Current Microscopy Contributions to Advances in Science and Technology (A. Méndez-Vilas, Ed.)

Quantitative image analysis of food microstructure G. Impoco*,1, N. Fucà1, L. Tuminello1 and G. Licitra1,2 1 2

CoRFiLaC, S.P. 25 Km 5, 97100 Ragusa, Italy DISPA, University of Catania, Catania 95100, Italy

In this chapter, we highlight some of the common pitfalls in the use of image analysis. We show a few examples of how image analysis can be successfully used, possibly in conjunction with sensory analysis and visual observation, to discover interesting properties about food products. Special attention is granted to segmentation, a ubiquitous image processing step, propaedeutic to most quantitative analysis methods. In fact, this is the most critical step of all the image analysis chain and is potentially disastrous for the quality of subsequent analyses. Unfortunately, this step is the one which people tend to run more carelessly. We discuss the need to carefully choose segmentation algorithms as well as to thoroughly analyse the outcome, before any quantitative analysis is made. Finally, a few good practices are discussed. Keywords image analysis; food microstructure; quantitative analysis.

1. Introduction Microstructure is one of the major controlling factors of texture (firmness, softness, cohesiveness, rubberiness, elasticity, pastiness, crumbliness) and functional properties of food products, also affecting physicochemical, transport and nutritional properties. As texture and functional properties are significant quality requirements for consumers, microstructure analysis plays an important role in the quality evaluation of dairy products: better quality usually brings higher revenues and consumer satisfaction Microscopists experience and accuracy are fundamental for a proper micrographs analysis. However, one of the most common pitfalls in microscopy is technicians’ unintentional tendency to confirm their own hypothesis in the micrograph observation. Visual observation is often biased by subjective evaluation of image content [1]. Since human judgement lacks objectivity, human natural imaging system is not completely trusted for quantitative estimations. Automatic measurements driven by image analysis are quantitative, objective and repeatable. It provides for quantitative information from digital micrographs, in the form of numerical data [1]. Unfortunately, the quality of computer analysis of digital images is far from being comparable to human visual capabilities. Human visual system has evolved a natural capability to recognize and isolate interest object features from the background. Humans are thus pretty clever to discriminate image features. This does not hold true for machine vision systems. On the other hand, computer measurements are objective and deterministic, whereas human natural visual system is not capable of taking reliable measurements. In recent years, automatic analysis of digital images has become popular for quantitative measurements of food microscopic features (micro-features) and has been paired up with many different imaging technologies, such as electron, confocal, and light microscopy. In this chapter, technical aspects of image analysis steps will be addressed. Segmentation will be discussed and common bad practices and wrong beliefs will be highlighted. Next, quantification of image features will be treated. Finally, some examples of applications of image analysis on food products will be reported.

2. Microscopy and image analysis Successful image analysis requires input micrographs to be conveniently processed before any measurement can be taken. Accurate detection of structural features is an important pre-requisite for reliable quantitative analysis of food morphology [10]. The measurements computed on binarised images are deeply influenced by the accuracy of feature detection, especially when shape descriptors are involved. Image analysis is thus a two-step process. First, the image is processed in order to mark relevant features (segmentation). Then, a series of measurements is taken on marked image regions. Both aspects are discussed in the following sections. 2.1 Image processing and segmentation Feature detection is mainly carried out by image segmentation, i.e., a partitioning of the image domain into nonoverlapping regions. The goal of a successful segmentation is to accurately mark one region for each interest feature in the micrograph. Each region should closely cover the image area occupied by the corresponding feature. The most common segmentation scheme used in analysis of food micrographs is thresholding: a scalar value is used to partition a greyscale image into two regions containing lighter and darker colours, respectively. A two-level labelling is

© 2012 FORMATEX

903

Current Microscopy Contributions to Advances in Science and Technology (A. Méndez-Vilas, Ed.)

obtained, partitioning the input image into background and foreground (Figure 1b). This approach is clearly naïve, whereas lots of more sophisticated and effective methods exist. However, it is fast and intuitive for the operator and it often gives acceptable results.

