Pas de résumé / This thesis is dedicated to the problem of training and integration strategies of several modalities (visual, textual), in order to perform an efficient Visual Concept Detection and Annotation (VCDA) task, which has become a very popular and important research topic in recent years because of its wide range of application such as image/video indexing and retrieval, security access control, video monitoring, etc. Despite a lot of efforts and progress that have been made during the past years, it remains an open problem and is still considered as one of the most challenging problems in computer vision community, mainly due to inter-class similarities and intra-class variations like occlusion, background clutter, changes in viewpoint, pose, scale and illumination. This means that the image content can hardly be described by low-level visual features. In order to address these problems, the text associated with images is used to capture valuable semantic meanings about image content. Moreover, In order to benefit from both visual models and textual models, we propose multimodal approach. As the typical visual models, designing good visual descriptors and modeling these descriptors play an important role. Meanwhile how to organize the text associated with images is also very important. In this context, the objective of this thesis is to propose some innovative contributions for the task of VCDA. For visual models, a novel visual features/descriptors was proposed, which effectively and efficiently represent the visual content of images/videos. In addition, a novel method for encoding local binary descriptors was present. For textual models, we proposed two kinds of novel textual descriptor. The first descriptor is semantic Bag-of-Words(sBoW) using a dictionary. The second descriptor is Image Distance Feature(IDF) based on tags associated with images. Finally, in order to benefit from both visual models and textual models, fusion is carried out by MKL efficiently embed. [...]

Detection and choice research have largely focused on the effects of relative reinforcer frequencies or magnitudes. The effects of punishment have received much less attention. This thesis investigated the effects of punishment on human signal-detection performance using a number of different procedures. These included punisher frequency and magnitude variations, different types of punishers (point loss & time-outs), variations in stimulus disparity, and different detection tasks (judgments of stimulus arrays containing either more blue or red objects, or judgments of statements that were either true or false). It examined whether punishers have similar, but opposite, effects to reinforcers on detection performance, and whether the effects of punishment were successfully captured by existing models of punishment and choice. Experiment 1 varied the relative frequency or magnitude of time-out punishers for errors using the blue/red task. Participants were systematically biased away from the response alternative associated with the higher rate or magnitude of time-out punishers in two of three procedures. Experiment 2 varied the relative frequency of point-loss punishers using the blue/red task and the true/false task. Participants were systematically biased away from the alternative associated with the higher rate of point-loss punishers for the true/false task. Experiment 3 examined the effects of punishment on response bias from a psychophysical perspective. Previous detection research which varied stimulus discriminability while holding reinforcers ratios constant and unequal (Johnstone & Alsop, 2000; McCarthy & Davison, 1984) found that a criterion location measure (e.g., c, Green & Swets, 1966) was a better descriptor of isobias functions compared to a likelihood ratio measure (e.g., log β[G], Green & Swets, 1966). Experiment 3 varied stimulus discriminability while holding punisher ratios constant and unequal. Like previous research, isobias functions were consistent with a criterion location measure. Experiments 4, 5, 6, and 7 examined contemporary models of choice and punishment. Experiments 4, 5, and 6 varied the relative reinforcer ratio in detection tasks, both with and without the inclusion of an equal rate of punishment. Experiment 7 held the reinforcer ratio constant and unequal, and varied the durations of time-out punishers. Increases in preference (for the richer alternative) from reinforcer-only conditions to reinforcer + punisher conditions would support a subtractive model of punishment, while decreases in preference would support an additive model of punishment. Experiment 4 was a between-groups study using time-out punishers. It supported the predictions of an additive model. Experiment 5 used three different procedures in a preliminary within-subjects design, evaluating which procedure was best suited for a larger within-subjects experiment (Experiment 6). In Experiment 6, participants sat four reinforcer-only and four reinforcer + punisher conditions where reinforcers were point-gains and punishers were point-losses. The results from Experiment 6 were mixed - some participants showed increased preference while others showed little change or a slight decrease. This appeared related to the order in which participants received the reinforcer-only and reinforcer + punisher conditions. Experiment 7 also found no consistent change in preference with increases in time-out durations. Instead, there was a slow increase in bias on the richer alternative across the eight sessions. Overall, punishers had similar, but opposite, effects to reinforcers in detection procedures (Experiments 1, 2, & 3). These effects were successfully captured by Davison and Tustin�s (1978) model of detection. The later experiments did not provide support for a subtractive model punishment model of choice, which had provided the best descriptor in corresponding concurrent-schedule research. Instead, Experiment 4 supported an additive model, and Experiments 5, 6, and 7 provided no evidence for either model - limitations and implications of these studies are discussed. However, the present thesis illustrates that the signal detection procedure is promising for studying the combined effects of reinforcement and punishment, and may offer a worthwhile complement to standard concurrent-schedule choice procedures.

This thesis presents a general, trainable system for object detection in static images and video sequences. The core system finds a certain class of objects in static images of completely unconstrained, cluttered scenes without using motion, tracking, or handcrafted models and without making any assumptions on the scene structure or the number of objects in the scene. The system uses a set of training data of positive and negative example images as input, transforms the pixel images to a Haar wavelet representation, and uses a support vector machine classifier to learn the difference between in-class and out-of-class patterns. To detect objects in out-of-sample images, we do a brute force search over all the subwindows in the image. This system is applied to face, people, and car detection with excellent results. For our extensions to video sequences, we augment the core static detection system in several ways -- 1) extending the representation to five frames, 2) implementing an approximation to a Kalman filter, and 3) modeling detections in an image as a density and propagating this density through time according to measured features. In addition, we present a real-time version of the system that is currently running in a DaimlerChrysler experimental vehicle. As part of this thesis, we also present a system that, instead of detecting full patterns, uses a component-based approach. We find it to be more robust to occlusions, rotations in depth, and severe lighting conditions for people detection than the full body version. We also experiment with various other representations including pixels and principal components and show results that quantify how the number of features, color, and gray-level affect performance.

With the development of mechanical fault diagnosis technology, complex mechanical systems do not need to be shut down periodically for the maintenance. The working condition of the mechanical systems can be monitored by analyzing the wear metal particles in the systems' lubricating oil. However, the output signals of the monitoring sensor are non-stationary. In some case the particle signals are overlapped with each other. The goal of this thesis is to find a method to decompose those overlapped particle signals, and then count the particle number in the lubricating oil. At the beginning EMD method was introduced in the experiment because of the character of the sensor signals. In this project, because EMD method is sensitive to the noise in the original signals, an improved version of EMD, EEMD method was implemented. Finally, a post processing method was used to get a better result.

With the development of mechanical fault diagnosis technology, complex mechanical systems do not need to be shut down periodically for the maintenance. The working condition of the mechanical systems can be monitored by analyzing the wear metal particles in the systems' lubricating oil. However, the output signals of the monitoring sensor are non-stationary. In some case the particle signals are overlapped with each other. The goal of this thesis is to find a method to decompose those overlapped particle signals, and then count the particle number in the lubricating oil. At the beginning EMD method was introduced in the experiment because of the character of the sensor signals. In this project, because EMD method is sensitive to the noise in the original signals, an improved version of EMD, EEMD method was implemented. Finally, a post processing method was used to get a better result.

Software maintenance accounts for the most part of the software development cost and efforts, with its major activities focused on the detection, location, analysis and removal of defects present in the software. Although software defects can be originated, and be present, at any phase of the software development life-cycle, implementation (i.e., source code) contains more than three-fourths of the total defects. Due to the diverse nature of the defects, their detection and analysis activities have to be carried out by equally diverse tools, often necessitating the application of multiple tools for reasonable defect coverage that directly increases maintenance overhead. Unified detection tools are known to combine different specialized techniques into a single and massive core, resulting in operational difficulty and maintenance cost increment. The objective of this research was to search for a technique that can detect dissimilar defects using a simplified model and a single methodology, both of which should contribute in creating an easy-to-acquire solution. Following this goal, a ‘Supervised Automation Framework’ named FlexTax was developed for semi-automatic defect mapping and taxonomy generation, which was then applied on a large-scale real-world defect dataset to generate a comprehensive Defect Taxonomy that was verified using machine learning classifiers and manual verification. This Taxonomy, along with an extensive literature survey, was used for comprehension of the properties of different classes of defects, and for developing Defect Similarity Metrics. The Taxonomy, and the Similarity Metrics were then used to develop a defect detection model and associated techniques, collectively named Symbolic Range Tuple Analysis, or SRTA. SRTA relies on Symbolic Analysis, Path Summarization and Range Propagation to detect dissimilar classes of defects using a simplified set of operations. To verify the effectiveness of the technique, SRTA was evaluated by processing multiple real-world open-source systems, by direct comparison with three state-of-the-art tools, by a controlled experiment, by using an established Benchmark, by comparison with other tools through secondary data, and by a large-scale fault-injection experiment conducted using a Mutation-Injection Framework, which relied on the taxonomy developed earlier for the definition of mutation rules. Experimental results confirmed SRTA’s practicality, generality, scalability and accuracy, and proved SRTA’s applicability as a new Defect Detection Technique.

