Lamb Carcass Classification System Based on Computer Vision Part 2: Texture Features and Neural Networks
Meegalla R. Chandraratne, Sandhya Samarasinghe, Don Kulasiri, Chris Miles Frampton, R. Bickerstaffe · Lincoln University Research Archive (Lincoln University) · 2003
In this study, the ability of neural network models for lamb carcass classification was compared with a multivariate statistical technique with respect to the classification accuracy. The lamb carcass classification system is based on image and texture analyses. Digital images of lamb chops were used to calculate twelve image geometric variables. In addition, a set of ninety textural features was used to extract the textural information from the acquired images. Texture analysis is based on the grey level co-occurrence matrix method. Principal component analysis (PCA) was used to reduce the dimensionality of feature spaces. Two feature sets were generated. These feature sets comprised of 14 principal component (PC) scores calculated from the original variables and 14 variables selected from the original set of variables. Both feature spaces were used for neural network and discriminant analysis. Several network configurations were tested and the classification accuracy of 93% was achieved from three-layer multilayer perceptron (MLP) network. Its performance was 14% better than that from the Discriminant function analysis (DFA). The study shows the predictive potential of combining neural networks with texture analysis for lamb grading.