Agricultural produce grading by computer vision using Genetic Programming

An approach to generating task-specific computer vision systems from generic components using machine learning is presented. With this system, it is possible to learn both feature segmentation and classification from training data. This approach is applied to a disparate range of problems in the domain of agricultural produce grading: mango surface inspection and maturity evaluation, apple variety discrimination, wheat and barley classification and purple sticky rice grading. It is shown that shape, colour and texture features together produce more accurate classification results than fewer categories of feature, and that these evolved classifiers are competitive with neural networks and support vector machines.