Automatic segmentation of overlapped fruits based on convex hull

Computer vision is widely used in the production processing industry of agricultural products, such as harvest, classification, quality inspection. Expect the foreground and background separation, overlapped objects segmentation is also an important part in agricultural application. The segmentation of overlapped fruits is essential to the location, recognition, classification of a single fruit. An overlapped fruits segmentation algorithm based on convex theory is proposed in this paper, and this method is suitable for nearly convex fruits overlapping segmentation in various circumstances, including overlap of same shape fruits, overlap of different shape fruits and adhesions between fruit and stem. There are five steps: 1) calculate the area ratio of the object itself and its convex hull in image, and judge if it needs to be segmented further; 2) detect candidate split points by using Harries corner detection algorithm, and eliminate some pseudo split points by judging distances to improve algorithm efficiency; 3) split lines are the tangent lines of the split points, where each point leads to two lines, resulting in a series of segment results; 4) calculate the area and the area ratio of each result, where the largest part satisfied with the set area ratio is selected for the final result. Besides, the automatically segmentation model has also been tested; 5) local adjustment has been applied on the result using Snake model, so the result would be more practical. This method has pretty well segment result on adhesive cucumbers, overlapping kiwi fruits, kiwi fruit adhesions to the branch, sweet pepper adhesion to the stem and so on. Results show that the average segmentation accuracy of overlapping situations can achieve 95%, minimum segment accuracy is 91%, and average segmentation time is 6.49 s, indicating the proposed method has preferable robustness and accuracy. © 2019, Asian Association for Agricultural Engineering. All rights reserved.