Recurrent neural networks for driving behaviour analysis (ongoing)
Recurrent neural networks are an ideal choice for driving behavior analysis by means of time series of measurements, obtained either from telematics sensors or mobile phones. Intuitively, different driving styles, i.e. aggressive, drowsy or normal, manifest primarily on inertial events, such as accelerations, brakings and turnings. Hybrid intelligent approaches considering both short-term patterns in times-series measurements and overall statistics of inertial events per route, could provide a valuable tool for automotive and car insurance industries.
CT/MRI-based segmentation for intervertebral disc boundary extraction (ongoing)
Intervertebral disc (IVD) localization and segmentation have triggered intensive research efforts in the medical image analysis community, since IVD abnormalities are strong indicators of various spinal cord-related pathologies. Despite the intensive research efforts to address IVD boundary extraction based on MR images, the potential of bimodal approaches, which benefit from complementary information derived from both MRI and CT, has not been fully realized. Furthermore, most existing approaches rely on learning, although sufficiently large and labelled 3D datasets are not always available. In this light, addressing unsupervised, bimodal segmentation methods is a promising direction for vertebrae and IVD boundary extraction.
Recognition of structural elements in floor plans (ongoing)
Computational tools for the analysis of architectural floor plans and the recognition of structural elements such as walls, windows and doors are an emerging research area in document image processing, related to a number of applications in the architectural domain such as re-utilization of previous designs, 3D reconstruction, and assistance in decoration. The design and implementation of such computational tools pose several challenges: the large variability of notation styles complicate the development of a generic approach, whereas a floor plan often appears as a noisy, non-vector image with hand-drawn strokes and distortions. Moreover, there is a lack of large annotated datasets, encompassing all notation styles in order to enable generic deep learning-based recognition.
Physics-inspired image processing (ongoing)
Camera sensors are inherently sensitive to the near-infrared (NIR) spectrum (700–1100 nm). Past work on shadow detection and removal has been based on three observations: First, shadows are generally darker than their surroundings, in both the visible and the NIR. Second, the majority of objects that are dark in the visible spectrum are much brighter in NIR. Third, most of the considered illuminants in the shadow formation process have a distinct behavior in the NIR (Rufenacht et al., 2014). This research aims to employ physics-inspired image processing concepts to address NIR-based image enhancement and pre-processing.
Pedestrian recognition based on LIDAR data
This research introduces encoding methods for pedestrian recognition, based on statistical shape analysis of 3D LIDAR data. The proposed approach has two variants, based on the encoding of local shape descriptors either in a spatially agnostic or spatially sensitive fashion. The latter method derives more detailed cues, by enriching the ‘gross’ information reflected by overall statistics of local shape descriptors, with ‘fine-grained’ information reflected by statistics associated with spatial clusters. Experiments on artificial LIDAR datasets, which include challenging samples, as well as on a large scale dataset of real LIDAR data, lead to the conclusion that both variants of the proposed approach 1) obtain high recognition accuracy, 2) are robust against low-resolution sampling, 3) are robust against increasing distance, and 4) are robust against non-standard shapes and poses. On the other hand, the spatially-sensitive variant is more robust against partial occlusion and bad clustering.
Predictive digitisation of cultural heritage objects
This research introduces the concept of predictive digitisation. Based on single partial scans of an object, the predictive digitisation platform (PDP) predicts the shape of digitized objects without having to perform full 3D scanning. The input point cloud data from the acquisition source are used to retrieve and fit the closest matching candidate shape from a digitized artefact repository onto the acquired geometry, thus predicting and automatically suggesting the geometry for the part that has not been scanned. In this work, PDP is realised by integrating two distinct components: a partial 3D object retrieval methodology (P3DOR) and a reshaping methodology which relies upon the rigid registration method and the registration-based k-sparse algorithm. Overall, PDP opens a whole new range of possibilities for decreased acquisition times and simplified procedures, leading to an order of magnitude reduction in time and cost. CH objects provide an ideal application domain for PDP, since such objects posses regularity, symmetries or repeated patterns and salient features.
