Search

Detection of Clostridioides difficile cells in scanning electron microscopy images is a challenging task due to the challenges of cell rotation and inhomogeneous illumination. Currently, orientation-invariance in deep ConvNets is achieved by data augmentation. However, training with all possible orientations increases computational complexity. Furthermore, conventional illumination-invariance models include pre-processing illumination normalization steps. However, illumination normalization algorithms remove important texture information which is critical for the analysis of SEM images. In this paper, RISEC (Rotational Invariant Segmentation of Elongated Cells in SEM images with Inhomogeneous Illumination) is proposed to address the challenges of cell rotation and inhomogeneous illumination. First, a generative adversarial network segments the candidate cell regions proposals, addressing the inhomogeneous illumination. Then, the region proposals are passed to two capsule layers where a rotation-invariant shape representation is learned for every cell type via dynamic routing. Our experiments indicate that RISEC outperforms the state of the art models (e.g., CapsNet, and U-net) by at least 11% improving the dice score.

Hidden conditional random fields (HCRFs) are a powerful supervised classification system, which is able to capture the intrinsic motion patterns of a human action. However, finding the optimal number of hidden states remains a severe limitation for this model. This paper addresses this limitation by proposing a new model, called robust incremental hidden conditional random field (RI-HCRF). A hidden Markov model (HMM) is created for each observation paired with an action label and its parameters are defined by the potentials of the original HCRF graph. Starting from an initial number of hidden states and increasing their number incrementally, the Viterbi path is computed for each HMM. The method seeks for a sequence of hidden states, where each variable participates in a maximum number of optimal paths. Thereby, variables with low participation in optimal paths are rejected. In addition, a robust mixture of Student’s t-distributions is imposed as a regularizer to the parameters of the model. The experimental results on human action recognition show that RI-HCRF successfully estimates the number of hidden states and outperforms all state-of-the-art models.

For many computer vision applications, such as image description and human identification, recognizing the visual attributes of humans is an essential yet challenging problem. Its challenges originate from its multi-label nature, the large underlying class imbalance and the lack of spatial annotations. Existing methods follow either a computer vision approach while failing to account for class imbalance, or explore machine learning solutions, which disregard the spatial and semantic relations that exist in the images. With that in mind, we propose an effective method that extracts and aggregates visual attention masks at different scales. We introduce a loss function to handle class imbalance both at class and at an instance level and further demonstrate that penalizing attention masks with high prediction variance accounts for the weak supervision of the attention mechanism. By identifying and addressing these challenges, we achieve state-of-the-art results with a simple attention mechanism in both PETA and WIDER-Attribute datasets without additional context or side information.

Segmentation of cells in scanning electron microscopy images is a challenging problem due to the presence of inhomogeneous illumination. Classical pre-processing methods for illumination normalization destroy the texture and add noise to the image. In this paper, we present a deep cell segmentation method using adversarial training that is robust to inhomogeneous illumination. Specifically, we apply a model based on U-net as the segmenter and a deep ConvNet as the discriminator for the adversarial training called SoLiD: “Segmentation of clostridioides difficile cells in the presence of inhomogeneous iLlumInation using a Deep adversarial network”. We also present an image augmentation algorithm to obtain the training images required for SoLid. The results indicate that SoLiD is robust to inhomogeneous illumination. The segmentation performance is compared to the U-net and the dice score is improved by 44%.

Simultaneous localization and mapping based augmented reality (AR) is trending in mobile AR applications. With the help of depth sensors, both accuracy and speed have improved. However, the method that performs the initial registration to align virtual objects with real scenery is not well developed, especially for some applications requiring very accurate registration (e.g., systems used in medical applications). For the iRay system, which is a mobile AR system using the Structure Sensor, we propose to use an iterative closest point algorithm to initially register the scanned mesh with the torso surface obtained pre-operatively, and then use SLAM to track pose. In this paper, a semi-automatic initial registration strategy is evaluated by a user study. This strategy is designed to help the user modify the selection of the 3D scanning area, so that the errors introduced by subjective differences can be reduced. The results indicate that the proposed strategy helps the users improve initial registration quality and reduces the average needed time.

This paper presents an efficient hybrid (top-down and bottom-up) framework for activity recognition based on analyzing group context in crowded scenes. The approach presented starts by discovering interacting groups of people using a graph based clustering algorithm. Given the interacting groups, a novel group context activity descriptor is computed that captures not only the focal person’s activity but also the behaviors of neighbors in the group. Finally, for a high-level of understanding of human activities, we propose a bottom-up approach using a random field model to encode activity relationships between people in the scene. We evaluate our approach on two public benchmark datasets and compare the utility of our proposed descriptor with other descriptors using the same baseline recognition framework. The results of both the steps show that our approach with the proposed descriptor achieves recognition rates comparable to state-of-the-art methods for activity recognition in crowded scenes.

