The present invention describes a procedure for multivariate hyperspectral image analysis in real time in accordance with patent claims.

Technical background and prior art

In many industrial settings, there is a need to measure and analyse whether at least one object's (product's) chemical content and physical characteristics vary and/or deviates from at least one predefined reference value (reference range) for the object (product).

Measurements can be made using several different techniques in which at least one information sets regarding the object's chemical composition and physical properties are collected and compared to at least one reference data set for the chemical content and physical properties of the item. One problem associated with the measurement of the object's chemical and physical properties is that the measurement is usually done with methods where the object is destroyed or carried out with methods where the object must be discarded after the measurement.

Data about an object's chemical composition and physical properties can be gathered by using spectral analysis. The use of spectral analysis for collecting data sets from analyzing object's chemical composition and physical properties has been widely used for a large number of procedures and applications. A standard technique for acquisition of spectral data is by measuring a single point in a sample using a spectrometer; hence it is referred to as a single-point measuring device.

When measuring with a single-point measuring device, measurements will only capture information from a limited section of the object's surface.

With only a small part of the entire surface of an object being measured with a single point measuring system, there is a requirement for samples to be homogenously distributed in terms of chemical and physical properties in order to obtain valid and representative results.

Another disadvantage with this approach is that it mixes information originating from noise, the sample itself and deviations within the sample. This is because it assumes that the points for which the measurements are taken to provide representative information for the property to be characterized, which is often not the case for an inhomogeneous substance. When multiple single point measuring devices are used in parallel there is also a problem to synchronize the instruments.

Another spectral analysis method is hyperspectral (image) analysis. With

hyperspectral image analysis can information regarding the chemical composition and physical properties of a whole objects surface be efficiently collected. The problem with using existing instrumentations for hyperspectral image analysis is that the collected amount of data is usually very large. The shear size and complexity of the data requires advanced processing capabilities for analyzing the amount of information. The requirement for processing capability means that the measurement and analysis cannot easily be done in real time as the data is generated and is done with a time lag between the time of sampling and the time of analysis.

For industrial production, the delay in identifying deviations in chemical composition or physical characteristics has the effect that a large number of defective products will be produced before the problem is identified and rectified. This may even mean, especially for sensitive products in food and pharmaceutical industry, that the entire batch must be disposed resulting in very high cassation costs.

One problem with performing hyperspectral image analysis in industrial processes is that the object to be analyzed is usually transported on a conveyor or similar linear transportation unit. A common problem with existing techniques for hyperspectral image analysis in real time is that these cannot sufficiently fast cope with the speeds at which the products to be analysed is transported. Lowering the carrier speed is not an option since it would result in decreased production efficiency.

There is also a problem that existing systems and procedures for performing hyperspectral image analysis is highly specific and intended to be used for only one or a few applications. Implementation of additional applications results in purchase and development of new systems, which is expensive.

Within the specific applications where the procedure under the present patent application is used, there are problems solved by the procedure. For preparation and production of food exists, for example problems with variations in the quality of the manufactured food, chemical, sensorical, structural and temporal.

As an example, cheese quality is today analyzed via destructive analysis where a sample from the cheese is taken and analyzed in a laboratory. A specific problem with the analysis of cheese is that it is inherently inhomogeneous (heterogeneous), which means that it can be difficult to obtain representative information from individual test samples. The cheese manufacturers do not with certainty know the variation in quality between individual cheeses, within or between batches. If the analysis of a cheese sample would show a deviation in quality it is still difficult to know the amount of products that need to be discarded. In the final analysis, the cheese quality variations lead to the customer's perceived satisfaction with the resulting cheese is not on par with either the customer's expectations or customer experience of the same type of cheese. This may even lead to the customer stops buying cheese from the

manufacturer in question. Uncontrolled quality variations can also lead to lowered of machine utilization in the manufacturing process.

