Person Re-Identification using Deep Convnets with Multi-task Learning

Person re-identification involves recognizing a person across non-overlapping camera views, with different pose, illumination, and camera characteristics. We propose to tackle this problem by training a deep convolutional network to represent a person’s appearance as a low-dimensional feature vector that is invariant to common appearance variations encountered in the re-identification problem. Specifically, a Siamese-network architecture is used to train a feature extraction network using pairs of similar and dissimilar images. We show that use of a novel multi-task learning objective is crucial for regularizing the network parameters in order to prevent over-fitting due to the small size the training dataset. We complement the verification task, which is at the heart of re-identification, by training the network to jointly perform verification, identification, and to recognise attributes related to the clothing and pose of the person in each image. Additionally, we show that our proposed approach performs well even in the challenging cross-dataset scenario, which may better reflect real-world expected performance. 

Non-contact Pressure-based Sleep/Wake Discrimination

Poor sleep is increasingly being recognised as an important prognostic parameter of health. For those with suspected sleep disorders, patients are referred to sleep clinics which guide treatment. However, sleep clinics are not always a viable option due to their high cost, a lack of experienced practitioners, lengthy waiting lists and an unrepresentative sleeping environment. A home-based non-contact sleep/wake monitoring system may be used as a guide for treatment potentially stratifying patients by clinical need or highlighting longitudinal changes in sleep and nocturnal patterns. This paper presents the evaluation of an under-mattress sleep monitoring system for non-contact sleep/wake discrimination. A large dataset of sensor data with concomitant sleep/wake state was collected from both younger and older adults participating in a circadian sleep study. A thorough training/testing/validation procedure was configured and optimised feature extraction and sleep/wake discrimination algorithms evaluated both within and across the two cohorts. An accuracy, sensitivity and specificity of 74.3%, 95.5%, and 53.2% is reported over all subjects using an external validation
dataset (71.9%, 87.9% and 56%, and 77.5%, 98% and 57% is reported for younger and older subjects respectively). These results compare favourably with similar research, however this system provides an ambient alternative suitable for long term continuous sleep monitoring, particularly amongst vulnerable populations.

Análise de sinais cerebrais utilizando árvores de decisão

Este trabalho propõe um estudo de sinais cerebrais aplicados em sistemas BCI (Brain-Computer Interface - Interfaces Cérebro Computador), através do uso de Árvores de Decisão e da análise dessas árvores com base nas Neurociências. Para realizar o tratamento dos dados são necessárias 5 fases: aquisição de dados, pré-processamento, extração de características, classificação e validação. Neste trabalho, todas as fases são contempladas. Contudo, enfatiza-se as fases de classificação e de validação. Na classificação utiliza-se a técnica de Inteligência Artificial denominada Árvores de Decisão. Essa técnica é reconhecida na literatura como uma das formas mais simples e bem sucedidas de algoritmos de aprendizagem. Já a fase de validação é realizada nos estudos baseados na Neurociência, que é um conjunto das disciplinas que estudam o sistema nervoso, sua estrutura, seu desenvolvimento, funcionamento, evolução, relação com o comportamento e a mente, e também suas alterações. Os resultados obtidos neste trabalho são promissores, mesmo sendo iniciais, visto que podem melhor explicar, com a utilização de uma forma automática, alguns processos cerebrais.

