Automatic thresholding for defect detection by background histogram mode extents

• Automated visual inspection and defect detection in manufactured parts or surfaces is addressed in this paper.
• The most common technique for defect detection from images, i.e. thresholding, is addressed.
• A new thresholding technique is proposed that effectively segments defects from the background and works for both large and small defects. The proposed technique uses the fundamental idea of thresholding. It conveniently allows for automated detection of defect-free regions by approximately locating defects. The results can be used for identifying the histogram mode of background. The thresholds are then automatically selected at the extents of the background mode.
• The proposed technique offers several free parameters with physical synthetics that can be intuitively adjusted to improve segmentation results or increase the applicability of the method to new images.
• The proposed method is compared to the conventional Otsu method and the results show significant improvements for several examples.

Automatic thresholding is a popular segmentation technique that is widely used for automated visual inspection of defects. Many methods have been proposed for appropriate selection of the threshold value. However, most of these methods perform well for images where defects and background have distinguishable histogram modes and select a threshold close to a valley between the two modes which is usually very hard to locate except for the clearly bi-modal histograms. Additionally, where defect detection requires bi-level segmentation, these methods require a prior knowledge of the number of thresholding levels. In this paper, a new approach for threshold selection is taken that aims to find the threshold value at the boundary of the intensity ranges of defects (object) and background by comparing the histogram modes of the background and defective regions. The proposed method automatically detects defective regions as well as defect-free regions or background. By study of the histogram of the background region, appropriate threshold values are automatically selected at the extents of the background histogram mode. The proposed method proved very effective on several standard images of surface defects. The significance of the method's efficiency is well seen in successfully segmenting defects in images with non-uniform background and with no visible bi- or multi-modal behavior. Another significance of the technique is segmentation of defects comprising of two intensity regions (such as bumps and pits) without specifying the number of threshold levels.