• No results found

The thesis is structured as follows. Chapter 2 presents preliminary background theory for the purpose of introducing corrosion and machine learning to readers unfamiliar with these topics. It begins with a short introduction to the basics of corrosion damages in Sec-tion 2.1. Machine learning and neural networks are then defined, and their inner mechanics explained in Section 2.2.

Chapter 3 follows with a literature study on image segmentation. Its purpose is to dis-cuss use-cases of image segmentation for industrial inspections and investigate available image segmentation methods. The chapter begins with Section 3.1 discussing use-cases of image segmentation, and particularly how segmentation of corrosion damages can aid automating industrial inspections. Next, necessary requirements for an image segmenta-tion framework to be used in industrial inspecsegmenta-tions are defined in Secsegmenta-tion 3.2. Finally, Section 3.3 reviews numerous algorithms and machine learning networks with potential for corrosion damage segmentation.

Chapter 4 details methods and implementation. How the dataset was constructed, as well as a statistical analysis of it, is covered in Section 4.1. This is followed by a discussion

of data augmentation methods in Section 4.2. Metrics used to evaluate the performance of different image segmentation algorithms are then discussed in Section 4.3. Implemen-tation, including hardware, software and hyperparameter configurations, are specified in Section 4.5. The chapter ends with an outline of the conducted experiments in Section 4.4 Results and corresponding discussions are presented in Chapter 5. Evaluation of PSP-Net and Mask R-CNN with varying training schemes are given in Section 5.1 and Sec-tion 5.2, respectively. A summary and overall discussion is provided in SecSec-tion 5.3.

Finally, conclusions and further work are found in Section 6.1 and Section 6.2 of Chap-ter 6.

Chapter 2

Preliminaries

This chapter presents preliminary theory on corrosion damages and machine learning in Section 2.1 and Section 2.2, respectively. Corrosion damages are only briefly reviewed as it is not the main focus of the thesis. Please also note that Section 2.2.1, Section 2.2.3 and partly Section 2.2.5 are loosely based on the authors project thesis written Autumn 2019 [14].

2.1 Controlling Corrosion on Steel Constructions

Corrosion is the process in which metals oxidize, i.e. react with the environment into a more chemically stable form with lower energy. This typically happens in humid and salty environments, such as on oil platforms and bridges. If the metal corroding is steel, the term rust is often used. This is by far the most common type of corrosion and is easily recognized by its red/brown color. Corroded copper and zinc, on the other hand, turns turquoise and white, respectively.

Corrosion is essentially an attack on the metal. It degrades properties such as thick-ness of cross section, strength, appearance and magnetic permeability. In worst case, metal constructions collapse due to undetected corrosion damages or lack of necessary mainte-nance, potentially leading to fatalities. Extensive inspections are therefore essential in order to detect damages and initiate maintenance as soon as possible.

Uniform corrosion, i.e. corrosion evenly distributed on the exposed surface of the metal, is often easily detected, repaired and future behavior predicted. An example is shown in Figure 2.1a. Corrosion pits, on the other hand, is much more challenging. As shown in Figure 2.1b, corrosion pits are local holes penetrating deep into the metal creating seemingly small damages, while causing potentially severe damages to the integrity of the construction.

To avoid development of corrosion, metal constructions are often protected with layers of paint. This protects the metal from the aggressive environment as long as the paint is intact. The paint will, however, degrade over time, as shown in Figure 2.1c. Early detection and repair of paint flaking can therefore avoid corrosion of the actual construction.

Constructions with very long design lifetimes, such as bridges, are additionally often

coated with a layer of metallic zinc in between the metal and paint. Thus, if the paint degrades or is damaged, the layer of zinc starts corroding instead of the steel construction itself. The zinc layer essentially works both as a sacrificial anode and a barrier, preventing corrosion of the construction.

(a) Uniform corrosion. (b) Corrosion pits. (c) Degraded paint.

Figure 2.1: Different types of damages. (a) Uniform corrosion, in which the entire metal surface is evenly corroded. (b) Corrosion pits, a type of corrosion damage in which holes penetrate into the metal. (c) Degraded paint, or paint flaking, where the protected metal is exposed to the environment enabling corrosion.

As construction damages are continuously forming, every little damage cannot be re-paired immediately. The Norwegian Public Roads Administration (NPRA) therefore uses estimated area of damaged coating as a criterion to initiate maintenance. Today the area fraction of damaged coating is estimated manually, which introduces a large level of un-certainty, due to the complexity of the task and the fact that many people are involved in the process. Automatic calculation of total damaged area from a set of images would increase the accuracy of the estimate by eliminating the "human factor". However, this is a very challenging task. One necessary component in such a system is the ability to de-tect and outline corrosion within images, i.e. performing image segmentation of corrosion damages. Current state-of-the-art methods for image segmentation are based on artificial intelligence and machine learning, topics of which are discussed in the next section.