A computer vision and classification system for fresh fruits and vegetables

Quality defects in fresh fruits and vegetables have adverse effects on the market value of the products. Increased consumer awareness of product quality has made the marketing of fresh fruits and vegetables so competitive that maintaining high quality is essential to economic survival in the business. Products with internal defects may go undetected and may damage the surrounding healthy fruits during storage. Thus, there is an urgent need for product classification based on internal quality.;Quality control requires the testing of individual units, a task that is possible only with rapid and nondestructive procedures. X-ray imaging is a nondestructive technique that has shown potential for detecting internal defects in horticultural products. However, a number of issues need to be addressed before a computer vision based classification system can be developed. These are (a) what kind of noise is introduced by the x-ray imaging systems and how can it be removed to minimize artifacts in later processing (filter design)? (b) which image features can be correlated with the product quality (feature selection) and how can these features be measured (feature extraction)? and (c) what kind of classifiers should be used for improved accuracy of product classification?;This research was initiated to finding answers to above noted questions for classification of two products (apples and onions). The image noise analyses suggested that a Gaussian filter is appropriate for removing noise from x-ray images. A 5 x 5 mask was found appropriate for detecting bruises in apples and internal defects in onions. Robert's cross edge detector worked well for detection of image features related to these defects. The morphological opening operation was successful in extracting the area features related to watercore defect in apples. No prefiltering was required for watercore feature extraction, however, as the opening operation removes the noise as well.;Two types of image features (spatial and transformed) were extracted for product classification. Area, intensity, and discrete cosine transform (DCT) coefficients were used for watercore detection in apples. Edge features and DCT coefficients were used for detecting bruises in apples and internal defects in onions. Morphological opening and Robert's cross were used for enhancing area and edge features, respectively.;Combining the transformed features (DCT coefficients) with the spatial features improved the accuracy of product classification. The Bayesian, Fuzzy Logic, and Artificial Neural Network (ANN) based classifiers were evaluated. The fuzzy logic classifier performed as well as the Bayesian classifier which is considered the standard. The ANN classifier, however, performed better than its counterparts. (Abstract shortened by UMI.).