| Line-scan imaging was investigated as a means for detection of surface defects on dried prunes. Spectral characteristics of the defects were measured. The degree of discontinuity in gray level was used as the sorting criterion.; A laboratory line-scan imaging system was developed to acquire and process prune images. An algorithm for prune defect detection was developed based on row gradients, local thresholds, and a nonparametric classifier. Emphasis in the algorithm development was on simplicity, to allow for an eventual sorting rate of 20 prunes/s. Classification errors were 8% for scab, 0% for exposed pit, 0% for mold, 1% for crack, 3% for insect injury, and 0% for good prunes with the defective pixel threshold value of TDP = 1.25%. Specular reflectance did not significantly compromise sorting accuracy.; A prototype automatic prune sorter was developed, including a feeder, singulator, illumination chamber, three line-scan cameras, three multiprocessor camera subsystems, master computer, and pneumatic rejector. The defect detection algorithm was simplified to only one feature of defective rows, DR, used for the classification. A typical prune image could be processed in less than 15 ms. When a prune was moved in air at a speed of 2 m/s through the illumination chamber, its entire surface was scanned by three line-scan cameras. High speed image frame grabber circuits with parallel structure were developed to allow the subsystem computers to process the current image while the next image was being digitized. An adaptive image sizing scheme developed was found important for high speed image acquisition, processing, and pneumatic rejection.; Real-time sorting errors were 5.2% for mold, 9.1% for crack, 16.5% for scab, and 3.7% for good prunes with the gradient threshold value of TGRAD = 64 and the defective row threshold value of TDR = 16 at a sorting rate of 10 fruit per second. Greater accuracy could be achieved with the original algorithm and faster computer hardware. |
