Continual learning for computer vision applications

One of the most visionary goals of Artificial Intelligence is to create a system able to mimic and eventually surpass the intelligence observed in biological systems including, ambitiously, the one observed in humans. The main distinctive strength of humans is their ability to build a deep understanding of the world by learning continuously and drawing from their experiences. This ability, which is found in various degrees in all intelligent biological beings, allows them to adapt and properly react to changes by incrementally expanding and refining their knowledge. Arguably, achieving this ability is one of the main goals of Artificial Intelligence and a cornerstone towards the creation of intelligent artificial agents. Modern Deep Learning approaches allowed researchers and industries to achieve great advancements towards the resolution of many long-standing problems in areas like Computer Vision and Natural Language Processing. However, while this current age of renewed interest in AI allowed for the creation of extremely useful applications, a concerningly limited effort is being directed towards the design of systems able to learn continuously. The biggest problem that hinders an AI system from learning incrementally is the catastrophic forgetting phenomenon. This phenomenon, which was discovered in the 90s, naturally occurs in Deep Learning architectures where classic learning paradigms are applied when learning incrementally from a stream of experiences. This dissertation revolves around the Continual Learning field, a sub-field of Machine Learning research that has recently made a comeback following the renewed interest in Deep Learning approaches. This work will focus on a comprehensive view of continual learning by considering algorithmic, benchmarking, and applicative aspects of this field. This dissertation will also touch on community aspects such as the design and creation of research tools aimed at supporting Continual Learning research, and the theoretical and practical aspects concerning public competitions in this field.

Post-processing and visualization techniques in 2D boundary element analysis

Most post-processors for boundary element (BE) analysis use an auxiliary domain mesh to display domain results, working against the profitable modelling process of a pure boundary discretization. This paper introduces a novel visualization technique which preserves the basic properties of the boundary element methods. The proposed algorithm does not require any domain discretization and is based on the direct and automatic identification of isolines. Another critical aspect of the visualization of domain results in BE analysis is the effort required to evaluate results in interior points. In order to tackle this issue, the present article also provides a comparison between the performance of two different BE formulations (conventional and hybrid). In addition, this paper presents an overview of the most common post-processing and visualization techniques in BE analysis, such as the classical algorithms of scan line and the interpolation over a domain discretization. The results presented herein show that the proposed algorithm offers a very high performance compared with other visualization procedures.

Proposal for DICOM Multiframe Medical Image Integrity and Authenticity

This paper presents a novel algorithm to successfully achieve viable integrity and authenticity addition and verification of n-frame DICOM medical images using cryptographic mechanisms. The aim of this work is the enhancement of DICOM security measures, especially for multiframe images. Current approaches have limitations that should be properly addressed for improved security. The algorithm proposed in this work uses data encryption to provide integrity and authenticity, along with digital signature. Relevant header data and digital signature are used as inputs to cipher the image. Therefore, one can only retrieve the original data if and only if the images and the inputs are correct. The encryption process itself is a cascading scheme, where a frame is ciphered with data related to the previous frames, generating also additional data on image integrity and authenticity. Decryption is similar to encryption, featuring also the standard security verification of the image. The implementation was done in JAVA, and a performance evaluation was carried out comparing the speed of the algorithm with other existing approaches. The evaluation showed a good performance of the algorithm, which is an encouraging result to use it in a real environment.

Characterisation of the three-dimensional structure of earthworm burrow systems using image analysis and mathematical morphology

The aim of this work was to exemplify the specific contribution of both two- and three-dimensional (31)) X-ray computed tomography to characterise earthworm burrow systems. To achieve this purpose we used 3D mathematical morphology operators to characterise burrow systems resulting from the activity of an anecic (Aporrectodea noctunia), and an endogeic species (Allolobophora chlorotica), when both species were introduced either separately or together into artificial soil cores. Images of these soil cores were obtained using a medical X-ray tomography scanner. Three-dimensional reconstructions of burrow systems were obtained using a specifically developed segmentation algorithm. To study the differences between burrow systems, a set of classical tools of mathematical morphology (granulometries) were used. So-called granulometries based on different structuring elements clearly separated the different burrow systems. They enabled us to show that burrows made by the anecic species were fatter, longer, more vertical, more continuous but less sinuous than burrows of the endogeic species. The granulometry transform of the soil matrix showed that burrows made by A. nocturna were more evenly distributed than those of A. chlorotica. Although a good discrimination was possible when only one species was introduced into the soil cores, it was not possible to separate burrows of the two species from each other in cases where species were introduced into the same soil core. This limitation, partly due to the insufficient spatial resolution of the medical scanner, precluded the use of the morphological operators to study putative interactions between the two species.

