REAL-TIME DEFECT DETECTION IN FRUIT .2. AN ALGORITHM AND PERFORMANCE OF A PROTOTYPE SYSTEM

An algorithm to acquire and analyze two combined near infrared (NIR) i mages of each fruit in real-time was developed and implemented with a pipeline image processing system. Information from the structured illu mination portion of each image was used to distinguish between defects and concavities which both appeared as dark spots in the diffusely il luminated scene. The total projected area of defects on each fruit war ; estimated, and subsequent classification was based on the defect pix el total. Apples and peaches were tested at a rate of 5 fruit/s to eva luate system performance. By adjusting the defect pixel threshold to a chieve a 25% error rate on good apples, classification errors for brui se, crack, and cut classes were 51%, 42%, and 46%, respectively. Compa rable results for bruise, scar; and cut peach classes were 48%, 22%, a nd 58%, respectively. Specular reflectance was the major source of err or in the apple data. Acquiring more than two images of each fruit and using more than six lines of structured illumination per fruit would reduce sorting errors. Potential sorting efficiencies were determined by removing fruit from the original data set in which a defect was not presented to the camera or the concavity was between consecutive line s of structured illumination. With a 25% sorting error rate for good c lasses, the classification error rates for bruise, crack, and cut appl e classes were reduced to 38%, 38%, and 33%, respectively. Similarly, error rates for the bruise, scar and cut peach classes were reduced to 9%, 3%, and 30%, respectively.