Known technology

Systems and methods for making and implementing hyperspectral imaging is already known. As an example is described in patent document WO2008099407 a method for implementation of hyperspectral imaging. The procedure described in the above-referred document differs widely from that of this patent

application. For example, is the described procedure not intended to be used for analysis in real time.

Another system for performing hyperspectral image analysis is also described in patent document WO2006102640 by the applicant Infotonics Technology. The patent document describes a system for using hyperspectral imaging in the medical context, or food contexts. The procedure described in the above-referred document differs widely from that of this patent application.

Line scanners are already known in a number of variations. The company Specim markets an example of line scanners. Also the companies Neo and Head Wall are marketing variants of line scanners. Line scanners are often based on so-called Push Broom technology.

Methods to monitor the cheese with the help of spectral analysis are already known. For example, described in the patent document US20090305423 is a method to evaluate cheese with at least one infrared spectra at least one point. The procedure is based on a sample taken from cheeses and analyzed alongside the production line. The method differs widely from the method according to the present patent application by the use of prior destructive sampling. It analyzes not the cheese's natural features such as holes, cracks or colour. The presence of foreign objects such as mold, plastic and other deviations cannot be identified with the procedure according to patent documents.

The patent document WO2007000166 describes a method to collect and analyze spectral information from samples such as cheese. In the patent documents described method differs widely from the method in accordance with the present invention. For example, sampling is conducted by piercing the cheese with a sensor probe. The patent documents describe a procedure using destructive sampling. Furthermore, there is no measurement or analysis of the cheese physical properties such as holes, cracks, colour. The presence of foreign objects such as mold, plastic and the like may also not be identified with the procedure according to patent documents.

The patent document WO0248687 describes a procedure for real time

hyperspectral analysis for measuring the chemical content and / or the physical properties of seeds in a non-destructive way. The patent document covers in great detail the general technical approach for hyperspectral analysis using a push-broom instrument. The patent document has a strong focus on the applicability of hyperspectral analysis for sorting and processing of seeds and grain to obtain desired chemical and physical properties. The statistical modeling procedure described is rather simplistic and without hierarchical decision support or explicit deviation detection, results in limitations in accuracy, robustness and general usefulness of the procedure. For example will contamination of deviations such as minerals or soil give non-representative spectral profiles for the objects resulting in erroneous predictions. Another example is that deviations must be included in the modelling procedure in order to be identified.

With the above problems and known technologies in regard there is a need for a non-destructive general procedure for hyperspectral image analysis where the analysis of the collected data is done with multivariate analysis.

The objectives of the present invention

The main purpose of the present invention is to provide a procedure and a system that solves the above problems. Another purpose of the present patent application is to provide a procedure that can be used for a wide variety of applications. A further purpose of the present invention is to create a procedure in which the test is done essentially by non-destructive testing (sampling). There is also a purpose of the present invention to provide a procedure whereby the hyperspectral image analysis is performed via multivariate analysis. A further purpose is to provide a procedure by which the hyperspectral image analysis can be done in real time. A yet further purpose of the present invention is that the calibration procedure can be performed based on the reference values of selected parameters or based on structural properties of objects, and the calibration procedure can be performed automatically in real time.

Brief description of figures referred to in the following paragraphs In the following detailed description of the present invention, the following figures will be referenced to. These are outlined in the following figure list. The figures in the exemplified embodiments are not limiting the scope of protection of the present patent application. Note that the figures are schematic and that the details thus can be omitted in these.

Figure 1 illustrates schematically a procedure for implementation of multivariate hyperspectral imaging, according to a first embodiment of the present invention. Figure 2 shows the identification and separation of the measured object from the background, and identification of structural deviation / characteristic of the measured object

Figure 3 shows a hierarchical model system for classification and quantification. Figure 4 shows schematically a device to be used for sample measurements.