a)

b)

Fig. 1 A scanning electron microscope (SEM) image of Ragusano cheese. The binarised image is obtained by manual thresholding.

a)

b))

Fig. 2 A detail of the micrograph in Fig. 1. The crack in the protein matrix is accurately detected and marked.

a)

b)

Fig. 3 A detail of the micrograph in Fig. 1. Most pores are missed.

a)

b)

c)

Fig. 4 A detail of the micrograph in Fig. 1. The binary image (b) is very noisy, whereas if noise removal (median filter) is applied before thresholding the resulting image shows less spurious spots.

As an example, consider the micrograph in Fig. 1a and the result of manual thresholding in Fig. 1b. Three close-ups are shown in Figures 2-4. The crack in the protein matrix is accurately marked in the region shown in Fig. 2, whereas thresholding fails to detect pores in Fig. 3. Spurious tiny structures are also erroneously segmented in Fig. 4, due to uneven matrix surface and possibly to acquisition noise. It is worth analysing these results. The first case shows a dark crack through a brighter matrix. The brightness of these two image regions is consistently different. Thus, an intensity

© 2012 FORMATEX

904

Current Microscopy Contributions to Advances in Science and Technology (A. Méndez-Vilas, Ed.)

level, T, can be easily found so that matrix pixels are brighter than T and crack pixels are darker than T. This is a perfect input image for thresholding algorithms. This does not hold true for the close-up in Fig. 3. Here, most pores are not significantly lighter or darker than the flat matrix. There is a smooth transition between the matrix and the bottom of pores. This violates an important assumption of thresholding algorithms on the distribution of pixel values: foreground and background should have well-separated brightness pixel values, i.e., one region should be dark and the other bright. Since the close-up in Fig. 3 does not conform to this assumption, there is no means to detect these pores by using thresholding alone. These considerations apply for Fig. 4, as well. Surface roughness and acquisition noise cause small white spots to appear on the binarised image. Although of no significance for feature detection, they are sufficiently darker than the great part of the matrix to be marked as pores. The message, here, is that one often has wrong expectations on what an image processing algorithm does. In the example reported, the problem is simply stated: distinguish compact matrix from pores. This task is straightforward for a human operator but it is very hard for a computer program. In fact, it turns out to be one of the most difficult unsolved problems in computer vision. There is a gap between this plain request and the series of simple mathematical operations onto image pixels that should implement the request. In the case of thresholding, the problem is solved by assuming that the image shows only two interest regions (background and foreground) with considerably different brightness characteristics. This is not quite the same goal the user had in mind. In order not to miss pores in Fig. 3, a segmentation algorithm should also mark circular brightness gradients as pores. A certain amount of domain knowledge could be used to handle these situations, either built in the algorithm itself or as additional information. Referring to the definition above, thresholding does not allow any of these. As for Fig. 4, white spots can be avoided only if spatial information is used. Namely, the brightness of neighbouring pixels should be analysed to decide if a pixel belongs to foreground or background. We name context-aware those algorithms that use neighbourhoods. Clearly, thresholding is not one of them. Understanding what kind of processing an algorithm does is thus essential to choose the right tool for one's needs and to foresee results. Basically, there is no general-purpose computer vision algorithm capable of solving all segmentation tasks. Hence, one should learn how to interpret segmentation errors and find more adapt algorithms for the task at hand. In short, try and understand what to expect from a segmentation software tool. Now, the question is how good my computer vision algorithm is to solve my segmentation problem. Is my segmentation good enough for successful (or even meaningful) quantitative analysis? Well, hold on tight: the harsh truth is that quantitative analysis on bad segmentations is meaningless. Actually, rubbish. Being able to somehow evaluate segmentation results is thus a key skill to run high quality image analysis experiments. Unfortunately, there is no simple answer to the question whether segmentation is good enough for further analysis. The quality of a segmentation is strongly related to the application, i.e., to the type of analyses that are to be performed on the segmented micrograph. Let us make this point clear with a simple example. Suppose we want to compute porosity and the number of pores in the micrograph of Fig. 1. The segmentation in Fig. 1b shows a relevant number of small spots, mainly due to acquisition noise and matrix surface roughness. The total area on the micrograph surface occupied by those spots is negligible with respect to the area of pores. Hence, the computation of porosity is not affected by this segmentation error. Thus, our segmentation is probably good enough for this task. On the other hand, these spots artificially increase the number of pores and constitute a non-negligible fraction of the total count of pores. This measurement is thus useless because it is affected by too much high an error. So, the quality of the segmentation in Fig. 1b is too low to reliably estimate the number of pores.

