Preclinical research for developing new drugs is a long and expensive procedure. Experiments relying on image acquisition and analysis tend to be low throughput and use reporter systems that may influence the studied cells. With image-based assays focusing on extracting qualitative information from microscopic images of mammalian cells, more cost-efficient and high-throughput analyses are possible. Furthermore, studying cell morphology has proven to be a good indicator of cell phenotype. Using hand-crafted feature descriptors based on cell morphology, label-free quantification of cell apoptosis has been achieved. These hand-crafted descriptors are based on cell characteristics translated to quantifiable metrics, but risk being biased towards easily observable features and therefore miss subtle ones.         This project proposes an alternative approach by generating a latent representation of cell features using deep learning models and aims to find if they can compete with pre-defined hand-crafted representations in classifying live or dead cells. For this purpose, three deep learning models are implemented, one autoencoder and two variational-autoencoder. We develop a core architecture shared between the models based on a convolutional neural network using a latent space with 16 dimensions. We then train the models to recreate single-cell images of SKOV3 ovarian cancer cells. The latent representation was extracted at specific checkpoints during training and later used for training a logistic regression classifier. Finally, comparing classification accuracy between the hand-crafted feature representations and generated representation was made with novel cell images. The generated representations show a slight but consistent increase in classification accuracy, up to 4.9 percent points, even without capturing all morphological details in the recreation. Thus, we conclude that it is possible for generated representations to outperform hand-crafted feature descriptors in live or dead cell classification.

Indiana University-Purdue University Indianapolis (IUPUI) / Domain adaptation is one of the hottest directions in solving annotation insufficiency problem of deep learning. General domain adaptation is not consistent with the practical scenarios in the industry. In this thesis, we focus on two concerns as below. First is that labeled data are generally collected from multiple domains. In other words, multi-source adaptation is a more common situation. Simply extending these single-source approaches to the multi-source cases could cause sub-optimal inference, so specialized multi-source adaptation methods are essential. The main challenge in the multi-source scenario is a more complex divergence situation. Not only the divergence between target and each source plays a role, but the divergences among distinct sources matter as well. However, the significance of maintaining consistency among multiple sources didn't gain enough attention in previous work. In this thesis, we propose an Enhanced Consistency Multi-Source Adaptation (EC-MSA) framework to address it from three perspectives. First, we mitigate feature-level discrepancy by cross-domain conditional alignment, narrowing the divergence between each source and target domain class-wisely. Second, we enhance multi-source consistency via dual mix-up, diminishing the disagreements among different sources. Third, we deploy a target distilling mechanism to handle the uncertainty of target prediction, aiming to provide high-quality pseudo-labeled target samples to benefit the previous two aspects. Extensive experiments are conducted on several common benchmark datasets and demonstrate that our model outperforms the state-of-the-art methods. Second is that data privacy and security is necessary in practice. That is, we hope to keep the raw data stored locally while can still obtain a satisfied model. In such a case, the risk of data leakage greatly decreases. Therefore, it is natural for us to combine the federated learning paradigm with domain adaptation. Under the source-private setting, the main challenge for us is to expose information from the source domain to the target domain while make sure that the communication process is safe enough. In this thesis, we propose a method named Fourier Transform-Assisted Federated Domain Adaptation (FTA-FDA) to alleviate the difficulties in two ways. We apply Fast Fourier Transform to the raw data and transfer only the amplitude spectra during the communication. Then frequency space interpolations between these two domains are conducted, minimizing the discrepancies while ensuring the contact of them and keeping raw data safe. What's more, we make prototype alignments by using the model weights together with target features, trying to reduce the discrepancy in the class level. Experiments on Office-31 demonstrate the effectiveness and competitiveness of our approach, and further analyses prove that our algorithm can help protect privacy and security.

Recently, the deep neural network structure caps-net was proposed by Sabouret al. [11]. Capsule networks are designed to learn relative geometry betweenthe features of a layer and the features of the next layer. The Capsule network’smain building blocks are capsules, which are represented by vectors. The ideais that each capsule will represent a feature as well as traits or subfeatures ofthat feature. This allows for smart information routing. Capsules traits are usedto predict the traits of the capsules in the next layer, and information is sent toto next layer capsules on which the predictions agree. This is called routing byagreement.This thesis investigates theoretical support of new and existing routing al-gorithms as well as evaluates their performance on the MNIST [16] and CIFAR-10 [8] datasets. A variation of the dynamic routing algorithm presented in theoriginal paper [11] achieved the highest accuracy and fastest execution time.