Detailed description of the invention

In the following paragraphs, a procedure for the measurement and analysis of objects (products) chemical composition and physical properties using hyperspectral image analysis will be described. The procedure can broadly be divided into two levels. In the first level, a model system is constructed

(structured system of statistical models) so that it can take on hyperspectral image data to measure (predict) the chemical and structural properties of an object (product). In the next level the model system is applied for real time measurement of in the previous defined level defined and modelled properties. The procedure can be used in a variety of application areas such as the food and pharmaceutical industries and plays an important role for quality control by the classification and quantification of relevant chemical and structural properties.

The by the procedure measured physical and chemical properties are then compared with at least one predefined value (range) of relevance for product quality. The procedure provides a unique advantage in that it can replace costly and time-consuming laboratory measurements with a fast and non-destructive analysis. In some cases, the procedure can in addition improving the efficiency of analysis also provide a method for measuring characteristics otherwise difficult to define, such as to describe the distribution of chemical and structural properties of a surface. In the procedures first level, analysis systems are built up and optimized to measure for the application area relevant variables and to describe important structural features. On a global prediction level, the procedure can be used to classify a product from its hyperspectral information content and automatically select a specific applicable model system to use for measuring the chemical and structural properties of the product. The procedure also enables users to automatically improve model systems predictive ability by determining whether the new measurements contain new potentially

informative variation. If it results in the model describing more variation and better predictions the model system can be automatically updated.

Figure 1 shows schematically one embodiment of the present procedure. The procedure of the illustrative procedure consists of eight steps. In alternative embodiments, this procedure can be defined with more or fewer steps. The steps can also be defined with other for this purpose appropriate names that are included in the basic intentions of the present patent application. Two or more step can also be combined into one step.

In the first step, referred to as problem definition, in the overall procedure analyzes the application-specific conditions and problems that exist within the specific application's technical area. As an example, parameters/properties of relevance to analyze for an object within the application scope, is identified and defined. These parameters can for example be of importance for determining the object's quality.

The objects (products) chemical composition and the overall properties for the object at whole can be referred to as such parameters. Furthermore, the parameters can refer to specific substances / compounds and its concentration in the object. In addition, parameters can also include factors relating to the object's physical form and the object's physical properties. Parameters can also refer to other measurable quantities that are crucial for object properties such as structural abnormalities, the presence of undesirable chemical compounds, and similar items.

In the second step of the procedure, referred to as study design, a study design is prepared based on the analyzed problem definitions and the identified parameters of relevance. The experimental design results in at least one sample measurement, with a non-destructive method such as hyperspectral

measurement, and at least one reference measurement is planned.

Preferably a number of different hyperspectral measurements and at least one reference measurement are planned.

In the third step of the procedure, referred to as the sample measurements, the samples are measured according to the study. Sample measurements are intended to determine the presence and amount of relevant chemical substances in the object from the collected spectral information (raw data) from the surface of the object or its depth wise proximity. Through the test measurements, ranges of the normal variations of the object's chemical composition is calculated and determined for each variable.

During data acquisition, objects are illuminated by at least one electromagnetic radiation source. The electromagnetically radiated surface absorbs and reflects light of different wavelengths to different extent. The reflected electromagnetic radiation from the electromagnetically radiated object (and the surface that the object is located on) is recorded by at least one sensor, such as preferably at least a hyperspectral sensor, camera or similar device. The reflected ligth detected by the camera (sensor) is converted into at least three-dimensional data sets (data cubes) that consists of at least two spatial dimensions and at least one spectral dimension. During data acquisition, both the surface of the object, deviations on the surface of the object and parts of the background are scanned (Figure 2). The collected data volumes are stored on a for the purpose suitable stationary or portable storage medium such as a hard drive on a computer connected to the camera.

In the exemplified procedure, the camera consists of a hyperspectral line scanner. The sequential acquisition of lines creates a three-dimensional volume of information (such as data cubes) including at least two spatial dimensions and at least one spectral dimension. During measurements the spectral dimension is preferably hyperspectral, resulting in spectrum in which the number of wavelengths (variables) exceeds 19, and in many cases even greater than 100 wavelengths.

