labor and Welfare is considering the introduction of more stringent regulatory requirements, such as mandatory reporting to the prefectural governor. The greater demand for food safety as well as the safety requirements will further increase the burden on food manufacturers. In the past, food manufacturers tried to automate the food inspection process by using a color camera, however, it is a very difficult process as the shape, color and size of the food particles and foreign objects differ. This has resulted in human visual inspections being utilized as the main approach for food inspections.
Under this backdrop, inline inspections using hyperspectral cameras are attracting much attention as the new technology for automated inspections.
Characteristics of hyperspectral cameras
Visible light is usually defined as having a wavelength in the range of 400–800 nanometers (nm), and the color of visible light depends on its wavelength. White light, for example, is composed of wavelengths of different colors. Standard cameras cannot reproduce all of the wavelengths, and the combination of all visible colors is only approximated. When you take a photograph of a tree with a RGB color camera, you will find strong brightness in G as shown in Figure 1, in which the colors are divided into three bands of red, green and blue. On the other hand, multispectral and hyperspectral imaging allows for more sophisticated measurement of wavelengths. While RGB divides the visible light into three bands, multispectral imaging divides the visible light into several tens of bands. Hyperspectral imaging can measure the visible light in several hundreds of bands.
Figure 1: Difference between color cameras and hyperspectral cameras
Hyperspectral cameras have spectral sensitivity to near-infrared light, and, thus, can distinguish objects made of different substances. Figure 2 shows the image of a stone and a potato by a color camera (upper image) and a hyperspectral camera (lower image). Since the color and shape of these two objects are very similar, it is difficult to distinguish them using the color camera. On the other hand, the differences of the two objects are clearly identifiable in the image taken with the hyperspectral camera. Figure 3 shows the wavelength (nm) and absorbance of the stone (green) and the potato (red). The absorbance of the potato jumped around a wavelength of 950nm. This is because water contained in the potato absorbs the wavelengths of this range. On the contrary, the graph of the stone shows no signs of change in this range because a stone does not contain water and, thus, does not absorb light. This is how the distinction between a potato and a stone can be made. Absorbance characteristics vary depending on the materials, and spectral sensitivity to near-infrared light of a hyperspectral camera makes it possible to distinguish similar looking objects made of different materials.
Figure 2: Pictures of a stone (upper) and a potato using a color camera and a hyperspectral camera
Figure 3: Spectral information of a potato and a stone
The characteristics of a new hyperspectral camera are very useful for the field of machine vision. However, a hyperspectral camera typically costs over JPY 10 million and is used in limited fields such as in the research for aerospace applications. Assembling a hyperspectral camera manually requires precise manual work. Its optical parts also need to be made by combining different optical parts available in the market. These factors contribute to the high cost of hyperspectral cameras.
Innovative approach of Specim, Spectral Imaging Ltd.
Specim, Spectral Imaging Ltd. of Finland has finally addressed these challenges and managed to lower the cost of a hyperspectral camera dramatically. They addressed the two challenges above by following the below approaches:
- Instead of combining readily available optical parts, they design the optical part specifically for a hyperspectral camera and reduce the cost
- Automate and/or simplify the production process, which used to be done by precise manual work, and develop a mass-production system
Thanks to these approaches, Specim has successfully managed to halve the cost of manufacturing a hyperspectral camera. At the same time, it has also managed to reduce the size of the camera, opening up the opportunity for using a hyperspectral camera for machine vision.
Applications of hyperspectral cameras
Here are some examples of the use of hyperspectral cameras in the food inspection process.
A selector is a machine to remove foreign objects from granular products such as rice, wheat and beans. Conventional machines are equipped with a 3-band RGB camera, but as outlined above, RGB cameras cannot distinguish objects with similar colors and/or shapes. Spectral information obtained by a hyperspectral camera is useful to identify foreign objects.
Figure 4: Detection of raisins, foreign objects and haulm (Photo courtesy of Perception Park)
Spectral information is also useful in inspecting meat products. As shown in Figure 5, meat, fat and bones are clearly identifiable in the image taken with a hyperspectral camera.
Figure 5: Classification of meat, fat and bones (Photo courtesy of Perception Park)
② Appearance inspection
Whenever the exterior of an apple is hit or damaged, cells in the impacted area are ldamaged and the moisture content increases. A standard camera is unable to detect this change. Hyperspectral cameras have spectral sensitivity to near-infrared light and, thus, can capture the difference between the damaged part and the intact part of the apple.
Figure 6: Appearance inspection of an apple (Photo courtesy of Perception Park)
③ Biting inspection
Hyperspectral cameras can be used not only for the inspection of the food itself, but also for the food packaging. Biting inspection is of high priority because the food can go bad during the shipping. In many cases text is printed on the packaging seal, and in such cases, biting cannot be detected by 2D image analysis. As shown below, hyperspectral cameras can detect the biting accurately.
Figure 7: Example of biting inspection
The use of hyperspectral cameras in tasks that have traditionally been done manually can save manpower. Also, their use can increase the accuracy of the inspection because spectral imagery delivers much richer information compared with standard cameras.