3D-2D silhouette-based image registration for comparative radiography-based forensic identification
Abstracts:Comparative radiography is a forensic identification technique traditionally involving the manual comparison of ante-mortem and post-mortem radiographs, thus being time consuming and error prone. The main objective is to propose and validate a computer-aided comparative radiography paradigm based on the 3D bone scan-2D radiograph superimposition process of any bone or cavity. The proposal follows an image registration methodology to automatically search for the ante-mortem radiograph acquisition parameters from the forensic object’s silhouette considering occlusions. The underlying optimization problem is complex since a close initialization cannot be assumed and the image intensities are not reliable or not captured. Several experiments were performed to validate the method. First, we study its accuracy and robustness with synthetic images of clavicles, patellae and frontal sinuses. Second, we study how optimization performance and both variability and differences in the segmentation performed by human operators affect the identification using synthetic and real images of frontal sinuses.
Reimagining the central challenge of face recognition: Turning a problem into an advantage
Abstracts:High inter-personal similarity has been universally acknowledged as the principal challenge of automatic face recognition since the earliest days of research in this area. The challenge is particularly prominent when images or videos are acquired in largely unconstrained conditions ‘in the wild’, and intra-personal variability due to illumination, pose, occlusions, and a variety of other confounds is extreme. Counter to the general consensus and intuition, in this paper I demonstrate that in some contexts, high inter-personal similarity can be used to advantage, i.e. it can help improve recognition performance. I start by a theoretical introduction of this key conceptual novelty which I term ‘quasi-transitive similarity’, describe an approach that implements it in practice, and demonstrate its effectiveness empirically. The results on a most challenging real-world data set show impressive performance, and open avenues to future research on different technical approaches which make use of this novel idea.
Sparse Generalised Principal Component Analysis
Abstracts:In this paper, we develop a sparse method for unsupervised dimension reduction for data from an exponential-family distribution. Our idea extends previous work on Generalised Principal Component Analysis by adding L 1 and SCAD penalties to introduce sparsity. We demonstrate the significance and advantages of our method with synthetic and real data examples. We focus on the application to text data which is high-dimensional and non-Gaussian by nature and discuss the potential advantages of our methodology in achieving dimension reduction.
Visual tracking using spatio-temporally nonlocally regularized correlation filter
Abstracts:Due to the factors like rapidly fast motion, cluttered backgrounds, arbitrary object appearance variation and shape deformation, an effective target representation plays a key role in robust visual tracking. Existing methods often employ bounding boxes for target representations, which are easily polluted by noisy clutter backgrounds that may cause drifting problem when the target undergoes large-scale non-rigid or articulated motions. To address this issue, in this paper, motivated by the spatio-temporal nonlocality of target appearance reoccurrence in a video, we explore the nonlocal information to accurately represent and segment the target, yielding an object likelihood map to regularize a correlation filter (CF) for visual tracking. Specifically, given a set of tracked target bounding boxes, we first generate a set of superpixels to represent the foreground and background, and then update the appearance of each superpixel with its long-term spatio-temporally nonlocal counterparts. Then, with the updated appearances, we formulate a spatio-temporally graphical model comprised of the superpixel label consistency potentials. Afterwards, we generate segmentation by optimizing the graphical model via iteratively updating the appearance model and estimating the labels. Finally, with the segmentation mask, we obtain an object likelihood map that is employed to adaptively regularize the CF learning by suppressing the clutter background noises while making full use of the long-term stable target appearance information. Extensive evaluations on the OTB50, SegTrack, Youtube-Objects datasets demonstrate the effectiveness of the proposed method, which performs favorably against some state-of-art methods.
Building discriminative CNN image representations for object retrieval using the replicator equation
Abstracts:We present a generic unsupervised method to increase the discriminative power of image vectors obtained from a broad family of deep neural networks for object retrieval. This goal is accomplished by simultaneously selecting and weighting informative deep convolutional features using the replicator equation, commonly used to capture the essence of selection in evolutionary game theory. The proposed method includes three major steps: First, efficiently detecting features within Regions of Interest (ROIs) using a simple algorithm, as well as trivially collecting a subset of background features. Second, assigning unassigned features by optimizing a standard quadratic problem using the replicator equation. Finally, using the replicator equation again in order to partially address the issue of feature burstiness. We provide theoretical time complexity analysis to show that our method is efficient. Experimental results on several common object retrieval benchmarks using both pre-trained and fine-tuned deep networks show that our method compares favorably to the state-of-the-art. We also publish an easy-to-use Matlab implementation of the proposed method for reproducing our results.