The procedure can also involve the analysis of multispectral data sets with a smaller number of wavelengths which do not necessarily represent a continuous spectral profile. References to hyperspectral sets of information in patent documents can instantly be changed to also apply for multi-spectral data sets.

In the exemplified procedure, both for the calibration procedure and for the prediction procedure, the hyperspectral data is automatically normalized against a reference object with of high and stable absorbance.

The normalisation procedure can be practically carried in many ways in which the reference object is automatically measured without affecting the overall measurement process. Examples of this include using a fold-down arm or a cradle.

Although the use of a hyperspectral line scanner is described in this patent, it should not limit the scope of the protection for the application. In alternative embodiments, other for the purpose suitable devices can be used to collect the hyperspectral data quantity.

After sample measurements the objects are labelled in a suitable manner. The labelling may be done by wrapping the object in a suitable container and that the container (or item) is provided with a label such as bar code or with another for this purpose suitable labelling method. In alternative embodiments, it is conceivable that the object in itself is directly labelled.

In step four, called the reference measurements, the samples hyperspectrally measured and labelled in step three are further reference measured in for example a laboratory.

In the laboratory, comprehensive measurement and analysis of the parameters assessed as relevant in step one is performed.

The reference measurements are performed with appropriate techniques, not necessary related to spectral analysis. The techniques are already known techniques or future-developed techniques. These techniques can be of destructive nature, destroying the measured object of. The laboratory techniques can also be more accurate and precise than spectral analysis but with the drawback of being expensive, time consuming and labour intense. Through the object labelling, information from the test measurements and from the reference measurements can be connected for subsequent statistical analysis.

In step five of the general procedure, called analysis phase, is statistical / mathematical analysis of the collected amount of information from the test measurements (spectral data cubes) performed.

Classification of the hyperspectral data content is initially performed, splitting the information into at least two classes using at least one classification algorithm. The information is preferably separated into multiple classes.

Figure 3 illustrates performing classification with a global model based on a number of local application-specific models. With the global model structure, classification can be performed at several different levels. At a top level it distinguishes between objects (samples) and background (Figure 2). The sample can then be classified by the type of object that the object is made up of such as for example food or drugs. The object can also be classified as unknown. The unknown substance can be identified in that it deviates from what is previously known and need not previously have been calibrated in any model. The food can then be classified in more depth as dairy products, meat, or unknown. Dairy product can be further divided into sub-class of cheese, which is exemplified in the below presented cheese application.

For the cheese products, chemical content such as water and fat and several other compounds can be quantified. Furthermore, the presence of physical structures of relevance for the quality and characteristics of cheese, such as holes, cracks and other structural variations are classified and quantified. By using a combination of multivariate modelling and traditional image analysis can and substructures and additional objects be identified based on its chemical characteristics and multivariate latent variables. This in turn is useful in the preparation of models that correlate the reference measured parameters and spectral data sets. Structural aberrations and non-representative sub-objects can then be excluded from the calibration procedure and for following predictions.

It is also unique for the presented procedure that a statistical selection of informative partial information sets (spectrum) can be made for each class. The selection of representative spectrum for the different classes are performed by statistical methods in order to maximize the information content and to ultimately give as good a representation as possible of the reference measured parameters. By the procedure, the number of partial information sets (spectrum) and thus the total data volume is reduced to a large extent, resulting in that less memory space is required for storage of the information quantity. For example, a data set of 100 MB is reduced to few MB. The processing of the files is also faster in the modelling step.

The selection process has also led to an unexpected technical effect in that it solves the imbalance problem with classification models such as PLS-DA. What this means is that if there is a large imbalance in the number of spectrum that the different classes contains, a statistical selection is performed to reduce the number of spectrum for the larger classes while still retaining the

information content in the data. An alternative embodiment is to use a simulation step to increase the number of spectrum for the less represented classes. Solving the imbalance problem gives more accurate and reliable predictions for the modelled parameters.