Every day robots are becoming more common in the society. Consequently, they must have certain basic skills in order to interact with humans and the environment. One of these skills is the capacity to understand the places where they are able to move. Computer vision is one of the ways commonly used for achieving this purpose. Current technologies in this field offer outstanding solutions applied to improve data quality every day, therefore producing more accurate results in the analysis of an environment. With this in mind, the main goal of this research is to develop and validate an efficient object-based scene understanding method that will be able to help solve problems related to scene identification for mobile robotics. We seek to analyze state-of-the-art methods for finding the most suitable one for our goals, as well as to select the kind of data most convenient for dealing with this issue. Another primary goal of the research is to determine the most suitable data input for analyzing scenes in order to find an accurate representation for the scenes by meaning of semantic labels or point cloud features descriptors. As a secondary goal we will show the benefits of using semantic descriptors generated with pre-trained models for mapping and scene classification problems, as well as the use of deep learning models in conjunction with 3D features description procedures to build a 3D object classification model that is directly related with the representation goal of this work. The research described in this thesis was motivated by the need for a robust system capable of understanding the locations where a robot usually interacts. In the same way, the advent of better computational resources has allowed to implement some already defined techniques that demand high computational capacity and that offer a possible solution for dealing with scene understanding issues. One of these techniques are Convolutional Neural Networks (CNNs). These networks have the capacity of classifying an image based on their visual appearance. Then, they generate a list of lexical labels and the probability for each label, representing the likelihood of the present of an object in the scene. Labels are derived from the training sets that the networks learned to recognize. Therefore, we could use this list of labels and probabilities as an efficient representation of the environment and then assign a semantic category to the regions where a mobile robot is able to navigate, and at the same time construct a semantic or topological map based on this semantic representation of the place. After analyzing the state-of-the-art in Scene Understanding, we identified a set of approaches in order to develop a robust scene understanding procedure. Among these approaches we identified an almost unexplored gap in the topic of understanding scenes based on objects present in them. Consequently, we propose to perform an experimental study in this approach aimed at finding a way of fully describing a scene considering the objects lying in place. As the Scene Understanding task involves object detection and annotation, one of the first steps is to determine the kind of data to use as input data in our proposal. With this in mind, our proposal considers to evaluate the use of 3D data. This kind of data suffers from the presence of noise, therefore, we propose to use the Growing Neural Gas (GNG) algorithm to reduce noise effect in the object recognition procedure. GNGs have the capacity to grow and adapt their topology to represent 2D information, producing a smaller representation with a slight noise influence from the input data. Applied to 3D data, the GNG presents a good approach able to tackle with noise. However, using 3D data poses a set of problems such as the lack of a 3D object dataset with enough models to generalize methods and adapt them to real situations, as well as the fact that processing three-dimensional data is computationally expensive and requires a huge storage space. These problems led us to explore new approaches for developing object recognition tasks. Therefore, considering the outstanding results obtained by the CNNs in the latest ImageNet challenge, we propose to carry out an evaluation of the former as an object detection system. These networks were initially proposed in the 90s and are nowadays easily implementable due to hardware improvements in the recent years. CNNs have shown satisfying results when they tested in problems such as: detection of objects, pedestrians, traffic signals, sound waves classification, and for medical image processing, among others. Moreover, an aggregate value of CNNs is the semantic description capabilities produced by the categories/labels that the network is able to identify and that could be translated as a semantic explanation of the input image. Consequently, we propose using the evaluation of these semantic labels as a scene descriptor for building a supervised scene classification model. Having said that, we also propose using semantic descriptors to generate topological maps and test the description capabilities of lexical labels. In addition, semantic descriptors could be suitable for unsupervised places or environment labeling, so we propose using them to deal with this kind of problem in order to achieve a robust scene labeling method. Finally, for tackling the object recognition problem we propose to develop an experimental study for unsupervised object labeling. This will be applied to the objects present in a point cloud and labeled using a lexical labeling tool. Then, objects will be used as the training instances of a classifier mixing their 3D features with label assigned by the external tool.