a)

b)

c)

Fig. 5 A confocal light scanning microscope (CLSM) image of a milk sample where fat is stained. In the binary image (b) fat globules are not distinguishable. The thresholding in (c) is of better quality but does not detect several globules.

As a further example, consider the micrograph in Fig. 5. A confocal light scanning microscope (CLSM) image of a milk sample is shown where fat is stained. What measurements can we use for image analysis? Clearly, the binarisation

© 2012 FORMATEX

905

Current Microscopy Contributions to Advances in Science and Technology (A. Méndez-Vilas, Ed.)

in Fig. 5b does not allow to compute the count of pores, since most of them appear as if they where fused into larger structures. Fig. 5c is a better choice, even if most pores are not detected. Hence, neither of these segmentations allows to compute a reliable estimate of the number of pores. The latter, however, would probably give acceptable results for applications that can tolerate relevant estimation errors. On the other hand, the segmented micrograph in Fig. 5c is not suitable for porosity, since most pores are not detected. Working with the binarisation in Fig. 5b would probably output a slightly over-estimated, but still acceptable, measurement. As for shape measures, let us only discuss roundness [10] with the help of Fig. 5. Clearly, it does not make any sense to compute roundness on the aggregate at the top left of the micrograph if Fig. 5b. Fig. 5c is much more amenable to the computation of any shape measure since most globules can be clearly distinguished. However, there are large disparities between the real shape of some globules and the shape of detected features. Although this would probably not influence porosity, shape features such as roundness might be strongly affected. If there are several such cases, it is better not to trust shape measurements. Conversely, if those cases concern a small fraction of the total number of features, reliable measurements can be obtained if outliers are correctly dealt with. These examples should make clear that care must be taken when selecting image analysis tools for an application. There is a plethora of segmentation algorithms. Unfortunately, most of them are out of reach for the majority of users. Either they are not available in (commercial as well as free) image analysis software packages, or they require fine tuning of parameters whose meaning is obscure for those who do not know the technicalities of the algorithms used. Far from purporting to attempt a coherent taxonomy of segmentation algorithms, here we briefly discuss just a few possible methods: thresholding, region-merging, trainable. A simple introductory guide can be found in [5]. Enough has been said about thresholding in the preceding discussion. We have learned that thresholding algorithms lack two important features, namely context-awareness and domain knowledge. Region merging and trainable methods are meant to address these issues, respectively. Region merging methods start by labelling each pixel as a region. Then, similar regions are iteratively merged into larger regions until no further merging is possible. This simple scheme simultaneously enforces similarity and adjacency, i.e., tends to output contiguous regions of pixels with homogeneous properties. The homogeneity criterion adopted is the major distinguishing feature among these methods. It should be clear that, using this scheme the label assigned to each pixel indirectly depends on that of neighbouring pixels. Another way to exploit neighbourhood information is to pre-process images using spatial filters, such as Gaussian blur, median, or more specific low-pass and high-pass filters [11]. The influence of neighbourhood in segmentation is further developed in iterative re-labelling methods. At first, a segmentation is produced with any algorithm. Then, at each iteration pixels are re-labelled using the labelling of pixel neighbourhoods produced in previous iteration. The iterative process stops when the segmentation does not undergo any significant changes. Clearly, here neighbourhoods are explicitly taken into account in the iterative process. Region merging algorithms often fail with images where regions show complex image textures such as the group of pores at the top of the micrograph in Fig. 1. Basically, these textures violate any common homogeneity criterion, which is the basis of region merging strategies. In those cases, domain knowledge can be used, either by directly encoding this knowledge into the very steps of the algorithm or, more commonly by exploiting learn-classify statistical tools. The statistical distribution of pixel features is automatically learned by the algorithm for each interest class (e.g., in Fig. 1 it could be matrix and pores). These statistical distributions are used to label pixels of unseen images. Learning can be unsupervised, i.e., the algorithm is equipped with mechanisms that try to develop an autonomous learning process, or supervised, where an expert user trains the statistical model of the algorithm providing manually-labelled images. Texture segmentation has been thoroughly studied in the literature. An introductory work on texture classification can be found in [12], whereas [10] and [5] give a basic introduction. More generally, learning techniques have been largely employed in the analysis of macroscopic food images. A good survey on learning image analysis methods for quality control can be found in [3]. Image analysis of food microscopy has attracted less interest, so far. We know of no published review about the microscopic world. A recent application of learning methods to the segmentation of SEM micrographs is in [9]. Most images can be conveniently pre-processed before segmentation. For example, there are plenty of pre-processing algorithms to remove noise. The simplest one computes the median of a pixel neighbourhood. Applying this method to the micrograph before thresholding (e.g., see the close-up in Fig. 4a) reduces the number of spots caused by noise or surface roughness (Fig. 4c). Other pre-processing techniques address background removal (for images showing large illumination gradients or light spots), contrast enhancement, filtering, and so on. Pre-processing often solves most segmentation problems, especially coupled with simple segmentation methods, such as thresholding. An analysis of pro-processing techniques is beyond the scope of this chapter. Useful solutions are clearly discussed in [5]. 2.2