In the analysis is also the spectral raw data from the measurements (information quantity) merged from the results from the reference measurements. In alternative embodiments, all or some of the reference values for the parameters can be retrieved from at least one database. As an example, values from the reference measurements are added to a database when they are measured in the lab. In the statistical analysis step is the data from the hyperspectral

measurements correlated with the results from the reference measurements. The aggregated information quantities are analyzed with at least one software.

In step six, referred to as the Model System, at least one specific model or model system is created. At least one application specific model system is developed (modelled) with the analysis software. The model system includes at least one classification model, and preferably at least one quantification model for each specific application. Furthermore, structured workflows are created to be used in real time for the specific application.

The dedicated calculation engine used for quantification and classification in real time is optimized for maximized CPU and memory utilization. The structural framework of the workflow system is designed to provide support for this functionality. This allows for the calculations to be parallelized by either dividing the calculations for individual models and / or partial information sets to different processors in order to obtain maximized computational efficiency. The processing of the amount of information is done sequentially or in parallel.

The combination of the multivariate modelling procedure and optimized processor utilization for the calculation engine makes it unnecessary to restrict the statistical analysis and prediction to be based on only a few informative wavelengths. The above limitation step is called variable selection and is sometimes performed because some statistical methods cannot analyze data sets with many correlated wavelengths and that the calculation engine do not have the performance to analyse the large amounts of information common in hyperspectral image analysis.

In step seven of the procedure, referred to as At-line prediction, the model system is used next to the production line or in similar setting to classify and quantify previously reference-measured parameters. Samples are taken from the production line or redirected to a parallel line (possibly miniaturized) and is measured by the hyper-spectral analysis equipment. In the case that the measured test samples contains information not represented in the model system, there is a continuous fine-tuning of the model system through

performing additional reference measurement and updating of statistical models by introducing the variation in the modelling procedure. The models are automatically updated to improve the prediction accuracy of the system.

The step eight of the procedure, referred to as Integration On-line, the applicable model system (developed in Step six) is integrated On-line in the production line or in similar setting. In the On-line mode, the parameters reference measured in step four and statistically modelled in step five can be automatically measured in real time through the application of the developed model system. The model system used for measurement is composed of at least one multivariate model. A unique feature of the present invention is that predictions are performed by at least one software including at least one algorithm instantly after data

acquisition by the camera. One possible application is that a quantity of data for an entire object is first collected before the prediction is performed. This can for example be necessary if substructures need to be identified prior to predictions or that a parameter is to be measured for a whole object. The software may be Umbio Real-time developed by the company Umbio AB. Upon capturing new unidentified variation in the hyperspectral information content from the on-line measurements of the object, an alarm system or similar will emit a signal. In order to further characterize the new variation, the item is taken from the production line and analyzed with a new reference measurement. Based on the new reference measurement, the deviation is further analysed. The spectral signature and definition of the deviation, such as type, is integrated into the model system. The next time the deviation is observed in an On-line measurement it is identified from its definition. Over time, an ever more complete model system evolves from the use of the procedure.

Application of hyperspectral imaging for the analysis of food products such as cheese

The presented procedure for hyperspectral image analysis in real time, can to advantage be used for analysis of food such as cheese, bread, meat or other types of food. In the exemplified embodiment of the present patent application the food type is any variety of hard cheese.

In an exemplified embodiment of the procedure, hyperspectral image analysis is used to measure the properties of cheese. In the first step of the procedure a number of parameters of interest to measure and that can help to provide a better understanding on product quality and its observed variation, is identified and defined. Within the specific application such as analysis of hard cheese, the objects can be further divided into specific subtypes of hard cheese. Each type of cheese has a specific chemical composition that affects certain characteristics such as taste, smell, structure and others properties. An initial consideration can be to define the type(s) of cheese to be analyzed.

