Autonomous Underwater Vehicles (AUVs) have been vital equipment in a wide area of sectors such as the oil and gas industry, offshore renewables, marine science, aquaculture as well as deep-sea mining. Over the years, AUVs have developed dramatically in terms of power requirements, manoeuvrability, navigation, and autonomy. Autonomous Underwater vehicles can operate in remote hazardous environments where human reach and intervention is too dangerous.
The advantages of using AUVs to perform any given task such as survey missions, mapping the seafloor or performing underwater inspections has been increased over the last decade. After the completion of the mission, the AUV returns to download the data to the operator and be re-programmed for the next mission. Thus, AUV's have the capabilities to observe autonomously, but they cannot physically intervene in a situation, and they need human instruction in order to carry out the given mission. One of the tasks that AUVs are used for in underwater intervention missions is vision.
The proposed project is focused on infrastructure inspection, detection of faults that could occur in a variety of underwater structures and taking the necessary actions using computer vision and decision making in an underwater robot. The research question of the project is: how can intervention by underwater vehicles be achieved using deep learning for computer vision to create a self-controlled AUV capable of critically analysing objects and acting accordingly for the given mission? Therefore, the aim is to investigate the feasibility of physical intervention by an AUV.
To achieve physical intervention from an underwater vehicle which would be able critically to decide how to act in any given task and to mitigate the problem, computer vision will be implemented using Machine Learning (ML) techniques. Using ML, the robot would be able to recognise the object of interest and then to inspect it for faults. From the acquired vision information, ML algorithms will analyse the data, "learn" from that data, and finally make critical decisions in terms of the AUV's actions.

Initially, the system's image recognition will be trained using an in-house developed dataset of images. The image recognition will be based on existing ML techniques using neural networks such as Convolutional Neural Networks.
The next step would be to introduce the system to the actual underwater environment, which adds further complexity to the vision system due to blurring and low visibility. Upon successful image recognition, the system will be further developed to make decisions on how to react or intervene upon recognising a "fault". Furthermore, throughout the different steps of the project, virtual simulations will take place in order to assess the system's performance and make the necessary adjustments. Finally, the fully developed system will be implemented on a physical prototype and will be tested in a laboratory environment to assess its performance.

The anticipated outcomes of the project will be the development of an algorithm capable of identifying and inspecting underwater objects similar to those in a real underwater environment, and of physically intervening when a fault is detected.

The project will have a positive impact in industrial sectors such as the oil and gas industry, and the shipping industry will also benefit since they are seeking autonomous intervention technologies to use for equipment monitoring, fault detection and actively taking actions for repairing malfunctions in a safest and cost-efficient manner with the minimum environmental disruption.