Influence of Dataset Character on Classification Performance of Support Vector Machines for Grain Analysis

Anding, Katharina; Linβ, G.; Brückner, Peter

doi:10.2316/p.2010.674-071

Cited by 1 publication

(1 citation statement)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We have applied these algorithms to the problem of automatic quality assurance of wheat. In the system developed by us, a color line scan camera with 2048 pixels and a specialised separation device [11] was used to acquire single object images from every particle in a wheat sample of nearly 500 g. It is necessary to choose a classifier for the relevance detection step which is fast in training and classification, and it is also sensitive to irrelevant features. Both requirements are fulfilled by the Naive bayes and a support vector machine (SVM) classifier.…”

Section: Feature Set Optimizationmentioning

confidence: 99%

Image Based Recognition - Recent Challenges and Solutions Illustrated on Applications

Garten¹,

Anding²,

Lerm³

2016

MLAIJ

View full text Add to dashboard Cite

In this paper, problems and solutions for the automatic recognition of miscellaneous materials, especially bulk materials are discussed. The fact that many materials, especially natural materials, have a strong phenotypic variability resulting in high intra-class and low inter-class variability of the calculated features poses a complex recognition problem. The recognition of components of a wheat sample or the classification of mineral aggregates serves as an example to demonstrate different aspects in segmentation, feature extraction, classifier design and complexity assessment. We present a technique for the segmentation of highly overlapping and touching objects into single object images, a proposal for feature selection and classifier parameter optimization, as well as a method to visualise the complexity of a highdimensional recognition problem in a three-dimensional space. Every step of the pattern recognition process needs to be optimized carefully with special attention to the risk of overfitting. Modern processors and the application of field-programmable gate arrays as well as the outsourcing of processing steps to the graphic processing unit speed up the calculation and make real-time computation possible also for highly complex recognition problems such as the quality assurance of bulk materials.

show abstract

Section: Feature Set Optimizationmentioning

confidence: 99%