Automated Corrosion Detection Using Crowd Sourced Training for Deep Learning

The recent improvements in Artificial Intelligence (A.I.) for object recognition is largely attributable to the emergence of so-called deep learning artificial neural networks. Deep learning has developed as the natural progression from 'shallow networks' to multi-layered networks of neurons that are able to transform representations (of data, including images) from simple to complex, with increasing layer depth [1]. A review of deep learning and its broader capabilities was presented by LeCun et al. [2], and a review summarising the utilisation of machine learning in the study of materials degradation was also recently reported [3]. The automated detection of corrosion from images, either photographs or video, which may be collected from simple or sophisticated devices (i.e., smart phones, drones, etc.), presents significant advantages in terms of corrosion monitoring. Such advantages include ready access to geographically remote locations, the mitigation of safety risks to inspectors, cost savings and in monitoring speed. The automated detection of corrosion requires deep learning to approach human level intelligence. In order to demonstrate proof of concept the current research is specifically focused on the detection of ferrous corrosion, commonly referred to as 'rust'. A number of emerging, and presently evolving, technologies have enabled the development of deep learning methods, including the increased computing power of graphical processing units (GPUs), algorithm developments, and the capability to build large training sets via the Internet. Supervised learning A.I. utilises 'labelled data' to train neural networks. Such data, which could be an image, will include identification of whether or not corrosion is present in that image. During training an objective function is optimized by adjusting weights of neurons in a neural network [4]. In the case of convolutional neural networks, the neurons are arranged into layers of filters, that convolve the input and then output a non-linear transform of the data. Fundamentally, supervised learning drives the model to develop its own filters to maximize its probability of succeeding at the task. This is an important distinction from traditional classifiers that rely on hand-coded feature detectors. In other words, A.I. image analysis to detect corrosion does not require hard coding, but a deep neural network 'learns' to identify corrosion through its own algorithms. However, one of the major drawbacks of supervised learning is the need for labelled training data. In general terms, more training data leads to better A.I. accuracy [5][6][7][8]. Furthermore, it has been demonstrated that using more training data outperforms A.I. models developed with more accurately labelled data; provided that the incidence of so-called adversarial labelling (i.e., incorrect labelling of training data) is low [9][10][11]. It is believed that this is because there is a need for a certain amount of data for the model to learn a good generalized relationship between the inputs and the desired outputs.

In the interest of driving research into A.I. research, certain communities or research groups are releasing will release ready-made datasets, which include labelled data. Some of the largest and most well-known datasets of labelled images (noting that they do not include labelling for corrosion) include ImageNet [12], MS COCO [13], and ADE 20K [14,15]. It is instructive to look at the effort involved in the creation of these datasets. ImageNet consists of 3.2 million images with a per picture label. These images were collected from the Internet and labelled using Amazon Mechanical Turk, where participants from across the globe vote on the label of each image, a label was not assigned until at least 10 votes were registered. Thus, ImageNet required at least 32 million online votes for labelling. From ImageNet a subset was used for the ImageNet Large Scale Visual Recognition Challenge (ILSVRC), where a monetary prize is awarded to the most accurate automated detection model [16]. The ILSVRC has proven very successful in motivating research into A.I. for image classification.

Without the benefit of a publicly available dataset, labelling large quantities of data is the first and most important step toward developing accurate deep learning models. Methods to predict the required dataset size for a desired accuracy have been developed [17,18] that help researchers forecast the effort involved. Specifically, for corrosion detection and segmentation (i.e., per pixel labelling) the present authors found that there is a need for > 65,000 labelled images required to achieve an essentially human level accuracy for an A.I. model [11]. To generate a large labelled dataset, a special purpose website was developed at www.corrosiondetector.com . This intended to increase the rate of training by enlisting the public to both provide images and to contribute to labelling pre-existing data. Not only

This content is AI-processed based on open access ArXiv data.