Automated monitoring of corrosion on piling sheets

Abstract

Various tasks in the construction industry are tedious due to the high amount of repetition or time-consuming nature. In recent years Deep Learning within computer vision has made it possible to automate various tasks using images. The Hoofdvaarweg Lemmer-Delfzijl has been assessed using images and a pointcloud. The images were being worked with two employees over a month. This is time-consuming and there are a lot of images to go through.

Our project statement is thus: Develop a tool using computer vision techniques to reliably detect problematic corrosion on piling sheet within 4-5 months to understand what the state is of this topic for asset managers.

We first start with an analysis in which we looked at the existing the literature, the data, the existing methods and how Witteveen+Bos is assessing the images. We then set the requirements to which the algorithm should adhere to. Literature study has shown that most models, with data-sets of above 3000 images, achieve above 90% for both accuracy and mean average precision. Afterwards we start writing the algorithm and model testing various model structures as part of the synthesis procedure. The models are variating in structures, filters, depth, and augmentation.

We created a classifier, of four and six classes, and an object detection algorithm and conducted various evaluation techniques. The four-class classifier performed better than the six-class classifier. This could be due to the six-class classifier being made up of less data, classes that are vague, parts of the data showing imbalance problems.

An object detection algorithm was created to detect dimensional features to estimate the height above water and distance of the bumps. To convert the pixel distance to actual distance, we trained the model to detect a reference object. The object detector performed well, but did not meet the requirements we set. The dimension estimation provided can only provide a rough estimation. This may be the result of not every image, in the training set, contained a reference object. Creating the data-set was a tedious task and our data-set with two classes, took around eight hours to finish training.

We can conclude that for image classification, the structure of the model and the trainable parameters play a role. The object detector can count elements, but the predicted bounding box is sometimes larger than expected. Some recommendations are to increase data and classes. A robust feasibility for Witteveen+Bos regarding AI. Repurposing the algorithm for progress monitoring and exploring the interoperability between software relevant for the manager.