Unfortunately, people die and get injured due to accidents in the traffic. Furthermore, statistics of road accidents is limited and mostly composed of serious accidents, making it difficult to draw conclusions about how to improve the safety in the traffic. Thus, there is an interest in obtaining information about critical events in the traffic, i.e. potential accident situations, since they occur much more frequently. One way of detecting critical events is to use sensors, such as accelerometers and gyroscopes. As the usage of cellphones with built-in sensors increases, it would be interesting to examine whether these sensors are good enough to detect critical events. This is where the focus of this thesis lies. An application that collects data from the cellphone’s built-in accelerometer, gyroscope and GPS was developed and tested. The data was then analysed and compared to data from accurate sensors, represented by a VBOX coupled to an IMU. The conclusions made in this thesis are that the sensors in the cellphone perform almost equivalent results compared to the VBOX. It is possible to use data from the sensors in orderto detect critical events.

Characterization of dispersion curves in plate-like structures is possible with guided Lamb waves. In this research, experimental development of dispersion curves relies on the spectrogram, which suffers from the Heisenberg Uncertainty Principle. Reassignment is capable of localizing ill--defined dispersion curves. Unfortunately, reassignment also introduces spurious components, which reduce reassignment performance. This research develops an algorithm that provides both localization of dispersion curves and elimination of spurious components. To achieve this, an alternative formulation of reassignment called differential reassignment is modified and superimposed with nonlinear anisotropic diffusion. This study first examines reassignment and diffusion components individually. Three different versions of differential reassignment are considered, two of which are modifications explicitly derived in this research. The combined algorithm is then applied to reassign experimentally measured spectrograms, leading to a significant increase in clarity and notch detection performance.

Electrocardiogram is a pictorial representation of the electrical activity of heart beats. Because of the direct relationship between the ECG waveform and interval of the heart beats, it is possible that doctor can diagnose cardiac disease and monitor patient conditions from the unusual ECG waveforms. Based on the wavelet transform, this work introduces an algorithm to detect QRS complex. In particular, the quadratic spline wavelet has been adopted. The thesis first reviews wavelet transform briefly, then develops a QRS detention algorithm, which is then tested by using the MIT-BIH arrhythmia database. It is hoped that the proposed QRS detection algorithm can be a useful tool for medical personnel who are interested in using QRS information to explore their research work.

In this study, an efficient new hybrid approach for multiple sensors data fusion and fault detection is presented, addressing the problem with possible multiple faults, which is based on conventional fuzzy soft clustering and artificial immune system (AIS). The proposed hybrid system approach consists of three main phases. In the first phase signal separation is performed using the Fuzzy C-Means (FCM) algorithm. Subsequently a single (fused) signal based on the information provided from the sensor signals is generated by the fusion engine. The information provided from the previous two phases is used for fault detection in the third phase based on the Artificial Immune System (AIS) negative selection mechanism. The simulations and experiments for multiple sensor systems have confirmed the strength of the new approach for online fusing and fault detection. The hybrid system gives a fault tolerance by handling different problems such as noisy sensor signals and multiple faulty sensors. This makes the new hybrid approach attractive for solving such fusion problems and fault detection during real time operations. This hybrid system is extended for early fault detection in complex mechanical systems based on a set of extracted features; these features characterize the collected sensors data. The hybrid system is able to detect the onset of fault conditions which can lead to critical damage or failure. This early detection of failure signs can provide more effective information for any maintenance actions or corrective procedure decisions.