An application of computer vision technologies to log defect determination

Abstract

Classifying logs has an essential economic impact on the forest industry. The general aim of this research is to demonstrate the feasibility of using computer vision technologies to non-destructively identify and quantify the outside knots on freshly harvested radiata pine (Pinus radiata D. Don) logs.

The main goals of this thesis are: • To summarise current knowledge on log defects and on modelling of defects to infer what would be required in a vision system for log making and grading of Radiata pine • To select key features that can be used to determine knot position and size • To extract the features from images of freshly harvested logs • To analyse the features for determining knot position and size through different approaches • To develop a prototype computer vision system which demonstrates the feasibility of computer vision technologies for log quality assessment based on an image library and neural network package.

This thesis provides new knowledge and procedures, which with further development, can be applied to the tasks of log making and grading in the forest industry. The main contributions are: • It identifies the parameters essential for a machine vision system for log making and grading. • It validates that texture information is important for detecting log defects, and that the log texture orientation field is the major feature for determining knot position and size. • This research establishes the importance of analysing oriented texture. The algorithms for computing symbolic descriptions of oriented textures proceeds in two stages. The first stage consists of the computation of an orientation field in the log texture. The second stage uses statistical pattern recognition, phase portraits and neural networks to analyse the orientation field, and to derive symbolic descriptions for the knot position and size. • This research developed several improved fast algorithms in computation and analysis of the orientation field for practical application in real time. The advanced and improved points comparing these algorithms with the traditional method are: (1) replace the entire image computation with some regional computation; (2) replace the non-linear least squares techniques with the linear least squares techniques; (3) replace the static estimation with the dynamic growth estimation; (4) replace the indirect computation with the direct computation; (5) replace the structure-fixed pattern classifier with the neural network which has a learning function and can be re-trained in practice. This work also provided an improved approach to reduce the influence of noise on computing the orientation field • This research developed two useful tools: (1) an image library and neural network package (Neuraler) for image analysis and pattern recognition, (2) a prototype computer vision system (KnotVision) to determine knot position and size. The experimental results indicated that the KnotVision system can accurately detect more than 95% of the knot positions and more than 90% of the knot sizes. • This research provides an effective solution to several problems in the traditional thresholding techniques for edge detection in log images.