Feature extraction

Features that have been marked during segmentation can be used to collect measurements for quantitative image analysis. In principle, features could be gathered from raw images. That is, a series of measurements could be computed on the original image, without running any pre-processing or segmentation algorithm. This method, however, presents

© 2012 FORMATEX

906

Current Microscopy Contributions to Advances in Science and Technology (A. Méndez-Vilas, Ed.)

serious challenges. We are aware of just a few successful applications to food image analysis. One such application will be discussed in Section 4.

a)

b)

Fig. 6 The scanning electron microscope (SEM) image of Fig. 1 where whey pockets (green), fat globules (red) and protein matrix (black) are marked: a) labelling superimposed to the micrograph; b) labelling alone.

Let us assume, then, that each feature is a collection of pixels with the same label in the segmented image (see Fig. 6). Pixels are often assumed to be contiguous, i.e. a feature is a (nearly) compact block of pixels. In Fig. 6, whey, fat and protein have been marked on the sample micrograph of Fig. 1. Each whey pocket or fat globule can be considered as a feature. With the discussion about segmentation accuracy in mind, the marked features cannot be assumed to be correctly shaped in comparison to what a human operator would mark. Segmentation errors can affect contiguous pixels in such a way as to modify size and shape of image features. Clearly, one should not rely on measurements computed on highly deformed features. Care should be taken when measurements are collected and analysed. Several different measures can be computed on image features [10]: size and size distribution, colour and brightness, spatial distribution, density distribution and gradients, shape, and so on. A few of these measures are relatively robust against segmentation errors. Conversely, measures such as shape descriptors are highly sensitive to shape deformations due to misclassified pixels. These descriptors are also sensitive to coarse quantisation, i.e., a feature is represented by too few pixels for its shape to be recognisable. As an example, let us assume we are interested in discovering if features in our image have a preferred orientation. As a measure for orientation, we compute directionality [10] as the angle between the x-axis and the main axis of the feature, corresponding to the segment that joins the two farthest pixels in the feature boundary. Now, let us study the effects of this measure on small features. Let us consider a 2x2 square of pixels. One would probably argue that this pixel region should be treated as a small circular feature. Hence, it should have no preferred direction. However, pixel couples do not have the same distance. There is 1-pixel distance between rows or columns of pixels, whereas the distance between pixels along diagonals is about 1.414 pixels. As a consequence, a preferred direction of 45° is assigned to this feature. Our definition of directionality has a flaw. Can we find a better definition? Probably, but no definition is fit for all seasons. One solution is to fit an ellipse to each feature and take the orientation of the main axis. Examples can be easily produced where a single misclassified pixel causes a completely wrong orientation to be assigned to a feature. Clearly, large feature are less affected by a few misclassified pixels. The most successful solution to this problem is common sense. Small features often have little impact on bulk quantitative analysis. Hence, they can be safely dropped. There is also a physical argument to support this choice. Optical lenses are not perfect and probably they will never be. Every lens has a characteristic Point Spread Function (PSF, see [5]). Namely, a number photons running along a line are spread by a small angle when going through the lens. This causes a small light cone to hit a small pixel region. As a consequence, the colour of each pixel is the result of several contributions from light directed to neighbouring pixels. This is one of the main causes of segmentation errors, particularly along object boundaries. Although electron micrographs exploit pretty different physical principles, they are built in such a way as to follow the same imaging principle. Thus, this discussion applies to electronic imaging, as well. It should be clear that features represented by a few pixels are particularly prone to errors. It is common practice to drop features smaller than 5 to 10 pixels. 2.3 Quantitative analysis A number of measurements are usually collected for each feature. From a mathematical point of view, each feature can be represented by a k-dimensional vector, where k is the number of measures. In a common experimental setting,

