Computer Analysis of Images and Patterns

Inhaltsverzeichnis

1. Frontmatter

2. Intelligent Systems

   1. Frontmatter

   2. HMDHBN: Hidden Markov Inducing a Dynamic Hierarchical Bayesian Network for Tumor Growth Prediction

      Samya Amiri, Mohamed Ali Mahjoub

      Abstract

      Radiomics transform medical images into a rich source of information and a main tool for the tumor growth survey, which is the result of multiple processes at different scales composing a complex system. To model the tumor evolution in both time and space we propose to exploit radiomic features within a multi-scale architecture that models the biological events at different levels. The proposed framework is based on the HMM architecture that encodes the relation between radiomic features as observed phenomena and the mechanical interactions within the tumor as a hidden process. On the other hand, it models the Tumor evolution through time thanks to its dynamic aspect. While, to represent the biological interactions, we use a Hierarchical Bayesian Network where we associate a level for each scale (Tissue, cell-cluster, cell scale). Thus, the HMM induces a Dynamic Hierarchical Bayesian Network that encodes the tumor growth aspects and factors.

   3. MIVIABot: A Cognitive Robot for Smart Museum

      Alessia Saggese, Mario Vento, Vincenzo Vigilante

      Abstract

      Cognitive robots are robots provided with artificial intelligence capabilities, able to properly interact with people and with the objects in an a priori unknown environment, using advanced artificial intelligence algorithms. For instance, a humanoid robot can be perceived as a plausible tourist guide in a museum. Within this context, in this work we present how the latest findings in the field of machine learning and pattern recognition can be applied to equip a robot with sufficiently advanced perception capabilities in order to successfully guide visitors through the halls and the attraction in a museum.

      The challenge of running all those algorithms on a mobile, embedded platform in real time is tackled on an architectural level, where all the artificial intelligence features are tuned to run with a low computational burden and a Neural Network accelerator is included in the hardware setup. Improved robustness and predictable latency is obtained avoiding the use of cloud services in the system.

      Our robot, that we call MIVIABot, is able to decode and understand speech as well as extract soft biometrics from its interlocutor such as age, gender and emotional status. The robot can integrate all those elements in a dialog, using basic Natural Language Processing capabilities.

   4. Two-Stage RGB-Based Action Detection Using Augmented 3D Poses

      Konstantinos Papadopoulos, Enjie Ghorbel, Renato Baptista, Djamila Aouada, Björn Ottersten

      Abstract

      In this paper, a novel approach for action detection from RGB sequences is proposed. This concept takes advantage of the recent development of CNNs to estimate 3D human poses from a monocular camera. To show the validity of our method, we propose a 3D skeleton-based two-stage action detection approach. For localizing actions in unsegmented sequences, Relative Joint Position (RJP) and Histogram Of Displacements (HOD) are used as inputs to a k-nearest neighbor binary classifier in order to define action segments. Afterwards, to recognize the localized action proposals, a compact Long Short-Term Memory (LSTM) network with a de-noising expansion unit is employed. Compared to previous RGB-based methods, our approach offers robustness to radial motion, view-invariance and low computational complexity. Results on the Online Action Detection dataset show that our method outperforms earlier RGB-based approaches.

3. Real-Time and GPU Processing

   1. Frontmatter

   2. How Does Connected Components Labeling with Decision Trees Perform on GPUs?

      Stefano Allegretti, Federico Bolelli, Michele Cancilla, Federico Pollastri, Laura Canalini, Costantino Grana

      Abstract

      In this paper the problem of Connected Components Labeling (CCL) in binary images using Graphic Processing Units (GPUs) is tackled by a different perspective. In the last decade, many novel algorithms have been released, specifically designed for GPUs. Because CCL literature concerning sequential algorithms is very rich, and includes many efficient solutions, designers of parallel algorithms were often inspired by techniques that had already proved successful in a sequential environment, such as the Union-Find paradigm for solving equivalences between provisional labels. However, the use of decision trees to minimize memory accesses, which is one of the main feature of the best performing sequential algorithms, was never taken into account when designing parallel CCL solutions. In fact, branches in the code tend to cause thread divergence, which usually leads to inefficiency. Anyway, this consideration does not necessarily apply to every possible scenario. Are we sure that the advantages of decision trees do not compensate for the cost of thread divergence? In order to answer this question, we chose three well-known sequential CCL algorithms, which employ decision trees as the cornerstone of their strategy, and we built a data-parallel version of each of them. Experimental tests on real case datasets show that, in most cases, these solutions outperform state-of-the-art algorithms, thus demonstrating the effectiveness of decision trees also in a parallel environment.

