Developing robust algorithms for feature extraction in images of polymer layers

Abstract

Automated manufacturing processes are an important component of today’s industries. Assuming the processes are properly maintained they allow for great efficiency when producing various goods. Image analysis can be used as a tool to monitor such processes and evaluate their results.

This Thesis treats development of algorithms for automatic evaluation of images acquired using a standardized procedure. The images contain polymer samples from which the relative positions of two lines are extracted. These positions are thought to be related to machine settings. The image analysis algorithms have to be robust to large variation among the acquired images. Lighting and colour in general as well as shape, location and colour of the two sought lines will vary between samples. Recurring features of the images can be used as a basis of evaluation.

The algorithms use several different techniques in conjunction to identify the lines. Pyramid reduction is used to reduce noise in the images and computational effort. Principal component analysis is used to reduce the number of dimensions treated from three to two. The polymer sample and one of the lines are found using thresholding and morphological operations with parameters automatically calculated using certain parts of the images. The second line is found as the path of most likely pixels considering their intensity, gradient magnitude, gradient direction and location. Finally the relative positions of the lines are examined using principal component analysis.

The resulting algorithm produces relatively good results but has room for improvement, and suggestions for further work are provided. Machine settings are found to influence but not fully explain the relative positions of the two lines.

Popular Abstract

Image analysis as a tool for automatic evaluation and optimization

Today’s industries are often dependent on automated processes. Image analysis can be an important tool to monitor such processes and certain features in the images are often used to quantify the quality.

While the important features in an image are often evident to the human eye, they are not generally trivial for a computer to find. Human vision can detect subtle changes of nuance, texture and colour. What information humans use to do this is not always evident or known, and thus it can be hard to get a computer to replicate the behaviour.

In general algorithms developed to detect certain features in images need to be robust to changes. The need for robustness varies depending on intended use, but might include for example the position, shape, colour or size of the features, as well as the lighting conditions. Occasionally the algorithms might need to detect features partially obstructed by foreground objects.

To find features in images several different techniques can be used and combined. Which techniques yield the best results is highly dependent on the evaluated images. An algorithm meant to detect two lines in images of polymer layers while being robust to changes in size, lighting and positioning has been developed and tested.

The results for the algorithm were good, and 25 out of 30 examined samples got satisfactory results. The quality of the extracted lines means that such an algorithm could be used in the industry given that it is supervised by a human operator. Whenever the algorithm produces an error the supervisor would need to manually correct it.

The algorithm could be further improved by adding so called error handling. Whenever an error occurs it should be automatically detected and the faulty pixels corrected. The resulting algorithm if such error handling was well implemented might be good enough for unsupervised evaluation.

To find one of the lines colour segmentation is used in conjunction with various morphological operations. The other line is found as the most likely line given by combined probability distributions considering pixel intensity, gradient magnitude, gradient direction and an approximate location.

The images of polymer layers evaluated depict polymer samples manufactured using different machine settings. The relative positions of the two lines for various machine settings were evaluated using principal component analysis.

It was found that the distance between the lines is dependent on the machine setting used as the sample was manufactured. Further it was found that there are differences in the relative positions of the two lines not explained by the machine settings, and to fully explain the process further classification would be needed. Since the machine settings used is correlated to the distance between the lines the optimal settings yielding a certain distance between the lines can be calculated.

• author: Wemmenborn, Anton
• supervisor: Johan Lindström
• organization: Mathematical Statistics
• course: FMSM01 20182
• year: 2019
• type: H2 - Master's Degree (Two Years)
• subject: Technology and Engineering
• language: English
• id: 8972470
• date added to LUP: 2019-04-25 11:03:28
• date last changed: 2019-04-25 11:03:28