Marbling, i.e. intramuscular fat, is an important factor contributing to beef tenderness and therefore applied as an indicator for beef grading. Although the degree of marbling is usually determined by experienced graders, subjective assessment and human error could lead to inaccurate and inconsistent decision. Therefore, a number of methods particularly based on image processing have been adopted to automatically evaluate meat marbling. Nevertheless, the accuracy of marbling score is frequently attenuated by noise signals embedded in meat image, e.g. the light reflectance of moisture and the ice crystals on meat surface.
We have developed a new image-based method that improves the accuracy of meat marbling grading. The random partial-image selection technique is first applied to meat image. Ten rectangle partial images, each with dimension of 20 percent of the width and the length of the original image, are randomly selected. Multiple thresholding algorithms are then simultaneously applied to each partial image to distinguish the boundaries between fat and muscle areas using ImageJ program version 1.43i. The examples of algorithms used include auto-isodata, isodata, histogram-based thresholding, minimum error thresholding, moment-preserving thresholding, and mixture modeling. The average percent fat area is calculated from results of the multiple thresholding algorithms of all partial images. The quality grade is finally assigned according to the calculated percent fat area. The performance of method has been tested with a total of 434 images of beep rip-eye pieces of various grades, both with and without and noise problems, and compared with whole-image and sliding windows image processing methods. The quality grades assigned by our technique are highly concordant with those assigned by qualified meat inspectors, indicating high accuracy of this technique, which is also higher than other image processing techniques especially in case of noisy images. Using the average fat percentage value calculated from multiple thresholding algorithms avoids the limitation and minimizes the possible error of each algorithm. Analysis of randomly chosen partial-images decreases the chance to capture interferences in meat images such as distorted color or camera noise signals, which are unavoidable in whole-image and sliding windows image processing.
We proposed a new method that can automatically and accurately assign marbling grades to both normal meat and meat images with noise, which cannot efficiently be solved using other image processing techniques. Our approach can be a useful technique to enhance the accuracy of quality grading in meat industry.
beef Marbling fat, Random image selection, multiple thresholding algorithms
Presenting author: Pavita Tipsombatboon
Organisation: BIOTEC, Thailand
co-authors: Ekachai Jenwitheesuk, Jittima Piriyapongsa, Sissades Tongsima