Traffic is one of the key elements nowadays that affect our lives more or less in a every day basis. Traffic is present when we go to work, is on week ends, on holidays, even if we go shopping in our neighborhood, traffic is present. Every year, we can see on TV how after a bank holiday (United Kingdom public holiday), the traffic incidents are a figure that all TV news are obliged to report. If we see the accident causes, the ones caused by mechanical failures are always a minimal part, being human causes the majority of times. In our society, the tasks where technology help us to complete those with 100% success are very frequent. We tune our TVs for digital broadcasting, robots complete thousands of mechanical tasks, we work with computers that do not crash for months or years, even weather forecasting is more accurate than ever. All those aspects in our lives are successfully carried out on a daily basis. Nowadays, in traffic and road transport, we are starting a new era where driving a vehicle can be assisted partially or totally, parking our car can be done automatically, or even detecting a child in the middle of the road can be automatically done instead of leaving those tasks to the prone-to-fail human. The same features that today amaze us (as in the past did the TV broadcast in colour), in the future, those safety features will be a common thing in our cars. With more and more vehicles in the roads, cars, motorbikes, bicycles, more people in our cities and the necessity to be in a constant move, our society needs a zero-car-accidents conception, as we have now the technology to achieve it. Computer Vision is the computer science field that since the 80s has been obsessed with emulating the way human see and perceive their environment and react to it in an intelligent way. One decade ago, detecting complex objects in a scene as a human was impossible. All we could do was to detect the edges of an object, to threshold pixel values, detect motion, but nothing as the human capability to detect objects and identify their location. The advance in GPUs technology and the development of neural networks in the computer vision community has made those impossible tasks possible. GPUs now being a commodity item in our lives, the increase of amount and speed of RAM and the new and open models developed by experts in neural networks, make the task of detecting a child in the middle of a road a reality. In particular, detections with 99.79% probability are now possible, and the 100% probability goal is becoming a closer reality. In this thesis we have approached one of the key safety features in systems for traffic analysis, that is monitoring pedestrian crossing. After researching the state-of-the-art in pedestrian movement detection, we have presented a novel strategy for such detection. By locating a fixed camera in a place where pedestrians move, we are able to detect the movement of those and their direction. We have achieved that task by using a mix of old and new methodologies. Having a fixed camera, allow us to remove the background of the scene, only leaving the moving pedestrians. Once we have this information, we have created a dataset of moving people and trained a CNN able to detect in which direction the pedestrian is moving. Another work that we present in this thesis is a traffic dataset and the research with state-of.the-art CNN models to detect objects in traffic environments. Crucial objects like cars, people, bikes, motorbikes, traffic signals, etc. have been grouped in a novel dataset to feed state-of-the-art CNNs and we carried out an analysis about their ability to detect and to locate those objects from the car point of view. Moreover, with the help of tracking techniques, we improved efficiency and robustness of the proposed method, creating a system capable of performing real-time object detection (urban objects). In this thesis, we also present a traffic sign dataset, which comprises 45 different traffic signs. This dataset has been used for training a traffic sign classifier that is used a second step of our urban object detector. Moreover, a very basic but important aspect in safety driving is to keep the vehicle within the allowed space in the road (within the limits of the road). SLAM techniques have been used in the past for such tasks, but we present an end-to-end approach, where a CNN model learns to keep the vehicle within the limits of the road, correcting the angle of the vehicle steering wheel. Finally, we show an implementation of the presented systems running on a custom-built electric car. A series of experiments were carried out on a real-life traffic environment for evaluating the steering angle prediction system and the urban object detector. A mechanical system was implemented on the car to enable automatic steering wheel control.

The tremendous technological advance experienced in recent years has allowed the development and implementation of algorithms capable of performing different tasks that help humans in their daily lives. Scene recognition is one of the fields most benefited by these advances. Scene recognition gives different systems the ability to define a context for the identification or recognition of objects or places. In this same line of research, semantic localization allows a robot to identify a place semantically. Semantic classification is currently an exciting topic and it is the main goal of a large number of works. Within this context, it is a challenge for a system or for a mobile robot to identify semantically an environment either because the environment is visually different or has been gradually modified. Changing environments are challenging scenarios because, in real-world applications, the system must be able to adapt to these environments. This research focuses on recent techniques for categorizing places that take advantage of DL to produce a semantic definition for a zone. As a contribution to the solution of this problem, in this work, a method capable of updating a previously trained model is designed. This method was used as a module of an agenda system to help people with cognitive problems in their daily tasks. An augmented reality mobile phone application was designed which uses DL techniques to locate a customer’s location and provide useful information, thus improving their shopping experience. These solutions will be described and explained in detail throughout the following document.

