Machine vision inspection of fresh market carrots

Typescript (photocopy).A machine vision system was developed to inspect fresh-market carrots. Software and hardware components were designed and developed into a working inspection system, INSPECT. The main objectives of this study were to characterize normal and defective carrots, develop feature extraction techniques, develop Bayes decision functions for classification, and finally, integrate and test all developments with a sample of carrots. The reflectance properties of fresh-market carrots were measured over the visible and near infrared portion of the electromagnetic spectrum. Characteristic reflectance curves were developed. Soft rot, dry rot, and black crown were significantly different from normal carrot flesh. Cavity spots were not significantly different. It was determined that the electromagnetic range for optimal contrast between normal and defect carrot tissue was 535 to 722 nm. An adaptive thresholding technique was developed to enhance surface defects while retaining edge information. In addition, feature extraction techniques were developed to characterize surface defects, forking, curvature, and brokenness. Using a modified connected components algorithm, the segmented image was divided into blocks. The relationship between blocks was used to extract features for surface defects and forking. A curvature profile was developed from which three features were extracted to measure curvature. Using the diameter profile, three features were derived to characterize brokenness. Using the different features, Bayes classifiers were developed and tested. Normal probability distributions were assumed and used to define the state-conditional probabilities for each of the features. For curvature and brokenness, a log-normal probability distribution was used as well. For curvature classification, the Bayes classifier was not improved. However when a log-normal probability distribution was used for brokenness classification, improve was noticed. Two different neural networks were developed. These provided better classification for both curvature and brokenness characteristics. For surface defects and forking, only the normal probability distributions were used. Using the Bayes classifier, overall misclassification for all features tested was 6.1%. Utilization of neural networks for curvature and brokenness features, reduced misclassification to 3.5%.