Argumentation-enhanced Explainable Image Quality Captioning (Arg-IQC)

Image Quality Assessment (IQA) is an Image Processing task aimed at estimating the quality of a given image. Conventionally, the quality of the given image is assessed by a group of participants who conduct a subjective study and yield Mean Opinion Scores (MOS) [1]. The previous work in IQA focused on predicting MOS but hard to interpret. Recently, the emergence of image captioning techniques [2] offers an alternative to describe the visual quality of an image. Therefore, Image Quality Captioning (IQC) has many potential applications such as helping mobile phone users to improve the quality of images they take by providing real-time feedback.

Computational Argumentation, and specifically argument mining, have recently emerged as some of the most challenging, but also most promising tasks in Natural Language Processing. Argument mining techniques allow to extract argumentative structures from natural language text and audio [3, 4]. These argumentative structures provide insightful information of the reasons behind human argumentative discourses. Furthermore, computational argumentation has proved its effectiveness on providing explainability when needed [5]. Therefore, bringing together both, Image Processing and Computational Argumentation seems to be a sensitive and promising way of addressing IQC from an explainable viewpoint.

This PhD project has the following objectives:

– Design, specify, and create a new dataset containing arguments captions justifying the overall quality assessment of the image based on specific features.

– Formally define and describe the task of Visual Argument Mining, an argument-based image-to-text task that builds on previous work in this field incorporating argumentative reasoning and explainability.

– Develop a multi-task learning model to predict quality score and generate caption to describe image visual quality.

For informal enquiries about the project, contact Hanhe Lin, SSEN, [email protected]

For general enquiries about the University of Dundee, contact [email protected]

Our research community thrives on the diversity of students and staff which helps to make the University of Dundee a UK university of choice for postgraduate research. We welcome applications from all talented individuals and are committed to widening access to those who have the ability and potential to benefit from higher education.

QUALIFICATIONS

Applicants must have obtained, or expect to obtain, a UK honours degree at 2.1 or above (or equivalent for non-UK qualifications), and/or a Masters degree in a relevant discipline. For international qualifications, please see equivalent entry requirements here: www.dundee.ac.uk/study/international/country/.

English language requirement: IELTS (Academic) overall score must be at least 6.5 (with not less than 5.5 in reading, listening, speaking and 6.0 in writing). The University of Dundee accepts a variety of equivalent qualifications and alternative ways to demonstrate language proficiency; please see full details of the University’s English language requirements here: www.dundee.ac.uk/guides/english-language-requirements.

APPLICATION PROCESS

Step 1: Email Hanhe Lin, SSEN, [email protected] (1) send a copy of your CV and (2) discuss your potential application and any practicalities (e.g. suitable start date).

Step 2: After discussion with Dr Lin, formal applications can be made via our direct application system. When applying, please follow the instructions below:

Candidates must apply for the Doctor of Philosophy (PhD) degree in Computing (3 year) using our direct application system:

Please select the study mode (full-time/part-time) and start date agreed with the lead supervisor.

In the Research Proposal section, please:

–         Enter the lead supervisor’s name in the ‘proposed supervisor’ box

–         Enter the project title listed at the top of this page in the ‘proposed project title’ box

In the ‘personal statement’ section, please outline your suitability for the project selected.