A5022576557- Advanced Vision and Imaging
- Anomaly Detection Techniques and Applications
- Advanced Image and Video Retrieval Techniques
- 3D Surveying and Cultural Heritage
- AI in cancer detection
- Remote Sensing and LiDAR Applications
- Medical Image Segmentation Techniques
- Target Tracking and Data Fusion in Sensor Networks
- 3D Shape Modeling and Analysis
- Robotics and Sensor-Based Localization
- Video Surveillance and Tracking Methods
- Retinal Imaging and Analysis
- Image and Object Detection Techniques
- Image Processing Techniques and Applications
- Visual Attention and Saliency Detection
- Domain Adaptation and Few-Shot Learning
- Machine Learning and Data Classification
- Digital Imaging for Blood Diseases
- Advanced Neural Network Applications
- COVID-19 diagnosis using AI
- Advanced Image Processing Techniques
- Optical measurement and interference techniques
- Natural Language Processing Techniques
- Handwritten Text Recognition Techniques
- Medical Imaging Techniques and Applications
RMIT University
2015-2024
The Royal Melbourne Hospital
2018-2023
MIT University
2020-2022
IBM Research - Australia
2016-2017
Swinburne University of Technology
2013-2014
Retinal swelling due to the accumulation of fluid is associated with most vision-threatening retinal diseases. Optical coherence tomography (OCT) current standard care in assessing presence and quantity image-guided treatment management. Deep learning methods have made their impact across medical imaging, many OCT analysis been proposed. However, it currently not clear how successful they are interpreting on OCT, which lack standardized benchmarks. To address this, we organized a challenge...
Identifying the underlying model in a set of data contaminated by noise and outliers is fundamental task computer vision. The cost function associated with such tasks often highly complex, hence most cases only an approximate solution obtained evaluating on discrete locations parameter (hypothesis) space. To be successful at least one hypothesis has to vicinity solution. Due hypotheses generated minimal subsets can far from model, even when samples are said structure. In this paper we...
This paper proposes a novel method in order to obtain voxel-level segmentation for three fluid lesion types (IR-F/SRF/PED) OCT images provided by the ReTOUCH challenge [1]. The is based on deep neural network consisting of encoding and de-coding blocks connected with skip-connections which was trained using combined cost function comprising cross-entropy, dice adversarial loss terms. results held-out validation set shows that architecture functions used has resulted improved retinal...
State-of-the-art stereo matching networks trained only on synthetic data often fail to generalize more challenging real domains. In this paper, we attempt unfold an important factor that hinders the from generalizing across domains: through lens of shortcut learning. We demonstrate learning feature representations in is heavily influenced by artefacts (shortcut attributes). To mitigate issue, propose Information-Theoretic Shortcut Avoidance (ITSA) approach automatically restrict...
Diabetic Retinopathy (DR) is one of the leading causes blindness worldwide. Detecting DR and grading its severity essential for disease treatment. Convolutional neural networks (CNNs) have achieved state-of-the-art performance in many different visual classification tasks. In this paper, we propose to combine CNNs with dictionary based approaches, which incorporates pathology specific image representation into learning framework, improved classification. Specifically, construct...
This paper presents a new solution for multi-target tracking over network of sensors with limited spatial coverage. The proposed is based on the centralized data fusion architecture. main contribution introduction track-to-track approach in which posterior distributions states, reported by various sensor nodes, are fused way that redundant information combined and rest complement each other. formulated within labeled random finite set framework incorporates all state label provided multiple...
Recently, deep neural networks have achieved remarkable performance in single-image localisation, where the location and orientation of camera is estimated using an independent image. The main bottleneck requirement large volumes annotated data that usually generated structure-from-motion approaches. In this work, we demonstrate convolutional (CNN) can learn from synthetic images to perform task localisation real images, are rendered texture-less 3D models. We represent both as either...
The quality of input images significantly affects the outcome automated diabetic retinopathy screening systems. Current methods to identify image rely on hand-crafted geometric and structural features, that does not generalize well. We propose a new method for retinal classification (IQC) uses computational algorithms imitating working human visual proposed leverages learned supervised information using convolutional neural networks (CNN), thus avoiding hand-engineered features. Our analysis...
The fovea is one of the most important anatomical landmarks in eye and its localization required automated analysis retinal diseases due to role sharp central vision. In this paper, we propose a two-stage deep learning framework for accurate segmentation colour fundus images. first stage, coarse performed localize image. location information from stage then used perform fine-grained region second stage. proposed method performs end-to-end pixelwise by creating model based on fully...
Identifying the underlying models in a set of data points that is contaminated by noise and outliers leads to highly complex multi-model fitting problem. This problem can be posed as clustering projection higher-order affinities between into graph, which clustered using spectral clustering. Calculating all possible computationally expensive. Hence, most cases, only subset used. In this paper, we propose an effective sampling method for obtaining accurate approximation full required solve...
This paper presents an innovative method for motion segmentation in RGB-D dynamic videos with multiple moving objects. The focus is on finding static, small or slow objects (often overlooked by other methods) that their inclusion can improve the results. In our approach, semantic object based and cues are combined to estimate number of objects, parameters perform segmentation. Selective object-based sampling correspondence matching used specific parameters. main issue such approach over...
Deep learning techniques often perform poorly in the presence of domain shift, where test data follows a different distribution than training data. The most practically desirable approach to address this issue is Single Domain Generalization (S-DG), which aims train robust models using from single source. Prior work on S-DG has primarily focused augmentation generate diverse In paper, we explore an alternative by investigating robustness linear operators, such as convolution and dense layers...
One of the core challenges in visual multi-target tracking is occlusion. This especially important applications such as video surveillance and sports analytics. While offline batch processing algorithms can utilise future measurements to handle occlusion effectively, online have rely on current past only. As such, it markedly more challenging applications. To address this problem, we propagate information over time a way that generates sense déjà vu when similar motion features are observed....
Autonomous vehicles in intelligent transportation systems must be able to perform reliable and safe navigation. This necessitates accurate object detection, which is commonly achieved by high-precision depth perception. Existing stereo vision-based estimation generally involve computation of pixel correspondences disparities between rectified image pairs. The estimated disparity values will converted into downstream applications. As most applications often work the domain, accuracy more...
Volumetric imaging is an essential diagnostic tool for medical practitioners. The use of popular techniques such as convolutional neural networks (CNN) analysis volumetric images constrained by the availability detailed (with local annotations) training data and GPU memory. In this paper, image classification problem posed a multi-instance novel method proposed to adaptively select positive instances from bags during phase. This uses extreme value theory model feature distribution without...
In recent years, major capability improvements at synchrotron beamlines have given researchers the ability to capture more complex structures a higher resolution within very short time. This opens up possibility of studying dynamic processes and observing resulting structural changes over However, such studies can create huge quantity 3D image data, which presents challenge for segmentation analysis. Here tomography experiments Australian source are examined, were used study bread dough...
Low visibility hazard detection in construction sites is a crucial task for prevention of fatal accidents. Manual monitoring workers to ensure they follow the safety rules (e.g., wear high-visibility vests) cumbersome and practically infeasible many applications. Therefore, an automated system both fundamental practical interest. This paper proposes intelligent solution that uses live camera images detect who breach by not wearing vests. The proposed formulated form anomaly algorithm...
The accuracy of handwritten text recognition may be affected by the presence struck-out in manuscript. This paper investigates and improves performance a widely used approach Convolutional Recurrent Neural Network (CRNN) on lines containing struck out words. For this purpose, some common types strokes were superimposed words line. A model, trained IAM line database was tested Character Error Rate (CER) increased from 0.09 to 0.11. model re-trained dataset text. performed well terms...
In autonomous vehicles, depth information for the environment surrounding vehicle is commonly extracted using time-of-flight (ToF) sensors such as LiDARs and RADARs. Those have some limitations that may potentially degrade quality utility of to a substantial extent. An alternative solution estimation from stereo pairs. However, matching often fails at ill-posed regions including areas with repetitive patterns or textureless surfaces which are found on planar surfaces. This paper focuses...
Abstract Prostate cancer (PCa) is the second most frequent type of found in men worldwide, with around one nine being diagnosed PCa within their lifetime. often shows no symptoms its early stages and diagnosis techniques are either invasive, resource intensive, or has low efficacy, making widespread detection onerous. Inspired by recent success deep convolutional neural networks (CNN) computer aided (CADe), we propose a new CNN based framework for incidental clinically significant prostate...
Point clouds produced by either 3D scanners or multi-view images are often imperfect and contain noise outliers. This paper presents an end-to-end robust spherical harmonics approach to classifying objects. The proposed framework first uses the voxel grid of concentric spheres learn features over unit ball. We then limit order level suppress effect In addition, entire classification operation is performed in Fourier domain. As a result, our model learned that less sensitive data...
Three-dimensional point clouds produced by 3D scanners are often noisy and contain outliers. Such data inaccuracies can significantly affect current deep learning-based methods reduce their ability to classify objects. Most neural networks-based object classification were targeted achieve high accuracy without considering robustness. Thus, despite great success, they still fail good with low levels of noise This work is carried out develop a robust network structure that solidly identify The...