Biometric systems compare biometric samples to produce matching scores. However, the corresponding distributions are often heterogeneous and as a result it is hard to specify a threshold that works well in all cases. Score normalization techniques exploit the score distributions to improve the recognition performance. The goals of this chapter are to (i) introduce the reader to the concept of score normalization and (ii) answer important questions such as why normalizing matching scores is an effective and efficient way of exploiting score distributions, and when such methods are expected to work. In particular, the first section highlights the importance of normalizing matching scores; offers intuitive examples to demonstrate how variations between different (i) biometric samples, (ii) modalities , and (iii) subjects degrade recognition performance ; and answers the question of why score normalization effectively utilizes score distributions. The next three sections offer a review of score normalization methods developed to address each type of variation. The chapter concludes with a discussion of why such methods have not gained popularity in the research community and answers the question of when and how one should use score normalization.

In the majority of the methods proposed for age classification from facial images, the preprocessing steps consist of alignment and illumination correction followed by the extraction of features, which are forwarded to a classifier to estimate the age group of the person in the image. In this work, we argue that face frontalization, which is the correction of the pitch, yaw, and roll angles of the headpose in the 3D space, should be an integral part of any such algorithm as it unveils more discriminative features. Specifically, we propose a method for age classification which integrates a frontalization algorithm before feature extraction. Numerical experiments on the widely used FGnet Aging Database confirmed the importance of face frontalization achieving an average increment in accuracy of 4.43%.

A probabilistic framework using two random fields, DeTEC (Detection of Touching Elongated Cells) is proposed to detect cells in scanning electron microscopy images with inhomogeneous illumination. The first random field provides a binary segmentation of the image to superpixels that are candidates belonging to cells, and to superpixels that are part of the background, by imposing a prior on the smoothness of the texture features. The second random field selects the superpixels whose boundaries are more likely to form elongated cell walls by imposing a smoothness prior onto the orientations of the boundaries. The method is evaluated on a dataset of Clostridium difficile cell images and is compared to CellDetect.

This paper presents a novel approach for multi-person tracking utilizing a model motivated by the human vision system. The model predicts human motion based on modeling of perceived information. An attention map is designed to mimic human reasoning that integrates both spatial and temporal information. The spatial component addresses human attention allocation to different areas in a scene and is represented using a retinal mapping based on the log-polar transformation while the temporal component denotes the human attention allocation to subjects with different motion velocity and is modeled as a static-dynamic attention map. With the static-dynamic attention map and retinal mapping, attention driven motion of the tracked target is estimated with a center-surround search mechanism. This perception based motion model is integrated into a data association tracking framework with appearance and motion features. The proposed algorithm tracks a large number of subjects in complex scenes and the evaluation on public datasets show promising improvements over state-of-the-art methods.

In this paper, we propose a method to improve nearest neighbor classification accuracy under a semi-supervised setting. We call our approach GS4 (i.e., Generating Synthetic Samples Semi-Supervised). Existing self-training approaches classify unlabeled samples by exploiting local information. These samples are then incorporated into the training set of labeled data. However, errors are propagated and misclassifications at an early stage severely degrade the classification accuracy. To address this problem, the proposed method exploits the unlabeled data by using weights proportional to the classification confidence to generate synthetic samples. Specifically, our scheme is inspired by the Synthetic Minority Over-Sampling Technique. That is, each unlabeled sample is used to generate as many labeled samples as the number of classes represented by its \(k\)-nearest neighbors. In particular, the distance of each synthetic sample from its \(k\)-nearest neighbors of the same class is proportional to the classification confidence. As a result, the robustness to misclassification errors is increased and better accuracy is achieved. Experimental results using publicly available datasets demonstrate that statistically significant improvements are obtained when the proposed approach is employed.

We present a framework to address the imbalanced data problem using semi-supervised learning. Specifically, from a supervised problem, we create a semi-supervised problem and then use a semi-supervised learning method to identify the most relevant instances to establish a well-defined training set. We present extensive experimental results, which demonstrate that the proposed framework significantly outperforms all other sampling algorithms in 67% of the cases across three different classifiers and ranks second best for the remaining 33% of the cases.