Separation of Touching Grain Kernels in an Image by Ellipse Fitting Algorithm

An ellipse-fitting algorithm was developed to separate touching grain kernels in images. The algorithm randomly tracks the edge of touching kernels to find the sample points for fitted ellipses. The fitted ellipses were generated by a direct least-squares ellipse-fitting method. Then, clustering was used to identify the best representative ellipse for each kernel of the touching instance. With representative ellipses, touching grain kernel images were separated by morphology transform. Typical touching kernel patterns of four grain types, namely barley, Canada Western Amber Durum (CWAD) wheat, Canada Western Red Spring (CWRS) wheat, and oats obtained from composite samples from several growing locations across the western Canadian prairies were used to test this algorithm. The accuracies of separation were: 92·4% (barley), 96·1% (CWAD), 94·8% (oats), and 97·3% (CWRS wheat). The kernels used for the touching grains images were separated physically to create non-touching instances and another set of images of kernels were acquired. Morphological features were extracted from images of physically separated kernels and compared with features extracted from software-separated kernels. To decide if the difference between two kinds of features was significant, the large sample Z test of hypotheses was employed. Except for Fourier descriptor 1 of barley and CWAD, software separation did not change the values of morphological features more than the tolerance limits of the measurement system. To assess the classification capability after software separation, the morphological features extracted from physically separated kernels were used as training and basic testing data sets, and the features from software-separated kernels were used as production testing data sets. A back-propagation neural network was employed for grain type classification with morphological features as inputs. Compared to 97·1% physically separated grain kernels being correctly classified, the mean classification accuracy for all the software-separated grain types was 96·6%. The morphological features of software-separated kernels were not distorted during software separation and can be successfully used in grain type classification.