A new region-primitive method for classification of colour meat image texture based on size, orientation, and contrast
Introduction
In the food industry, texture is among the most important information in the analysis of the surface of objects in images (Amadasun & King, 1989), and has already played a significant part in the evaluation of meat qualities using computer vision (Zheng, Sun, & Zheng, 2006). However, the concept of texture in images, which is generally referred to as the born-in properties of images such as fineness, coarseness, smoothness, and graininess (Li, Tan, & Shatadal, 2001), is totally different from the one used in food products, which is a sensory property for describing the mouthfeel characteristics of food products (Chrystall, 1994). The preliminary work on image texture can be found as early as 1950s when an autocorrelation function was first proposed to characterise image texture (Kaizer, 1955). After years of study, however, strictly scientific definitions for texture are still not possible probably due to the infinite diversity of texture pattern, thus research in image texture becomes much more difficult, and so far there is not yet an ideal texture characterisation method available (Patel, Davies, & Hannah, 1996). Nevertheless, some unshaped ideas on how texture could be characterised were developed in two different aspects. In the first one, texture is generally referred to as the dependency between pixels and its neighbouring pixels in images; while in the second, texture is defined as the variation of the intensity values across pixels (Haralick, 1979). Meanwhile, research on image texture characterisation is still continuing mainly based on two approaches, i.e. statistical texture and structural texture. Furthermore, signal theoretic approaches such as wavelet transform (Mallat, 1989) and Gabor filter (Daugman, 1985) also played a significant part in texture analysis in recent years.
Statistical texture tries to extract texture features from higher order of the histogram of an image. Traditional statistical methods include co-occurrence matrix (Haralick et al., 1973, Zucker and Terzopoulos, 1980), run-length matrix (Galloway, 1975), and neighbouring dependence matrix (Sun & Wee, 1982). More recently, researchers concentrated on another two statistical technologies, i.e. Markov random fields (Wang & Liu, 1999) and autoregression models (Pietikäinen et al., 2000, Sarkar et al., 1997). The primary problem of statistical textures is their poor performance in dealing with macro texture patterns, yet this problem can be overcome by structural texture (Amadasun & King, 1989). Structural methods evaluate texture based on some deterministic primitives or textural elements that are found to occur iteratively in a certain arrangement or placement rules (Starovoitov, Jeong, & Park, 1998). Although, structural methods are found to be applicable to fabric and textile manufacturing and quality control in which texture patterns are more or less regular, they are not qualified for analysing texture of meat images because the textural patterns of meat images are usually complicated and very irregular, whereas the primitives that are proposed for description are mostly inerratic (Bodnarova, Bennamoun, & Kubik, 2000). Therefore, it is important to find some primitives with flexible and irregular shape so that the structural primitives are capable of describing irregular texture patterns in meat images.
Consequently, the objective of the current work was to develop a new structural texture extraction method, called region-primitive, and to demonstrate the performance of the proposed method in texture discrimination of colour meat images. In the region-primitive method, region-growing was used to construct structural primitives whose shape are changeable, corresponding to various texture patterns in meat images. Image features for texture pattern analysis were finally extracted from the primitives by consideration of three elementary factors in human perception of texture, i.e. size, orientation, and contrast.
Section snippets
Texture primitives
A region-growing method, by which neighbouring pixels are examined one at a time and added to the growing region if the pixel is found to be similar enough to the growing region, was proposed to segment images into small primitives (Russ, 1999). In the current work, the region growing procedure was seeded at the entry of the images, i.e. the top-left pixel in the images, and the construction procedure of the primitives is described as the following steps:
- 1.
Pixels from top to bottom and from left
Texture features
Once the image is segmented into a series of primitives, a two-layer structure can obviously be observed: the first layer is specified by the local properties within the primitives, while the second layer can be characterised from the interrelationship between these primitives, in other words, the organization of these primitives (Haralick, 1979). Vision research related to the human perception of texture reveals that size, orientation, and contrast are the elementary factors in texture
Experiments
After a total of eight texture features were obtained from the primitives, their ability in texture pattern recognition was evaluated in a classification experiment and classification results were compared with those using another group of textural features from the method of run-length matrix (Galloway, 1975).
Primitives construction results
Fig. 4 shows an image of each type of meat together with its region primitives constructed by the region-growing method. As illustrated the texture pattern in the meat images are quite irregular and appear difficult to be characterised by the traditional texture primitives, which have fixable size and shape. Instead, the region-growing method produces regions formed by similar neighbouring pixels, which seems suitable for describing the intensity variation in these images and thus for
Conclusions
Textural region-primitives were constructed using region-growing method based on an automatically calculated threshold from images. Eight texture features regarding to factors of human perception of texture, i.e. size, orientation, and contrast, were obtained from the region primitives. These texture features were then employed to discriminate the texture pattern of a set of meat images with the non-parametric discrimination function. Results were compared with those using features from the
Acknowledgements
The authors wish to acknowledge the financial assistance of the Food Institutional Research Measure, administered by the Irish Department of Agriculture and Food, Dublin.
