Crack Detection Of Eggshell Featuring An Improved Anisotropic Diffusion Filter And Support Vector Machine
Loading...
Date
2018-06-01
Authors
Abdullah, Mohd Hafidz
Journal Title
Journal ISSN
Volume Title
Publisher
Universiti Sains Malaysia
Abstract
Cracks on eggshell are categorized into two types: (i) macro-crack, and (ii) micro-crack. Unlike macro-crack, the detection of micro-crack is very difficult and
challenging since this type of defect is invisible to naked eyes. This problem has been
partially solved by utilizing a custom made candling light in the background
illumination set-up. Even though this has improved the visibility of micro-crack pixels,
however this imaging technique has also enhanced anomalies and other unwanted
pixels, leading to a very cluttered and noisy images. A three-stage window-free
method was proposed to solve this problem. In the first stage, line enhancement was
implemented in order to enhance the quality of line in the image. Next, the crack
enhancement was performed using an improved anisotropic diffusion filter. In this
case, cracks are characterized by pixels having high intensity and high gradient values.
Using these characteristics, the detection system has been developed to inspect
eggshells and classify them into one of the following three possible classes: (i) intact,
(ii) micro-crack, and (iii) macro-crack. In the third stage, a modified double
thresholding was employed to further highlight crack pixels. Results indicate that the
proposed method is competitive when compared with existing techniques and achieved
better performance in terms of FOM. On average the method has resulted in FOM of
0.73 compared to 0.67, 0.57 and 0.42 produced by the original and two recent variants
of anisotropic diffusion filter for crack enhancement, and 0.52, 0.68 and 0.48 produced
by Otsu, Sobel and Canny techniques for image segmentation. Meanwhile the
classifications has been performed using the state of the art twin bounded support
vector machine (TBSVM) and the results have been compared with the standard
support vector machine (SVM) utilizing three different approaches: (i) one-versus-all
(OVA), (ii) one-versus-one (OVO), and (iii) directed acyclic graph (DAG). Results
reveal that DAG outperforms OVA and OVO with sensitivity, specificity and accuracy
averaging at 93.1%, 96.5% and 93.0% for TBSVM compared to 90.7%, 95.4% and
90.7% for standard SVM. Meanwhile the ROC performance indicates that this
classifier can distinguish between intact and macro-crack samples with 100% certainty.
The performance decreases insignificantly when distinguishing intact from micro-crack and micro-crack from macro-crack samples. Therefore, these results suggest that
the proposed detection system is useful and effective for applications in egg
processing.