Automatic detection of patch-like defects on apples

This paper presents a machine vision system for the detection of patch-like defects on apples. The system consists of three parts: initial segmentation, stalk and calyx identification and refinement of defect segmentation. Dark patches which may include both defects and stalks/calyxes are first segmented out with a flooding algorithm. To identify stalks and calyxes so as to distinguish them from defects, a structured light and neural network approach is adopted. The structured light provides qualitative 3D shape information, and with the information and the features extracted from apple grey-level images, the neural network classifies each segmented patch as defective or non-defective. For defective ones, the segmentation is refined by a snake algorithm, which improves the accuracy of boundary localization. The experimental results with sample apples show that the proposed system can accurately detect patch-like defects and distinguish them from stalks and calyxes