PhD Student
CSIRO Data61
UOW AMRL VILA
School of Computing & IT
University of Wollongong

Current Office:
Room No. 3.237, Building 3
UOW Wollongong Campus

Contact me:
¹  sr801uowmail.edu.au
²  saimun.rahmandata61.csiro.com
  (+61) 4 1374 0811

Find me on :
• • • •

Image and Video Recognition in the Wild

Although image and video recognition systems have achieved impressive performance in recent years, their use in real-world applications remains challenging, especially when the image and videos are captured under non-ideal conditions, as is common in surveillance-based and machine vision based applications. Images and videos captured in such circumstances are often contaminated by blur, non-uniform lighting, low quality, and low inter-class difference. My research addresses these problems and proposes solutions to enable image and video recognition in real-world applications (in the wild) with high precision.

My work focuses on the extraction of effective and efficient features from images and videos. My works are in the areas of Computer Vision, Pattern Recognition and Machine Learning. My contributions are mostly in the field of deep covariance-matrix based feature learning, fine-grained image recognition, spatio-temporal data analysis, texture classification, and action recognition in low quality videos.

My works can be divided into three groups based on their types:

• Medical Image Classification
• Fine-grained Image Recognition
• Machine Vision Inspection
• Action Recognition in Low Quality Videos

---

Medical Image Classification

Medical images classification is a significant research area that receives growing attention from both the research community and medicine industry. It addresses the problem of diagnosis, analysis and teaching purposes in medicine. One of the primary objective of medical images classification is not only to achieve good accuracy but to understand which parts of anatomy are affected by the disease to help clinicians in early diagnosis of the pathology and in learning the progression of a disease. My work focuses on the understanding of medical images and perform their accurate classification with advance machine learning models.

---

Fine-grained Image Recognition

The general object classification task distinguishes very different object categories, such as a house and a bird. In contrast, fine-grained image classification aims to answer the question of given a bird image: which bird species is it? In a more specific way, it is about species and sub-category classification. This is a challenging task for two reasons. Firstly, some classes (species) from the same category, such as fish, have a very similar appearance leading to low inter-class variation. Secondly, a high degree of variability is prone to occur even within the same class due to large pose, lighting, and illumination variations in the natural environment. My research explores methods to improve fine-grained classification in challenging general and medical image datasets.

---

Machine Vision Inspection

While human inspectors working on assembly lines visually inspect parts to judge the quality of workmanship, machine vision systems use cameras and image processing software to perform similar inspections. Machine Vision inspection plays an important role in achieving 100% quality control in manufacturing, reducing costs and ensuring a high level of customer satisfaction. Machine vision system inspection consists of narrowly defined tasks such as counting objects on a conveyor, reading serial numbers, and searching for surface defects. Manufacturers often prefer machine vision systems for visual inspections that require high speed, high magnification, around-the-clock operation, and/or repeatability of measurements.

Following are some key projects in which I was involved during my tenure at Vitrox Corporation. Note that due to commercial reasons most of the findings in these projects were never published or demoed in front of the public.

---

Action Recgnition in Low Quality Videos

Human action recognition (HAR) in video is presently one of fastest growing areas in computer vision research due to the ubiquity and abundance of video data. With the society's increasing awareness for personal safety, the need for better security in public spaces has heightened the importance of automated video surveillance, especially recognising human activities, events or interactions. In recent years, most state-of-the-art techniques for HAR have been designed to perform well under constrained and highly controlled conditions, with many producing high accuracies in good quality videos. These capabilities may not be easily replicable in real-world surveillance conditions (via devices such as CCTV or web cameras) where video quality may be naturally poor.

This research was aimed to investigate the visual recognition of human activities in low quality surveillance videos. Hence, the first crucial task involves examining feasible fundamental representations that utilize spatio-temporal information inherent in videos. Currently, spatio-temporal information (in the form of volumes, trajectories, local features) possesses various notable drawbacks in real-world surveillance videos. Meanwhile, the sensitivity of existing state-of-the-art methods towards video quality demands for the proposal of new learning models that can robustly characterize human activities under such conditions. The effectiveness and robustness of the proposed methods will be evaluated in benchmark surveillance datasets with uncontrolled, low quality video scenes. There is currently no systematic, extensive work analysing this problem domain relating to video quality. Various techniques have been proposed to tackle various poor video quality problems for effective action recognition. Following are the outcomes from this research:

---

Copyright notice: This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. These works may not be reposted without the explicit permission of the copyright holder.

© 2018 Saimunur Rahmnan. CC-NC-SA-4.0. Built with Bootstrap v3.3.6. Style adopted from Animesh Garg.