Seed texture classification by random forest and neural networks

Image classification is a crucial problem for many image processing problems. Images that have close textures are challenging to be classified with high accuracy rate. Especially in natural images, classification is a difficult problem when considered independently from the color. In this study, seeds are classified based on textural features obtained from a database with 22 grades of seed. Feature extraction is achieved with the 3 basic feature extraction methods. The attributes are classified by neural network separately and the features yielding the best results are selected. Feature vectors from chosen method are further classified with random forest method. Random forest can be used for data classification with tree structure which has attributes like the number of trees, depth and the number of branches. As a result of experimentations, it is observed that the local binary pattern outperforms other feature descriptors in recognition rate after neural network classification and accuracy rates are further improved after classifying the same attributes with random forest. Seed type and/or defects could be classified with an average error rate of 0.454%.