© 2012 FORMATEX

907

Current Microscopy Contributions to Advances in Science and Technology (A. Méndez-Vilas, Ed.)

hundreds of features might be detected for each micrograph. This amounts to a large dataset, especially when dealing with several images. Qualitative analysis is easily performed thanks to the natural synthetic ability of human mind, without focussing on, or even noticing every single detail. Hence, humans are little subject to overwhelming complexity. On the other hand, computers must analyse details. Thus, a mathematical tool is needed to get an overall view of the data. Statistical analysis is probably the most appropriate method to extract synthetic indicators from data. It is common practice to start observing a graphical representation of data. Histograms give useful hints on the distribution of measurements, highlighting trends, modality and so on. Scatter plots can be used to detect possible outliers that may arise from segmentation errors. They are thus a basis to perform a deeper statistical analysis and to choose the most appropriate tools. As an example, let us suppose to compute roundness of fat globules in the CLSM micrograph in Fig. 5. What can we learn from outlier analysis? We expect that fat globules are approximately round, i.e., they have a roundness value close to 1 [10]. Low roundness values indicate outliers. Irregularly shaped features can be due to segmentation errors or to a fat globules aggregation, as in Fig. 5. Thus, outlier analysis on globule roundness could spot fat aggregates in our sample. Graphical representations can be used to interpret results, as well. Principal Component Analysis (PCA) can be employed to reduce the dimensionality of feature vectors. In image analysis, PCA can be conveniently exploited to detect the measures that better describe features in the sample. This observation suggests a working scheme. First, a subset of all micrographs is picked and a large set of measures is computed on each sample. Using PCA, only the most significant measures are retained. This reduced set of measures is then used to collect useful information on all images.

a)

b)

c)

d)

Fig. 7 Scanning electron microscope (SEM) images of Wagashi cheese before (a) and after (c) boiling. Small and large pores are marked in blue and red, respectively.

Quantitative analysis is not confined to measurements of feature properties. Interesting facts can be observed by analysing a group of features as a whole. Let us consider the SEM micrographs of Wagashi cheese in Fig. 7. Fig. 7a shows a cheese sample before boiling, whereas, after boiling the cheese microstructure appears as in Fig. 7c. Here, we are interested in grouping features in two sets: small and large pores. A simple thresholding method was used to detect

© 2012 FORMATEX

908

Current Microscopy Contributions to Advances in Science and Technology (A. Méndez-Vilas, Ed.)

features. Then, an operator selected a threshold to partition small and large features, by means of an interactive software tool. The ratio of small and large pores was analysed before (Fig 7a) and after (Fig. 7c) boiling. After boiling, protein matrix is more compact (area occupied varies from 73% to 67%) and porous (from 27% to 33%). More interestingly, small pores occupy slightly less area (from 15% to 11%), whereas large pores are larger and more numerous (area varies from 12% area to 22%). An interesting property could be demonstrated, by simply grouping pores according to their area.