Datasets, features, learning, and models in visual recognition

Visual recognition is a fundamental research topic in computer vision. This dissertation explores datasets, features, learning, and models used for visual recognition. In order to train visual models and evaluate different recognition algorithms, this dissertation develops an approach to collect object image datasets on web pages using an analysis of text around the image and of image appearance. This method exploits established online knowledge resources (Wikipedia pages for text; Flickr and Caltech data sets for images). The resources provide rich text and object appearance information. This dissertation describes results on two datasets. The first is Berg’s collection of 10 animal categories; on this dataset, we significantly outperform previous approaches. On an additional set of 5 categories, experimental results show the effectiveness of the method. Images are represented as features for visual recognition. This dissertation introduces a text-based image feature and demonstrates that it consistently improves performance on hard object classification problems. The feature is built using an auxiliary dataset of images annotated with tags, downloaded from the Internet. Image tags are noisy. The method obtains the text features of an unannotated image from the tags of its k-nearest neighbors in this auxiliary collection. A visual classifier presented with an object viewed under novel circumstances (say, a new viewing direction) must rely on its visual examples. This text feature may not change, because the auxiliary dataset likely contains a similar picture. While the tags associated with images are noisy, they are more stable when appearance changes. The performance of this feature is tested using PASCAL VOC 2006 and 2007 datasets. This feature performs well; it consistently improves the performance of visual object classifiers, and is particularly effective when the training dataset is small. With more and more collected training data, computational cost becomes a bottleneck, especially when training sophisticated classifiers such as kernelized SVM. This dissertation proposes a fast training algorithm called Stochastic Intersection Kernel Machine (SIKMA). This proposed training method will be useful for many vision problems, as it can produce a kernel classifier that is more accurate than a linear classifier, and can be trained on tens of thousands of examples in two minutes. It processes training examples one by one in a sequence, so memory cost is no longer the bottleneck to process large scale datasets. This dissertation applies this approach to train classifiers of Flickr groups with many group training examples. The resulting Flickr group prediction scores can be used to measure image similarity between two images. Experimental results on the Corel dataset and a PASCAL VOC dataset show the learned Flickr features perform better on image matching, retrieval, and classification than conventional visual features. Visual models are usually trained to best separate positive and negative training examples. However, when recognizing a large number of object categories, there may not be enough training examples for most objects, due to the intrinsic long-tailed distribution of objects in the real world. This dissertation proposes an approach to use comparative object similarity. The key insight is that, given a set of object categories which are similar and a set of categories which are dissimilar, a good object model should respond more strongly to examples from similar categories than to examples from dissimilar categories. This dissertation develops a regularized kernel machine algorithm to use this category dependent similarity regularization. Experiments on hundreds of categories show that our method can make significant improvement for categories with few or even no positive examples.

CONCURRENT LOCALIZATION AND MAPPING WITH SONAR SENSORS AND CONSIDERATION OF VEHICLE MOTION

Simultaneous Localization and Mapping (SLAM) is a procedure used to determine the location of a mobile vehicle in an unknown environment, while constructing a map of the unknown environment at the same time. Mobile platforms, which make use of SLAM algorithms, have industrial applications in autonomous maintenance, such as the inspection of flaws and defects in oil pipelines and storage tanks. A typical SLAM consists of four main components, namely, experimental setup (data gathering), vehicle pose estimation, feature extraction, and filtering. Feature extraction is the process of realizing significant features from the unknown environment such as corners, edges, walls, and interior features. In this work, an original feature extraction algorithm specific to distance measurements obtained through SONAR sensor data is presented. This algorithm has been constructed by combining the SONAR Salient Feature Extraction Algorithm and the Triangulation Hough Based Fusion with point-in-polygon detection. The reconstructed maps obtained through simulations and experimental data with the fusion algorithm are compared to the maps obtained with existing feature extraction algorithms. Based on the results obtained, it is suggested that the proposed algorithm can be employed as an option for data obtained from SONAR sensors in environment, where other forms of sensing are not viable. The algorithm fusion for feature extraction requires the vehicle pose estimation as an input, which is obtained from a vehicle pose estimation model. For the vehicle pose estimation, the author uses sensor integration to estimate the pose of the mobile vehicle. Different combinations of these sensors are studied (e.g., encoder, gyroscope, or encoder and gyroscope). The different sensor fusion techniques for the pose estimation are experimentally studied and compared. The vehicle pose estimation model, which produces the least amount of error, is used to generate inputs for the feature extraction algorithm fusion. In the experimental studies, two different environmental configurations are used, one without interior features and another one with two interior features. Numerical and experimental findings are discussed. Finally, the SLAM algorithm is implemented along with the algorithms for feature extraction and vehicle pose estimation. Three different cases are experimentally studied, with the floor of the environment intentionally altered to induce slipping. Results obtained for implementations with and without SLAM are compared and discussed. The present work represents a step towards the realization of autonomous inspection platforms for performing concurrent localization and mapping in harsh environments.

A contribution to semantic description of images and videos: an application of soft biometrics

