Machine vision using artificial neural networks with local 3D neighborhoods

Several approaches have been reported previously to identify internal log defects automatically using computed tomography (CT) imagery. Most of these have been feasibility efforts and consequently have had several limitations: (1) reports of classification accuracy are largely subjective, not statistical; (2) there has been no attempt to achieve real-time operation; and (3) texture information has not been used for image segmentation, but has been limited to region labeling. Neural network classifiers based on local neighborhoods have the potential to greatly increase computational speed, can be implemented to incorporate textural features during segmentation, and can provide an objective assessment of classification performance. This paper describes a method in which a multilayer feed-forward network is used to perform pixel-by-pixel defect classification. After initial thresholding to separate wood from background and internal voids, the classifier labels each pixel of a CT slice using histogram-normalized values of pixels in a 3 x 3 x 3 window about the classified pixel. A post-processing step then removes some spurious pixel misclassifications. Our approach is able to identify bark, knots, decay, splits, and clear wood on CT images from several species of hardwoods. By using normalized pixel values as inputs to the classifier, the neural network is able to formulate and apply aggregate features, such as average and standard deviation, as well as texture-related features. With appropriate hardware, the method can operate in real time. This approach to machine vision also has implications for the analysis of 2D gray-scale images or 3D RGB images.