   3. A Compact Light Field Camera for Real-Time Depth Estimation

      Yuriy Anisimov, Oliver Wasenmüller, Didier Stricker

      Abstract

      Depth cameras are utilized in many applications. Recently light field approaches are increasingly being used for depth computation. While these approaches demonstrate the technical feasibility, they can not be brought into real-world application, since they have both a high computation time as well as a large design. Exactly these two drawbacks are overcome in this paper. For the first time, we present a depth camera based on the light field principle, which provides real-time depth information as well as a compact design.

   4. A Real-Time Processing Stand-Alone Multiple Object Visual Tracking System

      Mauro Fernández-Sanjurjo, Manuel Mucientes, Víctor M. Brea

      Abstract

      Detection and tracking of multiple objects in real applications requires real-time performance, the management of tens of simultaneous objects, and handling frequent partial and total occlusions. Moreover, due to the software and hardware requirements of the different algorithms, this kind of systems require a distributed architecture to run in real-time. In this paper, we propose a vision based tracking system with three components: detection, tracking and data association. Tracking is based on a Discriminative Correlation Filter combined with a Kalman filter for occlusions handling. Also, our data association uses deep features to improve robustness. The complete system runs in real-time with tens of simultaneous objects, taking into account the runtimes of the Convolutional Neural Network detector, the tracking and the data association.

   5. Demo: Accelerating Depth-Map on Mobile Device Using CPU-GPU Co-processing

      Peter Fasogbon, Emre Aksu, Lasse Heikkilä

      Abstract

      With the growing use of smartphones, generating depth-map to accompany user acquisitions is becoming increasingly important for both manufacturers and consumers. Depth from Small Motion (DfSM) has been shown to be suitable approach since depth-maps can be generated with minimal effort such as handshaking motion, and without knowing camera calibration parameter. Direct porting of a desktop PC implementation of DfSM on mobile devices propose a major challenge due to its long execution time. The algorithm has been designed to run on desktop computers that have higher energy-efficient optimizations compared to mobile device with slower processors.

      In this paper, we investigate ways to speed up the DfSM algorithm to run faster on mobile devices. After porting the algorithm to the mobile platform, we applied several optimization techniques using mobile CPU-GPU co-processing by exploiting OpenCL capabilities. We evaluate the impact of our optimizations on performance, memory allocation, and demonstrate about 3\(\times \) speedup over mobile CPU implementation. We also show the portability of our optimizations by running on two different ANDROID devices.

4. Image Segmentation

   1. Frontmatter

   2. Skin Lesion Segmentation Ensemble with Diverse Training Strategies

      Laura Canalini, Federico Pollastri, Federico Bolelli, Michele Cancilla, Stefano Allegretti, Costantino Grana

      Abstract

      This paper presents a novel strategy to perform skin lesion segmentation from dermoscopic images. We design an effective segmentation pipeline, and explore several pre-training methods to initialize the features extractor, highlighting how different procedures lead the Convolutional Neural Network (CNN) to focus on different features. An encoder-decoder segmentation CNN is employed to take advantage of each pre-trained features extractor. Experimental results reveal how multiple initialization strategies can be exploited, by means of an ensemble method, to obtain state-of-the-art skin lesion segmentation accuracy.

   3. Hough Based Evolutions for Enhancing Structures in 3D Electron Microscopy

      Kireeti Bodduna, Joachim Weickert, Achilleas S. Frangakis

      Abstract

      Connecting interrupted line-like structures is a frequent problem in image processing. Here we focus on the specific needs that occur in 3D biophysical data analysis in electron microscopy (EM). We introduce a powerful framework for connecting line-like structures in 3D data sets by combining a specific semilocal Hough transform with a directional evolution equation. The Hough transform allows to find the principal orientations of the local structures in a robust way, and the evolution equation is designed as a partial differential equation that smoothes along these principal orientations. We evaluate the performance of our method for enhancing structures in both synthetic and real-world EM data. In contrast to traditional structure tensor based methods such as coherence-enhancing diffusion, our method can handle the missing wedge problem in EM, also known as limited angle tomography problem. A modified version of our approach is also able to tackle the discontinuities created due to the contrast transfer function correction of EM images.