3. Best practices In the preceding discussion, we have highlighted that results from the image analysis chain cannot be used uncritically. A few examples have been reported to support this claim. In section 2.1, lots of space has been dedicated to demonstrate how critical segmentation errors can invalidate measurements and prevent from further analysis. We have suggested the need for suitable image pre-processing and noise removal, prior to image segmentation. In section 2.2, feature selection was briefly presented and possible removal of small features was discussed, as well as the opportunity to handle outliers. All these examples should suggest a careful scientific attitude towards image analysis. Let us further stress this point. There are cases when low quality data are obtained, in spite of all efforts. Experimenters should refrain from extracting information from those data, since any interpretation of the results would be flawed and would possess poor scientific value. This was recognised in [7]. Here, CLSM was used to analyse the microstructure of fat globules in Cheddar cheese. A three-dimensional volume was imaged for each sample and size and shape of globules were analysed using simple measures. The volume was obtained as a stack of images 40μm thick. A large portion of globules lying at boundaries was cropped from the volume. Computing shape measurements on these boundary globules would have strongly affected results. Correctly, the authors do not venture to compute measurements of any kind on these globules. The lesson here is that shape descriptors should not be computed on badly-shaped features, such as the boundary globules in the example above. More generally, bad data should be filtered in order not to compromise the outcome of analyses and the subsequent arguments. On the other hand, a widespread bad practice is to threshold raw microscope images using an interactive tool. Nothing bad with it, if binarised images would be carefully analysed and processed in order to filter out badly-shaped as well as small features and spots. Unfortunately, this does not always happen. As a consequence, the number of features can be overestimated and shape measurements are often useless. Clearly, the scientific value of conclusions originated from such data is questionable. In the context of image analysis of cheese microstructure, Impoco et al. [8] proposed a simple algorithm to binarise SEM micrographs of Ragusano cheese. After pre-processing the input image to remove Gaussian as well as speckle noise, morphological operators are applied before and after segmentation. Finally, small features are not included in the dataset used for quantitative analysis. Despite the extra effort to obtain good quality binarised images, segmentation error can be as high as to involve 20% of pixels [9]. Unpublished results on simple thresholding, without any pre- or post-processing, showed that more than 50% of pixels had been misclassified. This clearly prevents any further analysis to give meaningful results. Once more, only high-quality segmented images should be used and suitable segmentation algorithms should be carefully selected for each application. Transmission electron microscope (TEM) was employed in [13] to study fat-free Feta cheese made from ultrafiltered milk. The effects of varying coagulation conditions on Feta microstructure were analysed on TEM micrographs. Images were thresholded and a skeleton was computed to analyse the compactness of the protein phase. It was expected that a more branching structure of the skeleton would reflect a coarser protein network. Before computing the skeleton, features with area less than 5 pixels were deemed irrelevant and dropped. This is probably a good choice, since they would have probably interfered with the computation of the skeleton structure. Moreover, such small structures make little contribution to the analysis of the protein matrix. The authors delete these features by using a morphological opening operator. This is successful in doing its job, but not quite the same as computing connected components and removing those small components. Basically, morphological operators are applied to the whole image, thus involving every feature. In detail, while the image is cleaned up from small features, the shape of all the remaining features changes. This is not desirable in many applications, especially those involving the computation of shape descriptors. For the computation of skeleton, as in [13], using morphology is probably safe. Its use when area is estimated is questionable.

4. Successful examples In the first two sections, we pointed out the need for a careful use of image analysis tools and, more important, a rigorous evaluation of the outcome of image segmentation and further analyses. In this section, we review a few published papers where image analysis was successfully employed.