Further on it is determined whether these parameters can be defined in a clear manner and whether there is potential to measure them by means of

hyperspectral image analysis. Among interesting chemical parameters

(properties) to measure are moisture content, salt content, fat content and protein content. Structural parameters relating to the cheese structure such as the number of holes, the distribution of holes and the presence of cracks are also interesting to measure. Structural variations can be detected even if the object has a homogeneous chemical composition. This is possible through effect of light scattering. Sensory parameters such as chewiness, stickiness and hardness can be analyzed with presented procedure can also be. Even deviations from normal cheese not already included in the model structure can be identified. Such deviations can be for example the unwanted presence of mold, plastics, stains, metal, insects and similar on the cheese. Detection of not already modelled deviations and objects can be done by analysing statistical residuals of multivariate models based on the analysed objects. As an example will other objects than cheese be identified by the cheese model as deviations. Mold can with this procedure be detected before the mold is visible for the naked eye or can be identified by a standard vision system. Deviations can also include variation in the distribution (homogeneity) of the raw material of the cheese. An overarching goal of the procedure may be to on a global scale describe the above parameters. A first step would be to identify the type of cheese and the type of cheese that should be included as an overall parameter for the problem definition and the study design.

In the second step an experimental study design is specified, defining what measurements to be performed in the following steps of the procedure. The study design is created to ensure that the amount of information generated by the hyperspectral test measurements and reference measurements can be analyzed to create an applicable model system. Of importance is to determine the number of sample measurements that need to be carried out so that a statistically informative sample series can be analyzed.

In the third step of the procedure step, at least one sample measurement is done capturing at least one hyperspectral set of information, using the equipment schematically shown in Figure 4. The equipment, known as the UmBio Inspector, preferably includes a hyperspectral line scanner with integrated illumination with the hyperspectral data sets collected from the reflected radiation from the cheese. The equipment further includes at least one controller, at least one computer and at least one software.

The line scanner preferably uses so-called Push Broom technology, capturing data with for example a frequency of 100 Hz. It is conceivable that the sampling rate for the scanner varies widely in the context of the present patent

application. Preferably the camera (scanner) is collecting data in the wavelength range 400-2500 nm. During scanning, objects are placed on a conveyor or similar and transported towards the camera (scanner). When scanning with a line scanner (line cameras), the cheese is scanned line by line. During the scan collected raw data is collected and stored in a suitable media such as a hard disk or the like. In alternative embodiments, it is conceivable that the cheese and the camera is relatively moved through the cheese being stationary and the camera moving or that both the camera and the cheese are moving. The scan can also be done with technology, where both the cheese and the camera is stationary, such as described in the patent document US6166373.

After sample measurements the cheese is labelled in a suitable manner. For example, the cheese is packaged in a suitable container provided with

appropriate labelling such as a barcode.

In step four, reference measurement of the cheese is carried out in a lab or similar establishment. In the lab is comprehensive measurement and analysis carried out of the chemical, sensory and structural parameters as in step one deemed relevant to investigate for the cheese in question.

In step five, statistical analysis of the results from the hyperspectral

measurements and the results from the reference measurements is undertaken. In the analysis, there is a processing (merging) of data from the reference measurements (information quantities) and the spectral raw data (information quantity) from sample measurements. The merged information quantities are analysed by a specific analysis software such as Umbio Evince Image, developed by the applicant. The software prepares (models) at least one specific model or model system, preferably at least one multivariate model is developed and workflows are developed for the specific application.

In step six, referred to as Model System, a model system is created that can be used to predict the quality related parameters identified in the problem definition and study design. The application-specific model system includes at least one classification model, and preferably at least one quantification model for each specific application. Carefully designed model systems can then be applied for real time measurement of the parameters included in the

experimental design. The developed model system is transferred to a structured workflow that is used by an optimized calculation engine to perform real time predictions both at-line and on-line. The calculation engine used for

quantification and classification in real time is optimized for maximized CPU and memory utilization. The structure of the complex workflow system is designed to provide support for this functionality. This allows for the calculations to be parallelized by either dividing the calculations for individual models and / or partial information sets to different processors in order to obtain maximized computational efficiency. The processing of the amount of information is done sequentially or in parallel.