Humans have a high ability to extract visual data information acquired by sight. Trought a learning process, which starts at birth and continues throughout life, image interpretation becomes almost instinctively. At a glance, one can easily describe a scene with reasonable precision, naming its main components. Usually, this is done by extracting low-level features such as edges, shapes and textures, and associanting them to high level meanings. In this way, a semantic description of the scene is done. An example of this, is the human capacity to recognize and describe other people physical and behavioral characteristics, or biometrics. Soft-biometrics also represents inherent characteristics of human body and behaviour, but do not allow unique person identification. Computer vision area aims to develop methods capable of performing visual interpretation with performance similar to humans. This thesis aims to propose computer vison methods which allows high level information extraction from images in the form of soft biometrics. This problem is approached in two ways, unsupervised and supervised learning methods. The first seeks to group images via an automatic feature extraction learning , using both convolution techniques, evolutionary computing and clustering. In this approach employed images contains faces and people. Second approach employs convolutional neural networks, which have the ability to operate on raw images, learning both feature extraction and classification processes. Here, images are classified according to gender and clothes, divided into upper and lower parts of human body. First approach, when tested with different image datasets obtained an accuracy of approximately 80% for faces and non-faces and 70% for people and non-person. The second tested using images and videos, obtained an accuracy of about 70% for gender, 80% to the upper clothes and 90% to lower clothes. The results of these case studies, show that proposed methods are promising, allowing the realization of automatic high level information image annotation. This opens possibilities for development of applications in diverse areas such as content-based image and video search and automatica video survaillance, reducing human effort in the task of manual annotation and monitoring.

Método para reconhecimento de vogais e extração de parâmetros acústicos para analises forenses

Forensic speaker comparison exams have complex characteristics, demanding a long time for manual analysis. A method for automatic recognition of vowels, providing feature extraction for acoustic analysis is proposed, aiming to contribute as a support tool in these exams. The proposal is based in formant measurements by LPC (Linear Predictive Coding), selectively by fundamental frequency detection, zero crossing rate, bandwidth and continuity, with the clustering being done by the k-means method. Experiments using samples from three different databases have shown promising results, in which the regions corresponding to five of the Brasilian Portuguese vowels were successfully located, providing visualization of a speaker’s vocal tract behavior, as well as the detection of segments corresponding to target vowels.

Modelling human behaviour and event detection from videos

Automatic analysis of human behaviour in large collections of videos is gaining interest, even more so with the advent of file sharing sites such as YouTube. However, challenges still exist owing to several factors such as inter- and intra-class variations, cluttered backgrounds, occlusion, camera motion, scale, view and illumination changes. This research focuses on modelling human behaviour for action recognition in videos. The developed techniques are validated on large scale benchmark datasets and applied on real-world scenarios such as soccer videos. Three major contributions are made. The first contribution is in the area of proper choice of a feature representation for videos. This involved a study of state-of-the-art techniques for action recognition, feature extraction processing and dimensional reduction techniques so as to yield the best performance with optimal computational requirements. Secondly, temporal modelling of human behaviour is performed. This involved frequency analysis and temporal integration of local information in the video frames to yield a temporal feature vector. Current practices mostly average the frame information over an entire video and neglect the temporal order. Lastly, the proposed framework is applied and further adapted to real-world scenario such as soccer videos. A dataset consisting of video sequences depicting events of players falling is created from actual match data to this end and used to experimentally evaluate the proposed framework.

EIR – sistema inteligente para detecção e classificação de calcificações em mamografias

Dissertação (mestrado)—Universidade de Brasília, Faculdade Gama, Programa de Pós-Graduação em Engenharia Biomédica, 2016.

Design and Pilot Study for an Efficient High-Throughput Automated Computer-Vision Guided Intelligent De-Calyxing Machine for Post-Harvest Strawberry Processing

Strawberries harvested for processing as frozen fruits are currently de-calyxed manually in the field. This process requires the removal of the stem cap with green leaves (i.e. the calyx) and incurs many disadvantages when performed by hand. Not only does it necessitate the need to maintain cutting tool sanitation, but it also increases labor time and exposure of the de-capped strawberries before in-plant processing. This leads to labor inefficiency and decreased harvest yield. By moving the calyx removal process from the fields to the processing plants, this new practice would reduce field labor and improve management and logistics, while increasing annual yield. As labor prices continue to increase, the strawberry industry has shown great interest in the development and implementation of an automated calyx removal system. In response, this dissertation describes the design, operation, and performance of a full-scale automatic vision-guided intelligent de-calyxing (AVID) prototype machine. The AVID machine utilizes commercially available equipment to produce a relatively low cost automated de-calyxing system that can be retrofitted into existing food processing facilities. This dissertation is broken up into five sections. The first two sections include a machine overview and a 12-week processing plant pilot study. Results of the pilot study indicate the AVID machine is able to de-calyx grade-1-with-cap conical strawberries at roughly 66 percent output weight yield at a throughput of 10,000 pounds per hour. The remaining three sections describe in detail the three main components of the machine: a strawberry loading and orientation conveyor, a machine vision system for calyx identification, and a synchronized multi-waterjet knife calyx removal system. In short, the loading system utilizes rotational energy to orient conical strawberries. The machine vision system determines cut locations through RGB real-time feature extraction. The high-speed multi-waterjet knife system uses direct drive actuation to locate 30,000 psi cutting streams to precise coordinates for calyx removal. Based on the observations and studies performed within this dissertation, the AVID machine is seen to be a viable option for automated high-throughput strawberry calyx removal. A summary of future tasks and further improvements is discussed at the end.