   4. Automated Segmentation of Nanoparticles in BF TEM Images by U-Net Binarization and Branch and Bound

      Sahar Zafari, Tuomas Eerola, Paulo Ferreira, Heikki Kälviäinen, Alan Bovik

      Abstract

      Transmission electron microscopy (TEM) provides information about Inorganic nanoparticles that no other method is able to deliver. Yet, a major task when studying Inorganic nanoparticles using TEM is the automated analysis of the images, i.e. segmentation of individual nanoparticles. The current state-of-the-art methods generally rely on binarization routines that require parameterization, and on methods to segment the overlapping nanoparticles (NPs) using highly idealized nanoparticle shape models. It is unclear, however, that there is any way to determine the best set of parameters providing an optimal segmentation, given the great diversity of NPs characteristics, such as shape and size, that may be encountered. Towards remedying these barriers, this paper introduces a method for segmentation of NPs in Bright Field (BF) TEM images. The proposed method involves three main steps: binarization, contour evidence extraction, and contour estimation. For the binarization, a model based on the U-Net architecture is trained to convert an input image into its binarized version. The contour evidence extraction starts by recovering contour segments from a binarized image using concave contour points detection. The contour segments which belong to the same nanoparticle are grouped in the segment grouping step. The grouping is formulated as a combinatorial optimization problem and solved using the well-known branch and bound algorithm. Finally, the full contours of the NPs are estimated by an ellipse. The experiments on a real-world dataset consisting of 150 BF TEM images containing approximately 2,700 NPs show that the proposed method outperforms five current state-of-art approaches in the overlapping NPs segmentation.

5. Image and Texture Analysis

   1. Frontmatter

   2. A Fractal-Based Approach to Network Characterization Applied to Texture Analysis

      Lucas C. Ribas, Antoine Manzanera, Odemir M. Bruno

      Abstract

      This work proposes a new method for texture analysis that combines fractal descriptors and complex network modeling. At first, the texture image is modeled as a network. Then, the network is converted into a surface where the Cartesian coordinates and the vertex degree is mapped into a 3D point in the surface. Then, we calculate a description vector of this surface using a method inspired by the Bouligand-Minkowski technique for estimating the fractal dimension of a surface. Specifically, the descriptor corresponds to the evolution of the volume occupied by the dilated surface, when the radius of the spherical structuring element increases. The feature vector is given by the concatenation of the volumes of the dilated surface for different radius values. Our proposal is an enhancement of the classic complex networks descriptors, where only the statistical information was considered. Our method was validated on four texture datasets and the results reveal that our method leads to highly discriminative textural features.

   3. Binary Tomography Using Variants of Local Binary Patterns as Texture Priors

      Judit Szűcs, Péter Balázs

      Abstract

      In this paper, we propose a novel approach for binary image reconstruction from few projections. The binary reconstruction problem can be highly underdetermined and one way to reduce the search space of feasible solutions is to exploit some prior knowledge of the image to be reconstructed. We use texture information extracted from sample image patches as prior knowledge. Experimental results show that this approach can retain the structure of the image even if just a very few number of projections are used for the reconstruction.

   4. Volumes of Blurred-Invariant Gaussians for Dynamic Texture Classification

      Thanh Tuan Nguyen, Thanh Phuong Nguyen, Frédéric Bouchara, Ngoc-Son Vu

      Abstract

      An effective model, which jointly captures shape and motion cues, for dynamic texture (DT) description is introduced by taking into account advantages of volumes of blurred-invariant features in three main following stages. First, a 3-dimensional Gaussian kernel is used to form smoothed sequences that allow to deal with well-known limitations of local encoding such as near uniform regions and sensitivity to noise. Second, a receptive volume of the Difference of Gaussians (DoG) is figured out to mitigate the negative impacts of environmental and illumination changes which are major challenges in DT understanding. Finally, a local encoding operator is addressed to construct a discriminative descriptor of enhancing patterns extracted from the filtered volumes. Evaluations on benchmark datasets (i.e., UCLA, DynTex, and DynTex++) for issue of DT classification have positively validated our crucial contributions.