Study of Neural Network Training Algorithms in Detection of Wood Surface Defects

Abstract

Accurate detection of defects through machine vision improves the economical growth of wood industry. In 
this paper six common defects on wood surface are considered for study. Quality of wood image is enhanced by 
Histogram Equalization method. Contrast enhanced images are subject to Thresholding segmentation which examines 
the objects in the image and identifies the defect. The segmented images are cropped in to small blocks. Segmentation
based Fractal Texture Analysis (SFTA) feature extraction method is accomplished to extract 21 texture features from 
the wood images. The extracted features are fed in to the training algorithms such as Levenberg-Marquardt, Scaled 
Conjugate Gradient, Gradient Descent with Adaptive Learning Rate, Bayesian Regularization and Resilent 
Backpropagation. The Performance of training algorithms are analyzed with several performance metrics. The result 
obtained shows a considerable improvement in accuracy of 98.2 % by Bayesian Regularization tool.