Local Adaptive Thresholding of Pecan X-Ray Images: Reverse Water Flow Method

Abstract

Segmentation of food product images by global thresholding methods has met with limited success due to unimodal image histograms and inhomogeneous backgrounds. Local adaptive thresholding methods have been successful in segmenting images with inhomogeneous backgrounds; however, there are limited studies applying them to food products. This article evaluates selected local adaptive thresholding methods and proposes a new faster method. The proposed method consists of modifications to the Oh water flow method. For pecan images, the proposed method and the Oh method were able to segment insect-eaten nutmeat and insect exit paths with orientation parallel to the x-ray beam. By adjusting the threshold, the proposed method was also able to segment insect exit paths with perpendicular orientation. The proposed method required only 38.9% of the computation time required by the Oh method. Two applications of the single Otsu threshold method and the Kim method both worked well for larger defects. Pixel misclassification error and relative foreground area error were used as objective indices to evaluate the segmentation results, and the results by the Oh method and the proposed method were comparable. The proposed method also performed well for sample citrus, metallic structure, and cell images. The proposed method is faster and simpler in approach, yet robust and accurate. Features can be extracted from the segmented images for defect classification, thus providing a step toward on-line non-destructive machine vision inspection. The proposed method should be extendable to images with unimodal histograms and inhomogeneous backgrounds.