This work also lead to a novel P3DOR method, applicable on both point clouds and structured 3D models, which is based on a shape matching scheme combining local shape descriptors with their Fisher encodings. Experiments on the SHREC 2013 large-scale benchmark dataset for partial object retrieval, as well as on the publicly available Hampson pottery dataset, demonstrate that the proposed method outperforms seven recently evaluated partial retrieval methods.
Machine learning and analysis of glial cell morphological progressions on fluorescence confocal microscopy images
This research addresses the development of intelligent multi-channel 3D image analysis methods, with emphasis on optical microscopy, for the quantification of structural associations, identification of critical events, as well as of spatial and temporal dependencies.
Α novel machine learning method has been introduced for astrocyte basal region detection and arbor reconstruction on 3D images obtained with fluorescence confocal microscopy. The proposed method is the first to address existing challenges in astrocyte imaging and quantification, which include the multi-scale nature of astrocytes, their structural similarity and spatial proximity with other types of cells, the prolific alterations in their morphology induced by activation states and the complex patterns of connectivity with other astrocytes, other cell types and foreign objects. An active learning algorithm is trained to identify the basal regions of astrocyte arbors and nuclei. Arbors are then reconstructed automatically using an efficient tracing algorithm which we term as, local priority-based parallel tracing. Arbor measurements are extracted using Scorcioni’s L-measure. These measurements are analyzed by unsupervised harmonic co-clustering to reveal the arbor morphological diversity.
Another machine learning method has been developed to aim the discovery of microglia morphological progressions in response to tissue perturbations, as well as of the underlying arbor features, from 3D multi-channel fluorescence confocal microscope images of rat brain tissue multiplex. Microglia are automatically traced, and a set of 131 arbor features are computed. An agglomerative clustering algorithm based on Pearson’s correlation is used to derive coherent modules of features. A k-NNG structural similarity analysis of feature modules enables the construction of a global similarity matrix, from which we derive an interactive progression chart through a modified Fruchterman-Reingold algorithm. The latter clearly reveals a progression from highly ramified microglia to round cells proximal to the injury site of an implanted neural recording device.
Automated active contour parameterization using local image geometry
A principled method for active contour (AC) parameterization remains a challenging issue in segmentation research, with a potential impact on the quality and objectivity of the segmentation results. This work introduces a novel framework for automated adjustment of region-based AC regularization and data fidelity parameters. Motivated by an isomorphism between the weighting factors of AC energy terms and the eigenvalues of structure tensors, we encode local geometry information by mining the orientation coherence in edge regions. In this light, the AC is repelled from regions of randomly oriented edges and guided toward structured edge regions. Experiments are performed on four state-of-the-art AC models and two image restoration models, which have been automatically adjusted and applied on benchmark datasets of natural, textured and biomedical images. The experimental results demonstrate that the obtained segmentation quality is comparable to the one obtained by empirical parameter adjustment, without the cumbersome and time-consuming process of trial and error.
Image analysis framework for infection monitoring
A novel intelligent framework has been introduced for the analysis of infection progress from time-series medical images, with application to pneumonia monitoring. In each image of a series, lung fields are detected and delineated by an active shape model (ASM) variant which is robust against the presence of infections. The relative extent of infections is assessed by supervised classification of samples acquired from the respective image regions. The samples are represented by multiple dissimilarity features fused according to a novel entropy-based weighted voting scheme offering non-parametric operation and robustness to outliers. The output of the proposed framework is a time-series of structured data quantifying the relative extent of infections over time. The results obtained indicate enhanced performance against state of the art. The effectiveness of the proposed framework to pneumonia monitoring, the generality and the adaptivity of its methods open perspectives for application to other medical imaging domains.