Image reconstruction affects computer tomographic assessment of lung hyperinflation

Objectives: Lung hyperinflation may be assessed by computed tomography (CT). As shown for patients with emphysema, however, CT image reconstruction affects quantification of hyperinflation. We studied the impact of reconstruction parameters on hyperinflation measurements in mechanically ventilated (MV) patients. Design: Observational analysis. Setting: A University hospital-affiliated research Unit. Patients: The patients were MV patients with injured (n = 5) or normal lungs (n = 6), and spontaneously breathing patients (n = 5). Interventions: None. Measurements and results: Eight image series involving 3, 5, 7, and 10 mm slices and standard and sharp filters were reconstructed from identical CT raw data. Hyperinflated (V-hyper), normally (V-normal), poorly (V-poor), and nonaerated (V-non) volumes were calculated by densitometry as percentage of total lung volume (V-total). V-hyper obtained with the sharp filter systematically exceeded that with the standard filter showing a median (interquartile range) increment of 138 (62-272) ml corresponding to approximately 4% of V-total. In contrast, sharp filtering minimally affected the other subvolumes (V-normal, V-poor, V-non, and V-total). Decreasing slice thickness also increased V-hyper significantly. When changing from 10 to 3 mm thickness, V-hyper increased by a median value of 107 (49-252) ml in parallel with a small and inconsistent increment in V-non of 12 (7-16) ml. Conclusions: Reconstruction parameters significantly affect quantitative CT assessment of V-hyper in MV patients. Our observations suggest that sharp filters are inappropriate for this purpose. Thin slices combined with standard filters and more appropriate thresholds (e.g., -950 HU in normal lungs) might improve the detection of V-hyper. Different studies on V-hyper can only be compared if identical reconstruction parameters were used.

Effect of image analysis software on neurofunctional activation during processing of emotional human faces

Functional brain imaging techniques such as functional MRI (fMRI) that allow the in vivo investigation of the human brain have been exponentially employed to address the neurophysiological substrates of emotional processing. Despite the growing number of fMRI studies in the field, when taken separately these individual imaging studies demonstrate contrasting findings and variable pictures, and are unable to definitively characterize the neural networks underlying each specific emotional condition. Different imaging packages, as well as the statistical approaches for image processing and analysis, probably have a detrimental role by increasing the heterogeneity of findings. In particular, it is unclear to what extent the observed neurofunctional response of the brain cortex during emotional processing depends on the fMRI package used in the analysis. In this pilot study, we performed a double analysis of an fMRI dataset using emotional faces. The Statistical Parametric Mapping (SPM) version 2.6 (Wellcome Department of Cognitive Neurology, London, UK) and the XBAM 3.4 (Brain Imaging Analysis Unit, Institute of Psychiatry, Kings College London, UK) programs, which use parametric and non-parametric analysis, respectively, were used to assess our results. Both packages revealed that processing of emotional faces was associated with an increased activation in the brain`s visual areas (occipital, fusiform and lingual gyri), in the cerebellum, in the parietal cortex, in the cingulate cortex (anterior and posterior cingulate), and in the dorsolateral and ventrolateral prefrontal cortex. However, blood oxygenation level-dependent (BOLD) response in the temporal regions, insula and putamen was evident in the XBAM analysis but not in the SPM analysis. Overall, SPM and XBAM analyses revealed comparable whole-group brain responses. Further Studies are needed to explore the between-group compatibility of the different imaging packages in other cognitive and emotional processing domains. (C) 2009 Elsevier Ltd. All rights reserved.

