D. Garten, P. Brückner, and G. Linβ (Germany)
Machine learning, feature selection, wheat, svm
Image-based recognition of natural objects is a sophisticated task. For the optical recognition of different impurities of a grain sample specialized image acquisition devices and image features are needed. We evaluated different colour and shape features as input for a support vector machine (SVM). The influence of the number of selected features on the classifier performance was also investigated. Feature selection was conducted using feature ranking in conjunction with a method for estimating the optimal number of features. For calculating the feature score we analyzed the well known Relief algorithm [1], Information Gain (InfoGain) [2] and Ambiguity-Measure [3]. As a result we got a 160 dimensional feature vector as input for the classifier. With the SVM classifier we reached total recognition rates of about 91% for this very complex recognition problem.
Important Links:
Go Back