© 2012 FORMATEX

909

Current Microscopy Contributions to Advances in Science and Technology (A. Méndez-Vilas, Ed.)

Christiansen et al. [2] compared microstructure and texture properties of salad dressings, produced according to a fractional factorial design where design variables were protein type, protein level, oil level, pH, addition of salt and sucrose, and processing temperature. Dressings were produced by high-pressure, and were emulsified and stabilized with different whey protein types. The dressings’ texture was characterized by sensory analysis. Image analysis of the dressings’ microstructure was run on SEM micrographs. Images were pre-processed by automatically adjusting contrast. Then, variations in the microstructure of emulsions were measured by two different descriptors, both related to contrast. One of them averages the absolute differences in image brightness between pixels at fixed distances, over the image. A spectrum is generated for different pixel distances. This is a robust measure that can capture image complexity and characterise image texture. Clearly, this measure is global, i.e., it involves all pixels in the image. Conversely, the other descriptor measures local contrast using a box-counting method. Basically, brightness differences are averaged over a pixel neighbourhood. The higher the value, the higher the local contrast. These descriptors were analysed using PCA and score plots to detect outliers. Then, descriptors were correlated each other and to sensory data. After outlier removal, images could be grouped into two sets, one where feature vectors were uniformly distributed, the other where two clusters could be clearly distinguished. Statistical models generated from each of the two measures showed a good prediction capability on some design variables. In this example, the proposed measures are robust to image noise, but not to poor image contrast (outlier images had often poor contrast). These measures do not suffer from interpretation errors, since they do not rely on segmentation. A good link between different types of analysis, also involving image analysis, has been demonstrated in this paper. In Fucà et al. [4], structural differences have been quantitatively measured among 6 full fat milk types subject to different industrial treatments and stored at two different temperatures. CLSM images were acquired from milk samples where fat was stained (Fig. 5a). Visual evaluation criteria, such as density and distribution, were established and used for visual evaluation by a panel of 6 trained people. This evaluation led to group milk types into two sets with different treatment characteristics. Namely, weakly- and strongly-treated milk. The same images were fed to a computer program that used different evaluation criteria. Those criteria were encoded as mathematical descriptors, mainly based on geometrical attributes of globules or of group of globules. The algorithm used by the program automatically partitioned images into the same two sets as human did. This method was proven to be trustful as to substitute human judgement. The algorithm had two stages. First, input images were converted to greylevel and noise was removed by a simple Gaussian filter. A well-known automatic thresholding algorithm was used to segment fat globules. Then, connected components related to small globules were removed. At this point, most images were correctly binarised. Some of them, though, showed large features made up of smaller globules too close to be distinguished. The watershed algorithm is usually employed to separate these blocks. Here, the authors decided to devise mathematical descriptors that explicitly accounted for globule aggregations and fusions, trying to distinguish between them. This example shows that complex and successful image analysis systems can be built from simple pieces, provided that care is taken to evaluate the outcome of each computational component. If one component at a stage is not successful on some inputs, another component at some other stage must be able to handle the problems that can arise. Gonzalez et al. [6] studied tissue integrity in onion samples subject to different high-pressure preservation methods. Samples were stained with red dye and then observed under the optical microscope. The staining of cell vacuoles allowed for the differentiation of viable from inviable cells in onion parenchyma tissue. Living cells on each image were counted using methods based on two different colour components: saturation in the HSB colour space and the green component in the RGB space. An edge image was computed from the saturation image component using a variance filter and then subtracted from the saturation component itself. Doing this way, cell contours were wiped off leaving well-separated features. Then, images were thresholded and the cells dilated to restore a more realistic size. Median filter was used to reduce speckle noise. The same procedure was used for the green component. Finally, small cells were filtered out on the basis of their area. The remaining cells were counted. These count estimates were compared to the conventional manual cell count. Statistical analysis showed that the number cells counted by this method was close to the real count when saturation was used. The green component did not discriminate sufficiently well viable from inviable cells. Although very simple, the algorithm proposed in [6] was carefully designed. Noise was removed as well as disturbing features, such as cell contours. Then, results were analysed and compared with a golden standard, i.e., manual cell count. Hence, the authors were able to demonstrate that their method, applied to the saturation component, is affordable enough to replace manual count.

