Anomalous Shape Recognition Using Computer Vision.

Chang-sik Kim, Louisiana State University and Agricultural & Mechanical College


In this research, a new method for anomalous shape recognition, named Curvature-Angular Descriptor (CAD), was written. The CAD generates shape feature vectors from a given curve shape. These vectors can be used to uniquely characterize the shape. The sweet potato has one of the most irregular shapes of any fruit or vegetable. The Standards for Grades of sweet potato shapes (The Code of Federal Regulation, title 7 Agriculture, 1995) are very subjective. For this reason, the sweet potato was chosen as an anomalous shape for evaluating the CAD. The CAD was applied using two methods for sweet potato shape recognition. In the first method (closed-boundary sweet potato shape grading), 407 sweet potatoes were randomly selected from several commercial sweet potato farms in Louisiana. These were inspected by four professional human inspectors and used to extract shape feature vectors using CAD. These extracted feature vectors were clustered using a Learning Vector Quantization (Kohonen, 1989) neural network. There was 27.23% average disagreement between inspectors. The method used in this application gave 24.39% average disagreement with each inspector. Based on experiments, this method achieved about the same ability as human inspectors within the subjective limits of human graders. In the second method (sweet potato grading using low level three dimensional computer vision), 240 random sample sweet potato shapes were selected and inspected by five human inspectors. The LVQ net was trained using extracted shape feature vectors. The trained LVQ net was compared with the human inspectors. The average inspection agreement between each inspector was about 85.8%. On the other hand, the average inspection agreement between the trained LVQ net and inspectors was about 78%. The anomalous shape description method developed in this study, the Curvature-Angular Descriptor, shows great potential as a method to distinguish irregular or defective sweet potato shapes and regular or non-defective sweet potato shapes. In its present form it can be used for research purposes to help determine the shape characteristics of new varieties and the effects on shape of cultural practices and disease. Further refinement in computation and hardware will lead to rapid commercial sorters.