Proteomics image analysis
This work introduced a novel active contour-based scheme for unsupervised segmentation of protein spots in two-dimensional gel electrophoresis (2D-GE) images. The proposed segmentation scheme is the first to exploit the attractive properties of the active contour formulation in order to cope with open problems in 2D-GE image analysis, including the presence of noise, streaks, multiplets and faint spots. In addition, it is unsupervised, providing an alternate to the laborious, error-prone process of manual editing, which is required in state-of-the-art 2D-GE image analysis software packages. It is based on the formation of a spot-targeted level-set surface, which uses spot locations identified by means of keypoint detection, as well as of morphologically-derived active contour energy terms, used to guide active contour initialization and evolution, respectively. The experimental results on real and synthetic 2D-GE images demonstrate that the proposed scheme results in more plausible spot boundaries and outperforms all commercial software packages in terms of segmentation quality.
Active contours guided by textural information
This research lead to novel LBP-guided active contour approaches to texture segmentation. The local binary pattern (LBP) operator is well suited for texture representation, combining efficiency and effectiveness for a variety of applications. In this light, two LBP-guided active contours have been formulated, namely the scalar-LBP Active Contour (s-LAC) and the vector-LBP Active Contour (v-LAC). These active contours combine the advantages of both the LBP texture representation and the vector-valued Chan-Vese model. s-LAC avoids the iterative calculation of active contour equation terms derived from textural feature vectors and enables efficient texture segmentation. v-LAC evolves utilizing regional information encoded by means of LBP feature vectors. It involves more complex computations than s-LAC but it can achieve higher segmentation quality. The computational cost involved in the application of v-LAC can be reduced if it is preceded by the application of s-LAC. The experimental evaluation of the proposed approaches on standard reference datasets (VisTex, Brodatz, Berkeley) demonstrates their segmentation performance on a variety of standard images of natural textures and scenes.
Thyroid ultrasound image analysis
This body of work is the first in image analysis literature which focused on the development of image analysis and pattern recognition methodologies for: 1) segmentation of thyroid ultrasound images and extraction of nodule boundaries, 2) boundary feature extraction for malignancy risk assessment, and 3) development of an expert system encompassing the previous methodologies.
The variable background active contour (VBAC) has been developed and applied for hypoechoic nodule boundary detection. VBAC is formulated so as to exclude from background calculations those regions that amplify inhomogeneity. An alternative active contour model has been developed for precise delineation of thyroid nodules of various shapes, which is invariant to echogenicity. The proposed model, named JET (joint echogenicity-texture), is based on a modified Mumford-Shah functional that, in addition to regional image intensity, incorporates statistical texture information encoded by feature distributions.
An optimization framework based on genetic algorithms has been proposed for automatic adjustment of VBAC and JET parameters. Experimental evaluation showed that VBAC and JET models result in precise nodule delineations, comparable with the ones obtained by medical experts and cope with the limitations of previous segmentation approaches.
As a follow-up, a novel machine learning approach has been proposed for malignancy risk assessment of thyroid nodules in ultrasound images. The proposed approach is based on boundary features and is motivated by the correlation which has been addressed in medical literature between nodule boundary irregularity and malignancy risk. k-nearest neighbor (kNN) and support vector machine (SVM) classifiers are employed for the classification tasks, utilizing compactness, fractal dimension and local echogenicity variance. The classification results indicate that the proposed approach is capable of discriminating between medium-risk and high-risk nodules.
Finally, this research concluded with the development of an expert system for the detection of thyroid gland boundaries, as well as for the detection, delineation and assessment of nodules in terms of malignancy risk. This system unifies the previous methodologies. Experimental evaluation shows that the proposed approach: 1) contributes to the objectification of the diagnostic process by the use of explicit image features, 2) is applicable on clinical practice and 3) may contribute to the reduction of false medical decisions.
Research in 3D image reconstruction, digital image and audio watermarking using fractional Fourier transform, content-based image retrieval, 3D fragment reassembly and machine learning-based gridding of DNA microarray images.