Non-rigid registration methods for skirting line detection in wool

Publication Type:
Thesis
Issue Date:
2024
Full metadata record
Automation has emerged as a transformative force in modern society, revolutionizing industries by enhancing efficiency, reducing labour costs, and increasing production output. In agriculture, automation has been particularly impactful, with examples such as precision farming and automated harvesting systems significantly improving crop production and management. Despite these advancements, the wool industry has remained largely untouched by automation. Automation in contaminant detection on fleece can severely improve wool production and reduce on-farm costs, by reducing the reliance on manual labour and enhancing the consistency of the product. To address this automation problem, the thesis introduces an approach to contaminant detection on fleece that centres on the development of an image registration technique. This technique compares images of wool before and after the skirting process, facilitating the detection of the skirting line by identifying the contaminated areas that had been removed in the after-skirt image. To align the after-skirt image with the before skirt image, it is essential to employ an image registration method capable of accommodating the transformations that happen during the skirting process. These transformations are detected with the use of feature matching algorithms and subsequently filtered to remove all outliers. Thus, a novel filtering process is utilised, which accounts for the deformations that happen to the wool during handling. With the use of these correspondences, image alignment is then achieved through the implementation of a non-rigid deformation method. The experiments demonstrate a lower error when aligning the wool images compared to rigid methods. Feature matching algorithms fail to find correspondences in areas with extreme deformation located near the skirting area. Thus, a method that integrates dense correspondences from damaged areas to the previously established correspondences is devised using a learning-based filtering method. By analysing the continuity in deformation, it becomes feasible to identify dense correspondences in the more affected areas found along the edges of the fleece. Finally, a physics-based simulation is employed to find correspondences at the edge of the fleece which undergoes significant stretching. Given the distinctive material properties of wool, feature matching algorithms proved inadequate in identifying correspondences on areas with significant stretching along the skirting line location. This simulation method replicates the elasticity observed at the fleece’s edge during the skirting process. As a result, it becomes possible to identify correspondences at the fleece’s edge, thereby increasing the accuracy of skirting line location and enabling accurate image registration for objects with extreme deformations.
Please use this identifier to cite or link to this item: