Back to search

FRINATEK-Fri prosj.st. mat.,naturv.,tek

Metrological texture analysis for hyperspectral images

Alternative title: Metrologisk teksturanalyse for hyperspektrale bilder

Awarded: NOK 3.1 mill.

Texture is everywhere in our everyday life. When going to a clothing store, we can differentiate between two white shirts, which one is made of silk or linen without having to touch them. The ripeness of fruit can also be determined by its texture. Beyond these daily tasks, texture is also used as cues for quality control in manufacturing industries to, e.g., detect surface defects, as well as in the medical field to aid medical doctors in providing cancer diagnoses. Our human visual system is limited to the visible range of the electromagnetic spectrum; we can only see the colors of an object or surface. We do not have the capability to "see", for example, heat or water content, because this information lies in the infrared range of the electromagnetic spectrum, thus invisible to us. Hyperspectral imaging (HSI) is an imaging technology that can capture such information at a much more detailed resolution. The aim of this project is to combine the potential of HSI in providing highly accurate measurement with new methods of texture analysis. However, the current ways of exploiting hyperspectral images will not allow the analysis results to live up to the said potential. These images must be treated as measurement data and that can be done by enforcing metrology to the framework of texture analysis. This means that a set of quality assessment protocols must also be developed, such that bias, uncertainty, and other metrological units can be quantified and controlled. Such protocols have been developed in this project, and the accompanying datasets have been made publicly available. In addition to developing metrological image processing (IP) tools for hyperspectral images, in this project, we have also taken the first step towards linking the human perception of a fundamental property of texture to its quantification. The usefulness and relevance of the IP tools have been shown in applications such as detection of surface defects, mapping of forest trees, and estimation of oil slick thickness in oil spills disaster. This project has also led to further collaborations in the more general research topics of image quality and hyperspectral imaging.

In line with the objectives of the project, new knowledge and methodologies have been developed in embedding metrology in texture analysis for spectral images. Additionally, this project has also started linking the measurement of texture to the human perception of it. Within the scope of the objectives, the project has significantly strengthened the competency of the project leader and established a lasting relationship between the participants. The formed collaborative relationship not only brings together complementary expertise in computer and imaging sciences, but also in color science, remote sensing and earth sciences, and cultural heritage. The project has shown results where the new knowledge and methodologies were applied to tasks such as estimating oil thickness in an oil spill disaster and the mapping of trees. Both cases demonstrate the potential impact of the results in the mitigation or mapping of risks caused by disaster or hazard, as well as contributing to climate research especially related to the mapping of forests and their health assessment.

Texture analysis is an important area of fundamental study in image processing, allowing to conduct more advanced tasks such as cancer tissue detection and land cover change analysis. Since its first development for the remote sensing field, hyperspectral imaging has been exploited in variety of domains for its potential gain of accuracy compared other imaging modalities. However, despite its capability of capturing very high spatial and spectral resolutions images, current hyperspectral texture analysis methods and algorithms are still not able to fully exploit the potential hyperspectral imaging has to offer. Based on the hypothesis that a hyperspectral image has to be treated as measurements rather than a mere mathematical object for its potential to be fully exploited, this project aims to develop the first metrological texture analysis framework for hyperspectral images. Building on the knowledge of grayscale and color texture analyses together with spectral mathematical morphology framework, theoretical tools in the framework will also be validated according to metrological constraints, e.g., bias, uncertainty, trueness, etc. Last but not least, contributions will be made to the remote sensing and cultural heritage fields by solving several application tasks, at the same time demonstrating the usefulness of the developed framework.

Publications from Cristin

No publications found

No publications found

No publications found

Funding scheme:

FRINATEK-Fri prosj.st. mat.,naturv.,tek