Estimating an object’s position can be of great interest in several applications,and there exists many different methods to do so. One approach is with Directionof Arrival (DOA) measurements from receivers to use the triangulation techniqueto estimate one or more transmitter’s position. One algorithm which can find theDOA measurements from several transmitters is the MUltiple SIgnal Classification(MUSIC) algorithm. However, this still leaves a ambiguity problem which givesfalse solutions, so called ghost points, if the number of receivers is not sufficient.In this report solving this problem with the help of deep learning is studied. Thethesis’s main objective is to investigate and study whether it is possible to performpositioning with measurements from the MUSIC-algorithm using deep learningand image processing methods. A deep neural network is built in TensorFlow and trained and tested using datagenerated from MATLAB. This thesis’s setup consists of two receivers, which areused to locate two transmitters. The network uses two MUSIC spectra from thetwo receivers, and returns a probability distribution of where the transmittersare located. The results are compared with a traditional method and are analysed.The results presented in this thesis show that it is possible to perform positioningusing deep learning methods. However, there is a lot of room for improvementwith accuracy, which can be an important future research direction to explore.

500 of 690 training images used in optimized training runs.

Intelligent transportation systems (ITS) have become a popular method to enhance the safety and efficiency of transportation. Pedestrians, as an essential participant of ITS, are very vulnerable in a traffic collision, compared with the passengers inside the vehicle. In order to protect the safety of all traffic participants and enhance transportation efficiency, the novel autonomous vehicles are required to detect pedestrians accurately and timely. In the area of pedestrian detection, deep learning-based pedestrian detection methods have gained significant development since the appearance of powerful GPUs. A large number of researchers are paying efforts to improve the accuracy of pedestrian detection by utilizing the Convolutional Neural Network (CNN)-based detectors. In this thesis, we propose a one-stage anchor-free pedestrian detector named Bi-Center Network (BCNet), which is aided by the semantic features of pedestrians' visible parts. The framework of our BCNet has two main modules: the feature extraction module produces the concatenated feature maps that extracted from different layers of ResNet, and the four parallel branches in the detection module produce the full body center keypoint heatmap, visible part center keypoint heatmap, heights, and offsets, respectively. The final bounding boxes are converted from the high response points on the fused center keypoint heatmap and corresponding predicted heights and offsets. The fused center keypoint heatmap contains the semantic feature fusion of the full body and the visible part of each pedestrian. Thus, we conduct ablation studies and discover the efficiency of feature fusion and how visibility features benefit the detector's performance by proposing two types of approaches: introducing two weighting hyper-parameters and applying three different attention mechanisms. Our BCNet gains 9.82% MR-2 (the lower the better) on the Reasonable setup of the CityPersons dataset, compared to baseline model which gains 12.14% MR-2 . The experimental results indicate that the performance of pedestrian detection could be significantly improved because the visibility semantic could prompt stronger responses on the heatmap. We compare our BCNet with state-of-the-art models on the CityPersons dataset and ETH dataset, which shows that our detector is effective and achieves a promising performance.

Connected vehicle networks and future autonomous driving systems call for characterization of risky driving events to improve safety applications and autonomous driving features. Precision of driving event characterization (\gls{dec}) systems in connected vehicles has become increasingly important with the responsive connectivity that is available to the modern vehicles. While risky behavior patterns entail potential safety issues on road networks, the advent of vehicular sensing and vehicular networks cannot guarantee accurate characterization of driving/movement behavior of vehicles and the precision of such systems still remains an open issue. Additionally, artificial intelligence-backed solutions are vital components towards highly accurate characterization systems in the modern transportation. However, such solutions require significant volume of driving event data for an acceptable level of performance. With this in mind, the proposal of this thesis is three-fold: 1) a reliable methodology to generate representative data under the scarcity of diverse anomalous sensory data, 2) classification of mobility/driving events of vehicles with attention-based deep learning methods, and 3) a modular prior-knowledge input method to the characterization methodologies in order to further improve the trustworthiness of the systems. Implementing the proposed steps, we are able to not only increase the consistency in the training process but also reduce the false positive detection instances compared to the previous models. One of the roadblocks against robust event characterization systems in connected vehicles that is tackled in this thesis is the scarcity of anomalous driving data to make the training of event classification models robust. To mitigate this issue an optimized deep recurrent neural network-based encoding model is introduced to extract the precise feature representation of the anomalous data. The utilization of the encoded input to the previous network allowed for a 12\% accuracy improvement. Furthermore, we introduced a framework for precise risky driving behavior detection that takes advantage of an attention-based neural networks model. Ultimately, the combination of prior knowledge modelling with our network and some optimizations to the network structure, the model outperforms the state-of-the-art solutions by reaching an average accuracy of 0.96 and F1-score of 0.92.