Detection and Grading of Prostate Cancer in Temporal Enhanced Ultrasound using Convolutional Neural Networks

Prostate cancer is the most common non-dermatological cancer amongst men in the developed world. The current definitive diagnosis is core needle biopsy guided by transrectal ultrasound. However, this method suffers from low sensitivity and specificity in detecting cancer. Recently, a new ultrasound based tissue typing approach has been proposed, known as temporal enhanced ultrasound (TeUS). In this approach, a set of temporal ultrasound frames is collected from a stationary tissue location without any intentional mechanical excitation. The main aim of this thesis is to implement a deep learning-based solution for prostate cancer detection and grading using TeUS data. In the proposed solution, convolutional neural networks are trained to extract high-level features from time domain TeUS data in temporally and spatially adjacent frames in nine in vivo prostatectomy cases. This approach avoids information loss due to feature extraction and also improves cancer detection rate. The output likelihoods of two TeUS arrangements are then combined to form our novel decision support system. This deep learning-based approach results in the area under the receiver operating characteristic curve (AUC) of 0.80 and 0.73 for prostate cancer detection and grading, respectively, in leave-one-patient-out cross-validation. Recently, multi-parametric magnetic resonance imaging (mp-MRI) has been utilized to improve detection rate of aggressive prostate cancer. In this thesis, for the first time, we present the fusion of mp-MRI and TeUS for characterization of prostate cancer to compensates the deficiencies of each image modalities and improve cancer detection rate. The results obtained using TeUS are fused with those attained using consolidated mp-MRI maps from multiple MR modalities and cancer delineations on those by multiple clinicians. The proposed fusion approach yields the AUC of 0.86 in prostate cancer detection. The outcomes of this thesis emphasize the viable potential of TeUS as a tissue typing method. Employing this ultrasound-based intervention, which is non-invasive and inexpensive, can be a valuable and practical addition to enhance the current prostate cancer detection.

A Case Based Methodology for Problem Solving aiming at Knee Osteoarthritis Detection

Knee osteoarthritis is the most common type of arthritis and a major cause of impaired mobility and disability for the ageing populations. Therefore, due to the increasing prevalence of the malady, it is expected that clinical and scientific practices had to be set in order to detect the problem in its early stages. Thus, this work will be focused on the improvement of methodologies for problem solving aiming at the development of Artificial Intelligence based decision support system to detect knee osteoarthritis. The framework is built on top of a Logic Programming approach to Knowledge Representation and Reasoning, complemented with a Case Based approach to computing that caters for the handling of incomplete, unknown, or even self-contradictory information.

Processamento e classificação de imagens biológicas: aplicação à histologia

A Histologia, o estudo de tecidos, é uma das áreas fundamentais da Biologia que permitiu enormes avanços científicos. Sendo uma tarefa exigente, meticulosa e demorada, será importante aproveitar a existência de ferramentas e algoritmos computacionais no seu auxílio, tornando o processo mais rápido e possibilitando a descoberta de informação que poderá não estar visível à partida. Esta dissertação tem como principal objectivo averiguar se um animal foi ou não sujeito à ingestão de um xenobiótico. Com esse objectivo em vista, utilizaram-se técnicas de processamento e segmentação de imagem aplicadas a imagens de tecido renal de ratos saudáveis e ratos que ingeriram o xenobiótico. Destas imagens extraíram-se inúmeras características do corpúsculo renal que após serem analisadas através de vários algoritmos de classificação mostraram ser possível saber se o animal ingeriu ou não o xenobiótico, com um reduzido grau de incerteza. ABSTRACT: Histology, the study of tissues, is one of the key areas of Biology that has allowed huge advances in Science. Being a demanding, meticulous and time consuming task, it is important to use the existence of computational tools and algorithms in its aid, making the process faster and enabling the discovery of information that may not be initially visible. The main goal of this thesis is to ascertain if an animal was subjected or not to the ingestion of a xenobiotic. With this in mind, were used image processing and segmentation techniques applied on images of kidney tissue from healthy rats and rats that ingested the xenobiotic. From these images were extracted several features of renal glomeruli that after being analyzed by various classification algorithms had shown to be possible to know, with an acceptable degree of certainty, if the animal ingested or not the xenobiotic.