5. Conclusions Correctly handling image analysis tools is essential to the success of quantitative analysis. Good practice and thorough scrutiny of the outcome of each image analysis step make a solid ground for reasoning about data and to draw scientifically valid conclusions. On the other hand, most people underestimate the importance of advisedly managing image manipulation and analysis algorithms. This is particularly true for image segmentation, probably the most critical step of the whole process, since it involves taking crisp decisions about feature shape and size. This chapter collected some examples, highlighting strengths and weaknesses of various approaches to practical research problems in food microscopy. The point of view of imaging experts is strongly present all over the text. The

© 2012 FORMATEX

910

Current Microscopy Contributions to Advances in Science and Technology (A. Méndez-Vilas, Ed.)

objective was to give hints on why and how image analysis should be carefully designed, as well as any other experimental setting. Image analysis is an experimental method itself. Acknowledgements

This work was funded by the Assessorato Agricoltura e Foreste della Regione Siciliana, Palermo, Italy.

References [1] Aguilera J. M., and Stanley D.W. Microstructural principles of food processing and engineering. (2nd ed.). Aspen. 1999. [2] Christiansen K.F., Krekling, T., Kohler, A., Vegarud, G., Langsrud, T., Egelandsdal, B. Microstructure and sensory properties of high pressure processed dressings stabilized by different whey proteins. Food Hydrocolloids, 20, 2006, pp. 650-662. [3] Du C.-J., Sun D.-W. Learning techniques used in computer vision for food quality evaluation: a review. Journal of food engineering, 72, 2006, pp. 39-55. [4] Fucà N., Pasta C., Impoco G., Caccamo M., Licitra G. Microstructural properties of milk fat globules, (to appear in) International Dairy Journal, 2012, doi:10.1016/j.idairyj.2012.03.010. Available at: http://www.sciencedirect.com/science/article/pii/S0958694612000775. [5] Gonzalez, R. C. and Woods, R. E. Digital Image Processing, 3rd ed. Prentice Hall, Upper Saddle River, NJ. 2008. [6] Gonzalez, M.E., Jernstedt, J.A., Slaughter, D.C., and Barrett, D.M. Microscopic Quantification of Cell Integrity in Raw and Processed Onion Parenchyma Cells. Journal of Food Science, Vol. 75, Nr. 7. 2010. [7] Gunasekaran S., Ding K. Three-dimensional characteristics of fat globules in Cheddar cheese. Journal of Dairy Science, Vol. 82 (9) pp. 1890-1896. 1999. [8] Impoco G., Carrato S., Tuminello L., Caccamo M., Licitra G. Quantitative analysis of cheese microstructure using SEM imagery. Minisymposium: Image Analysis Methods for Industrial Application, Communications of the SIMAI 2006 congress, Baia Samuele, Italy. [9] Impoco G., Tuminello L., Fucà N., Caccamo M., Licitra G. Segmentation of structural features in cheese micrographs using pixel statistics, Computers and electronics in agriculture, 79 (2), pp. 199-206. 2011. [10] Russ J.C. Image Analysis of Food Microstructure. CRC Press. 2004. [11] Varma M. and Zisserman A. Classifying images of materials: Achieving viewpoint and illumination independence. In Proceedings of the 7th European Conference on Computer Vision, Copenhagen, Denmark, Vol. 3, pp. 255-271. SpringerVerlag. 2002. [12] Varma M. Statistical approaches to texture classification. Ph.D. thesis, University of Oxford. 2004. [13] Wium H., Pedersen P.S., Qvist K.B. Effect of coagulation conditions on the microstructure and the large deformation properties of fat-free Feta cheese made from ultrafiltered milk. Food Hydrocolloids, 17, pp. 287-296. 2003.

© 2012 FORMATEX

911