Apples – bitter pit detection
Problem: Bitter pit is one of the most severe disorders in apples. We observed Golden Delicious apples, which are more susceptible to bitter pit than others.
Goal: detection of bitter pit as early as possible.
Approach: Without knowing if and where bitter pit will develop, we captured a series of hyperspectral images of different apples over several days. Some of them developed bitter pit after a couple of days, so we were able to observe problematic areas at an earlier point in time.
Findings: We were able to detect and locate bitter pit areas at a fairly early stage.
Lemons – early mold detection
Problem: Parasitic mold pathogens begin to develop on the lemon’s surface before the changes are visible to the naked eye.
Goal: early mold detection and removal of unhealthy fruits; spread prevention.
Approach: We captured images of lemons with our hyperspectral imaging system for several consecutive days. Every day, we observed and recorded changes in the quality of the samples. One lemon (out of 20) developed extensive mold patches after three days. We looked for any significant spectrum differences between healthy and moldy lemons in the images captured before the mold was visible.
Findings: Using principal component analysis, we derived a component that indicates a clear difference between moldy and healthy lemons.
Oranges and other citrus fruit – surface defects detection
Problem: Freezing damage, physical injuries, scars and other types of surface irregularities can lead to faster fruit decay.
Goal: to calculate a health score based on the severity and extent of the defects
Approach: Images of oranges were captured using our hyperspectral system. The spectrum of an orange with no irregularities was used as a reference spectrum. We calculated the distance from the reference spectrum for each pixel of all captured oranges.
Findings: Classic computer vision using RGB images can only be used to detect various types of defects when they are already clearly visible. An HSI system allows us to detect irregularities earlier when they are not yet pronounced.
Tomatoes – damage detection
Problem: Tomatoes are a relatively delicate fruit, highly susceptible to damage due to shocks during harvesting, packaging and transport. Quality inspection is very important at each step of the process.
Goal: to detect damaged tomatoes at an early stage (when defect is not yet clearly visible).
Approach: We captured images of tomatoes using our hyperspectral system and derived a reference spectrum based on the healthiest sample. For evaluation purposes, the fruits were observed for several days afterwards.
Findings: Visible injuries such as cracks and bruises are easily detected at an early stage. Any damage quickly degrades the condition of the whole fruit, especially along the stalk. Timely detection is key to proper quality assurance.
Buckwheat flour – purity assessment
Problem: In order to reduce purchase costs and increase profit, fraud occurs in the purity of buckwheat flour. Instead of pure buckwheat flour, some retailers or suppliers may sell a mixture of buckwheat and cheaper wheat flour.
Goal: to obtain an estimate of the purity of a flour sample.
Approach: We analyzed two bowls filled with flour – one with pure buckwheat and one with mixed flour.
Findings: Seemingly the same, differences between the two were obvious after analyzing hyperspectral images. Pixel-based classification is not homogenous as the flour does not mix, but an average score for the whole bowl shows a significant difference between pure and mixed samples.
Chicken meat – increased water content
Problem: To achieve greater weight and consequently price, meat suppliers deliberately inject water into meat.
Goal: to detect pieces of meat with increased water content (calculate the percentage of water in samples).
Approach: We analyzed two samples of chicken meat using a hyperspectral camera. One piece had a normal water content, the other one was injected with water. Knowing that water absorption reaches its peak at certain wavelengths, we focused only on that part of the spectra.
Findings: The shape of the spectral curve is significantly different in samples with increased water content.
Chocolate – plastic particles detection
Problem: Pieces of plastic mold are present in chocolate bars. Detection is difficult due to add-ins (nuts, fruit, …).
Goal: to detect foreign (plastic) particles on the surface of the chocolate bar, even if the particle is smeared with chocolate.
Approach: Plastic particles of various sizes were placed on the chocolate bar; some of them were smeared with chocolate.
Findings: Even small and chocolate-smeared plastic particles are easily separable from hazelnuts and pure chocolate.
Wrapping Foil Detector
Problem: One threat to food safety is foreign object contamination, for which product recalls are very common; imposing significant costs on food processors, damaging their reputations, and putting consumers at risk.
Goal: to detect small flakes of plastic wrapping foil, particularly in minced meat, dried meat, and other sliced meat products as well as on frozen blocks of meat products after unwrapping.
Our solution: Our product uses multi-spectral imaging technology to detect flakes of plastic wrapping foil, bones, cartilage and fat lumps in meat products in high resolution and with a rapid detection speed.
Roasted coffee beans
Problem: Foreign objects may be present among roasted coffee beans. They must be removed before grinding.
