Skip to Main Content
Automatic detection of ductwork is very desirable to replace manual inspection. Visual method is normally employed in this area, in which a reliable segmentation of the duct image is essential. This paper presents a hierarchical coarse-to-fine image segmentation method in a noise environment. False alarms could progressively be eliminated by sequentially using Otsu's method based on global adaptive thresholding, level set with local information and prior shape knowledge, and parameterized mathematical morphology. This approach is accurate and robust, thus can be applied in strongly noisy condition. Moreover, different defects with various shapes, such as crack, joint and hole, are segmented separately by the shape analysis and mathematical morphology, such that the geometrical features are easily extracted for automated defect detection. Experimental results validate the effectiveness and completeness of the proposed image segmentation method.