Fine localization of wood knots using computer vision methods

Samo Šela (2016) Fine localization of wood knots using computer vision methods. EngD thesis.

Preview

PDF Download (12Mb)

Abstract

The aim of this thesis is to find a viable solution for the fine segmentation of wood knots and lumber. Technological flawlessness is not inherent to the Slovenian wood processing industry and there is recognition that advancement is needed in order to ensure the industry’s competitiveness. This work and contribution is concerned with the integration of computer vision in the automated process of manufacturing shuttering panels. The process of manually patching wood knots represents a bottleneck in the manufacturing process while the effective localization of wood knots is one of the key components in automating this type of system. This paper presents a broad outline of the problem of segmentation. Further, it proposes and evaluates a method of segmentation based on determining the minimum cut, or rather, the maximum flow, on a graph. Using a roughly localized wood knot as its basis, the model determines the probability distribution of the Gaussian mixture for the wood knot and background. Section and border weights are determined on the basis of the acquired models, a graph is constructed and the region outside of the rough section of the wood knot is considered as a base and worked into the background. Calculating the minimum cut of the graph simultaneously presents a solution for segmentation. The result are two separate regions, where one region belongs to the knot and the other to the wood. An evaluation of the proposed method was presented with a collection of wood knots obtained from a Slovenian manufacturer of shuttering panels during the manufacturing process. For the purpose of validating the method, the collection of 119 wood knots was suitably annotated and made available as a public good. Over the entire collection, the proposed method achieved 99.00% accuracy for a precision of 0.94 and recall 0.98.

Actions (login required)