A benchmark and comparison of active learning for logistic regression
Abstracts:Logistic regression is by far the most widely used classifier in real-world applications. In this paper, we benchmark the state-of-the-art active learning methods for logistic regression and discuss and illustrate their underlying characteristics. Experiments are carried out on three synthetic datasets and 44 real-world datasets, providing insight into the behaviors of these active learning methods with respect to the area of the learning curve (which plots classification accuracy as a function of the number of queried examples) and their computational costs. Surprisingly, one of the earliest and simplest suggested active learning methods, i.e., uncertainty sampling, performs exceptionally well overall. Another remarkable finding is that random sampling, which is the rudimentary baseline to improve upon, is not overwhelmed by individual active learning techniques in many cases.
Polyp detection during colonoscopy using a regression-based convolutional neural network with a tracker
Abstracts:A computer-aided detection (CAD) tool for locating and detecting polyps can help reduce the chance of missing polyps during colonoscopy. Nevertheless, state-of-the-art algorithms were either computationally complex or suffered from low sensitivity and therefore unsuitable to be used in real clinical setting. In this paper, a novel regression-based Convolutional Neural Network (CNN) pipeline is presented for polyp detection during colonoscopy. The proposed pipeline was constructed in two parts: 1) to learn the spatial features of colorectal polyps, a fast object detection algorithm named ResYOLO was pre-trained with a large non-medical image database and further fine-tuned with colonoscopic images extracted from videos; and 2) temporal information was incorporated via a tracker named Efficient Convolution Operators (ECO) for refining the detection results given by ResYOLO. Evaluated on 17,574 frames extracted from 18 endoscopic videos of the AsuMayoDB, the proposed method was able to detect frames with polyps with a precision of 88.6%, recall of 71.6% and processing speed of 6.5 frames per second, i.e. the method can accurately locate polyps in more frames and at a faster speed compared to existing methods. In conclusion, the proposed method has great potential to be used to assist endoscopists in tracking polyps during colonoscopy.
Deep learning for image-based cancer detection and diagnosis − A survey
Abstracts:In this paper, we aim to provide a survey on the applications of deep learning for cancer detection and diagnosis and hope to provide an overview of the progress in this field. In the survey, we firstly provide an overview on deep learning and the popular architectures used for cancer detection and diagnosis. Especially we present four popular deep learning architectures, including convolutional neural networks, fully convolutional networks, auto-encoders, and deep belief networks in the survey. Secondly, we provide a survey on the studies exploiting deep learning for cancer detection and diagnosis. The surveys in this part are organized based on the types of cancers. Thirdly, we provide a summary and comments on the recent work on the applications of deep learning to cancer detection and diagnosis and propose some future research directions.
Curvature-aware manifold learning
Abstracts:One of the fundamental assumptions of traditional manifold learning algorithms is that the embedded manifold is globally or locally isometric to Euclidean space. Under this assumption, these algorithms divided manifold into a set of overlapping local patches which are locally isometric to linear subsets of Euclidean space. Then the learnt manifold would be a flat manifold with zero Riemannian curvature. But in the general cases, manifolds may not have this property. To be more specific, the traditional manifold learning does not consider the curvature information of the embedded manifold. In order to improve the existing algorithms, we propose a curvature-aware manifold learning algorithm called CAML. Without considering the local and global assumptions, we will add the curvature information to the process of manifold learning, and try to find a way to reduce the redundant dimensions of the general manifolds which are not isometric to Euclidean space. The experiments have shown that CAML has its own advantage comparing to other traditional manifold learning algorithms in the sense of the neighborhood preserving ratios (NPR) on synthetic databases and classification accuracies on image set classification.
Auto-weighted 2-dimensional maximum margin criterion
Abstracts:As a hot topic in machine learning, supervised learning is applied to both classification and recognition frequently. However, parameter-tuning in most supervised methods is a laborious work due to its complexity and unpredictability. In this paper, we propose an auto-weighted approach, termed as auto-weighted 2-dimensional maximum margin criterion, which updates the introduced weight in each iteration automatically to leverage the associated terms, so that the weight becomes insensitive to initialization. In addition, the proposed method extracts features from 2-order data directly, i.e., image data. Moreover, we have an observation that the objective value in the proposed method could directly reflect the performance in classification task under the varying dimensionality, which is much beneficial to selection of the optimal dimensionality. Extensive experiments on several datasets are conducted to validate that our method is of great superiority compared to other approaches.