Goal: to detect all non-coffee bean matters.
Approach: Some materials can be easily separated from coffee beans via color (RGB image), but some require more information. Combining visible color information, spectral curve shape and variability, we assign a coffee bean similarity rating to each pixel.
Findings: We were able to identify coffee beans based on combined color and spectral curve data.
Raw coffee beans
Problem: Different types of coffee beans may look alike, but the aroma is different.
Goal: to separate two types of raw coffee beans.
Approach: We captured two types of raw coffee beans using a hyperspectral camera system. We analyzed whether there is a significant difference in spectrum between the two.
Findings: Using principal component analysis and unsupervised classification, we determined that there are significant differences between the observed coffee types.
Cartilage detection in minced meat
Problem: After processing and mincing there is a chance that small cartilage particles are still present in the mince. Due to its small size and non-prominent color, manual detection is not effective.
Goal: to detect small cartilage particles in minced meat (burgers etc.).
Approach: We captured burgers using a hyperspectral camera. Cartilage particles of various sizes were manually inserted into the sample.
Findings: The spectrum of the cartilage differs from meat, fat and spices, so its automatic and reliable detection on the surface of the meat is possible.
Detection of plastic foil particles in pate
Problem: Frozen meat is usually wrapped in foil. Before processing and mincing, the foil is removed, but there is a chance of small torn foil particles entering the mix.
Goal: to detect small foil particles in the pate before packaging.
Approach: We filled the pate into the cuvette and added small particles of various types of foil. The foil was placed in the middle of the pate (was not visible on the surface). Light transmission was observed.
Findings: Foil particles one square millimeter or larger can be detected in up to 6 millimeters thick pate. The foil must not be transparent (we observed dark blue foil).
Coffee, cereals and nuts – analysis of spectrum differences
Problem: Grains of coffee, cereals and nuts can be very similar in shape and color.
Goal: to find out whether grains of different varieties can be separated by spectrum.
Approach: Using our hyperspectral system, we captured images of various types of grains: raw and roasted coffee, soy, peanuts, barley, oats, corn, proso millet, spelt, buckwheat, quinoa, amaranth, kamut, rye, black and brown rice.
Findings: Using the combination of principal component analysis and unsupervised clustering algorithm, imaged grains were appropriately classified into 16 different classes. In combination with grain shape and size detection, reliable automatic classification and foreign matter detection are possible.
Detection of can damage
Problem: Cans must be undamaged before the products can be delivered to retailers. Manual inspection is possible, but is often unreliable and slow.
Goal: automatic detection of damaged cans.
Approach: Damaged and undamaged cans were scanned using a laser-based 3D scanner. A 3D reconstruction of each can was analyzed.
Findings: Automatic detection of dents on the surface is possible.
Detection of damaged vials
Problem: The pharmaceutical industry has strict packaging standards. The packaging must be undamaged, the content must be accurately dosed and of an appropriate structure, not containing any foreign objects. We focused on transparent glass vials filled with fragile pills (the dusty inside of the vial walls is considered normal).
Goal: to detect cracks, scratches, chips and other defects on the glass.
Approach: We developed two innovative image capturing techniques, based on which irregularities on all parts of the vial can be detected (bottom, torso, shoulder, neck).
Findings: Cracks and chips are easily detected. Shallow scratches are more challenging, but also detectible with a combination of the two techniques.
Detecting defects on spruce boards
Problem: Wood is an organic material, so each board has unique features. Knots, cracks and resins are normal to some extent, but some types of irregularities can affect wood performance and ornamental value.
Goal: to detect, classify and measure defects.
Approach: Hyperspectral cameras allow us to distinguish between resin and a crack, which may look alike on an RGB image. Based on edge detection algorithms, we distinguish live wood knots from dead ones.
Findings: We found that a reliable classification of defects into 4 main classes is possible (crack, resin, live knot, dead knot). Using a combination of classification, precise measurement and user-defined rules, it is possible to automatically remove boards that are unsuitable for further use. Using an automatic system, manual inspection would no longer be necessary.
Steel lamination stacks – defect detection
Problem: Some irregularities may occur during production – such as scratches, excessive spacing between individual laminations etc. Manual inspection is possible, but slow, subjective and therefore often unreliable.
Goal: automatic detection of any deviation from the expected / tolerated form.
Approach: We employed an unsupervised machine learning algorithm to detect anomalies in the lamination stacks.
Findings: The method proved effective on an independent test set – even small defects were accurately detected.
We'd love to talk about your project
Our experts and developers would love to contribute their
expertise and insights to your potential projects.