An image processing application for quantification of protein aggregates in Caenorhabditis elegans

Protein aggregation became a widely accepted marker of many polyQ disorders, including Machado-Joseph disease (MJD), and is often used as readout for disease progression and development of therapeutic strategies. The lack of good platforms to rapidly quantify protein aggregates in a wide range of disease animal models prompted us to generate a novel image processing application that automatically identifies and quantifies the aggregates in a standardized and operator-independent manner. We propose here a novel image processing tool to quantify the protein aggregates in a Caenorhabditis elegans (C. elegans) model of MJD. Confocal mi-croscopy images were obtained from animals of different genetic conditions. The image processing application was developed using MeVisLab as a platform to pro-cess, analyse and visualize the images obtained from those animals. All segmenta-tion algorithms were based on intensity pixel levels.The quantification of area or numbers of aggregates per total body area, as well as the number of aggregates per animal were shown to be reliable and reproducible measures of protein aggrega-tion in C. elegans. The results obtained were consistent with the levels of aggrega-tion observed in the images. In conclusion, this novel imaging processing applica-tion allows the non-biased, reliable and high throughput quantification of protein aggregates in a C. elegans model of MJD, which may contribute to a significant improvement on the prognosis of treatment effectiveness for this group of disor-ders

Road lane detection for autonomous robot guidance

Computational Vision stands as the most comprehensive way of knowing the surrounding environment. Accordingly to that, this study aims to present a method to obtain from a common webcam, environment information to guide a mobile differential robot through a path similar to a roadway.

Road lane detection for autonomous robot guidance

Computational Vision stands as the most comprehensive way of knowing the surrounding environment. Accordingly to that, this study aims to present a method to obtain from a common webcam, environment information to guide a mobile differential robot through a path similar to a roadway.

A comparison of Sinogram Affirmed Iterative Reconstruction and filtered back projection on image quality and dose reduction in paediatric head CT: a phantom study

Background: Computed tomography (CT) is one of the most used modalities for diagnostics in paediatric populations, which is a concern as it also delivers a high patient dose. Research has focused on developing computer algorithms that provide better image quality at lower dose. The iterative reconstruction algorithm Sinogram-Affirmed Iterative Reconstruction (SAFIRE) was introduced as a new technique that reduces noise to increase image quality. Purpose: The aim of this study is to compare SAFIRE with the current gold standard, Filtered Back Projection (FBP), and assess whether SAFIRE alone permits a reduction in dose while maintaining image quality in paediatric head CT. Methods: Images were collected using a paediatric head phantom using a SIEMENS SOMATOM PERSPECTIVE 128 modulated acquisition. 54 images were reconstructed using FBP and 5 different strengths of SAFIRE. Objective measures of image quality were determined by measuring SNR and CNR. Visual measures of image quality were determined by 17 observers with different radiographic experiences. Images were randomized and displayed using 2AFC; observers scored the images answering 5 questions using a Likert scale. Results: At different dose levels, SAFIRE significantly increased SNR (up to 54%) in the acquired images compared to FBP at 80kVp (5.2-8.4), 110kVp (8.2-12.3), 130kVp (8.8-13.1). Visual image quality was higher with increasing SAFIRE strength. The highest image quality was scored with SAFIRE level 3 and higher. Conclusion: The SAFIRE algorithm is suitable for image noise reduction in paediatric head CT. Our data demonstrates that SAFIRE enhances SNR while reducing noise with a possible reduction of dose of 68%.

People counting system using existing surveillance video camera

The Casa da Música Foundation, responsible for the management of Casa da Música do Porto building, has the need to obtain statistical data related to the number of building’s visitors. This information is a valuable tool for the elaboration of periodical reports concerning the success of this cultural institution. For this reason it was necessary to develop a system capable of returning the number of visitors for a requested period of time. This represents a complex task due to the building’s unique architectural design, characterized by very large doors and halls, and the sudden large number of people that pass through them in moments preceding and proceeding the different activities occurring in the building. To achieve the technical solution for this challenge, several image processing methods, for people detection with still cameras, were first studied. The next step was the development of a real time algorithm, using OpenCV libraries and computer vision concepts,to count individuals with the desired accuracy. This algorithm includes the scientific and technical knowledge acquired in the study of the previous methods. The themes developed in this thesis comprise the fields of background maintenance, shadow and highlight detection, and blob detection and tracking. A graphical interface was also built, to help on the development, test and tunning of the proposed system, as a complement to the work. Furthermore, tests to the system were also performed, to certify the proposed techniques against a set of limited circumstances. The results obtained revealed that the algorithm was successfully applied to count the number of people in complex environments with reliable accuracy.