References (31)
- et al.
Image texture analysis: methods and comparisons
Chemometrics and Intelligent Laboratory Systems
(2004) Texture analysis using grey level run lengths
Computer Vision, Graphics, and Image Processing
(1975)- et al.
Machine vision based quality evaluation of Iyokan orange fruit using neural networks
Computers and Electronics in Agriculture
(2000) - et al.
Classification of tough and tender beef by image texture analysis
Meat Science
(2001) - et al.
A review on image segmentation techniques
Pattern Recognition
(1993) Color texture classification by integrative co-occurrence matrices
Pattern Recognition
(2004)- et al.
Fuzzy segmented image coding using orthonormal bases and derivative chain coding
Pattern Recognition
(1999) - et al.
Prediction of sensory texture quality attributes of cooked potatoes by NMR-imaging (MRI) of raw potatoes in combination with different image analysis methods
Journal of Food Engineering
(2004) - et al.
A region growing and merging algorithm to color segmentation
Pattern Recognition
(1997) - et al.
Texture classification using multi-resolution Markov random field models
Pattern Recognition Letter
(1999)
Gentropy: evolving 2D textures
Computers and Graphics
Recent applications of image texture for evaluation of food qualities – a review
Trends in Food Science and Technology
Region growing: childhood and adolescence
Computer Graphics and Image Processing
Finding structure in co-occurrence matrices for texture analysis
Computer Graphics and Image Processing
Textural features corresponding to textural properties
IEEE Transaction on Systems, Man and Cybernetics
Cited by (16)
The application of computer vision systems in meat science and industry – A review
2022, Meat ScienceCitation Excerpt :In the analysis of image texture, statistical methods, mathematical (fractals), transformational (Fourier, wavelet), and structural models are used. The texture characteristics of images of meat samples can be determined both in black and white and color images (Arsalane et al., 2019; Zheng, Sun, & Zheng, 2007). Information on size, shape, color, and texture obtained from images are subject to classification, which may be conducted based on various methods, including statistics, neural networks and fuzzy logic (Jackman, Sun, Du, & Allen, 2009).
Study on wear state evaluation of friction stir welding tools based on image of surface topography
2021, Measurement: Journal of the International Measurement ConfederationCitation Excerpt :The average tensile strength of the joints at the three positions was used to evaluate the weld quality. Since the texture is a visual phenomenon in which local structural features appear repeatedly [25], The changes of the weld surface contour caused by tool wear will be reflected in the changes of the smallest repeating structural unit in texture image, which is called texture primitive. Then a method is needed to extract the texture features from the original pixel value of the weld surface image.
Use of image analysis to evaluate the effect of high hydrostatic pressure and pasteurization as preservation treatments on the microstructure of red sweet pepper
2015, Innovative Food Science and Emerging TechnologiesCitation Excerpt :In this case, we also observed that increased magnification resulted in a significant decrease (p < 0.05) in the contrast for all of the treatments considered in the study. Several publications can be found in the literature that indicate the pertinence of the textural feature contrast to numerically describe microstructural changes (Mendoza et al., 2007; Park & Chen, 2001; Zheng, Sun, & Zheng, 2007). By contrast, Laddi et al. (2013) when studying the contribution of different textural features (entropy, contrast, energy, homogeneity, and correlation) to estimate the tea quality concluded that contrast have no effect on tea quality.
Applying an intelligent model and sensitivity analysis to inspect mass transfer kinetics, shrinkage and crust color changes of deep-fat fried ostrich meat cubes
2014, Meat ScienceCitation Excerpt :The color of the fried product which is affected by the operation conditions is the most important quality parameter evaluated by consumers and is critical in the acceptance of the product, even before it enters the mouth (Brewer, 2004). Recently image analysis has become an encouraging approach for determining food quality attributes objectively, especially in meat and meat products (Damez & Clerjon, 2008; Del Moral, O'Valle, Masseroli, & Del Moral, 2007; Valous, Mendoza, Sun, & Allen, 2009; Zheng, Sun, & Zheng, 2007). Assessment of color can be carried out with different methods including visual inspection, instrumental and computer vision (CV) methods.
Discrimination of crumb grain visual appearance of organic and non-organic bread loaves by image texture analysis
2008, Journal of Food Engineering