Using efficient discriminative algorithm in automated visual inspection of feed

A method for classifying the top and the side view of the cylindrical fish-feed pellets is proposed. The top and the side view of cylindrical pellet are circle-like and rectangle-like. The circle-like shape in the top view is not a perfect circle, the circle detection algorithm cannot be effectively applied. In addition, the corner of the feed pellets is not sharp; thus, the corner cannot be detected directly. In this paper, a corner detecting algorithm is applied as follows. Given a point in the pellet´s boundary, its left and right line segments (along the boundary) is the line having the given point as the rightmost and the leftmost pixels, respectively. An angle between the left and the right line segments of every edge point is estimated. A corner is detected as the point whose angle is the local maximum. An approximating rectangle is created from four corners having the four highest angles. The top and the side views are classified according to the two features: (1) the fitness of the approximating rectangle to the edge in the image and (2) the mean angle of the four corners in the approximating rectangle. SVM is used as the classifier. In the experiment on 1000 pellets (500 testing, 500 training), our proposed method achieved 95% classification accuracy.