In step seven of the procedure, referred to as At-line prediction, the model system is used next to the production line or in similar setting to classify and quantify previously reference-measured parameters. Samples are taken from the production line or redirected to a parallel line (possibly miniaturized) and is measured by the hyper-spectral analysis equipment. In the case that the measured test samples contains information not represented in the model system, there is a continuous fine-tuning of the model system through

performing additional reference measurement and updating of statistical models by introducing the variation in the modelling procedure. The models are automatically updated to improve the prediction accuracy of the system.

The deviations may for example be due to seasonal variations in chemical properties of the raw material. The seasonal variations can for example be explained by cows eating different type of feed in summer than in winter, which can results in variations in the chemical composition of milk. The deviations may also include an item or similar that ended up on the cheese.

A unique feature of the present invention is that predictions are performed by at least one software including at least one algorithm instantly after data

acquisition by the camera. It is also unique with the present invention that the processing of the information (including for example spectral data) and that the calculations are performed with parallel processors. The parallel processing is achieved as an example through multicore processors and/or several processors. The software may be Umbio Real-time developed by the company Umbio AB. Also can metadata for batches, operators, production plants and other

information be collected.

In step eight, the procedure is implemented on-line in production. This step is very similar to step seven, but instead of carrying out predictions next to the production line using the UmBio Inspector system, the equipment is integrated directly over the actual production line. In this way the procedure is used as an applicable measuring device, directly monitoring the production process. The through the procedure measured properties can then be saved for comparison with metadata in the form of process parameters.

In the detailed description of the present invention, design details and parts of the procedure may be omitted if obvious to a person skilled in the field for which the procedure and the device are aimed for. Such obvious design details and parts of the procedure are included to such extent required for a proper function to be obtained for this procedure and device.

Although some preferred embodiments are described in more detail, some variations and modifications of the procedure and the device may be evident to professionals in the area of the invention. All such modifications and variations are considered to be within the framework of the subsequent patent claims. For example, the stated uses do not limit the scope of protection for the procedure in accordance with the present patent application, but the procedure can be used to identify deviations in a very large number of items and products in which the procedure is appropriate to use. It is also conceivable that information acquisition can be done using another for the purpose appropriate technology.

It is also conceivable that the measurements are capturing information from other wavelengths than the near infrared region, depending on fitness for purpose. Within the description of hyperspectral data collection, it is also conceivable that the data is acquired by varying the frequency of the source of illumination. It is comparable to a procedure using a source of constant character and that a sensor that measures several wavelengths in parallel is used. These two approaches both result in a hyperspectral data set which is what is of main importance for the described procedure. Another plausible option is that instead of measuring the absorbance or reflectance of light shining through an object, so- called transmittance.

Advantages with the invention

With the present invention, a number of advantages is achieved. The most obvious is that a comprehensive approach is created for creating a variety of application-specific procedures, tailored to the specific application, for the prediction of deviation in the chemical content and physical properties for application-specific objects. Another advantage is that the prediction is done in real time without the need for manual destructive tests. The invention, with the above characteristics, allows for faster, less resource intensive and in some cases measurements not previously possible to perform. Another advantage of the present invention is that the procedure combines imaging and spectroscopic analysis which provides a unique opportunity to identify structures from its chemical properties. An additional advantage of the present invention is that the statistical analysis is done with multivariate methods which have many desirable properties for the analysis of hyperspectral data, such as that the analysis can handle many correlated variables. An important advantage of the procedure is that it helps to migrate from methodology development and application in a lab environment to be used as a stable method in real time directly in a production process.