No description available.

No description available.

Trimap-free natural image matting problem is an important computer vision task in which we extract foreground objects from given images without extra trimap input. Compared with trimap-based matting algorithms, trimap-free algorithms are easier to make false detection when the foreground object is not well defined. To solve the problem, we design a novel structure (SegMatting) to handle foreground segmentation and alpha matte prediction simultaneously, which is able to produce high-quality mattes based on RGB inputs alone. This entangled structure enables information exchange between the binary segmentation task and the alpha matte prediction task interactively, and we further design a hybrid loss to adaptively balance two tasks during the multitask learning process. Additionally, we adopt a salient object detection dataset to pretrain our network so that we could obtain a more accurate foreground segment before our training process. Experiments indicate that the proposed SegMatting qualitatively and quantitatively outperforms most previous trimap-free models with a significant margin, while remains competitive among trimap-based methods. / Thesis / Master of Science in Electrical and Computer Engineering (MSECE)

Anomaly detection is the process of finding unusual events in a given dataset. Anomaly detection is often performed on datasets with a fixed set of predefined features. As a result of this, if the normal features bear a close resemblance to the anomalous features, most anomaly detection algorithms exhibit poor performance. This work seeks to answer the question, can we deform these features so as to make the anomalies standout and hence improve the anomaly detection outcome? We employ a Deep Learning and an Active Learning framework to learn features for anomaly detection. In Active Learning, an Oracle (usually a domain expert) labels a small amount of data over a series of training rounds. The deep neural network is trained after each round to incorporate the feedback from the Oracle into the model. Results on the MNIST, CIFAR-10 and Galaxy Zoo datasets show that our algorithm, Ahunt, significantly outperforms other anomaly detection algorithms used on a fixed, static, set of features. Ahunt can therefore overcome a poor choice of features that happen to be suboptimal for detecting anomalies in the data, learning more appropriate features. We also explore the role of the loss function and Active Learning query strategy, showing these are important, especially when there is a significant variation in the anomalies.

Sound source localization is an important task for several applications and the use of deep learning for this task has recently become a popular research topic. While the majority of the previous work has focused on static sound sources, in this work we evaluate the performance of a deep learning classification system for localization of high-speed moving sound sources. In particular, we systematically evaluate the effect of a wide range of parameters at three levels including: data generation (e.g., acoustic conditions), feature extraction (e.g., STFT parameters), and model training (e.g., neural network architectures). We evaluate the performance of multiple metrics in terms of precision, recall, F-score and confusion matrix in a multi-class multi-label classification framework. We used four different deep learning models: feedforward neural networks, recurrent neural network, gated recurrent networks and temporal Convolutional neural network. We showed that (1) the presence of some reverberation in the training dataset can help in achieving better detection for the direction of arrival of acoustic sources, (2) window size does not affect the performance of static sources but highly affects the performance of moving sources, (3) sequence length has a significant effect on the performance of recurrent neural network architectures, (4) temporal convolutional neural networks can outperform both recurrent and feedforward networks for moving sound sources, (5) training and testing on white noise is easier for the network than training on speech data, and (6) increasing the number of elements in the microphone array improves the performance of the direction of arrival estimation.

No description available.

Fractures are ubiquitous in the subsurface, and they provide dominant pathways for fluid flow in low permeability formations. Therefore, fractures usually play an essential role in many engineering fields, such as hydrology, waste disposal, geothermal reservoir and petroleum reservoir exploitation. Since fractures are invisible and have variable sizes from micrometers to kilometers, there is limited knowledge of their structure. We aim to deepen the understanding of fracture networks in the subsurface from their topological structures, hydraulic connectivity and characteristics at different scales. We adopt the discrete fracture network method and develop an efficient C++ code, HatchFrac, to make in-depth investigations possible. We start from generating stochastic fracture networks by constraining fracture geometries with different stochastic distributions. We apply percolation theory to investigate the global connectivity of fracture networks. We find that commonly adopted percolation parameters are unsuitable for the characterization of the percolation state of complex fracture networks. We implement the concept of global efficiency to quantify the impact of fracture geometries on the connectivity of fracture networks. Furthermore, we constrain the fracture networks with geological data and geomechanics principles. We investigate the correlation of fracture intensities with different dimensionality and find that it is not feasible to obtain correct 3D intensity parameters from 1D or 2D samples. We utilize a deep-learning technique and propose a pixel-based detection algorithm to automatically interpret fractures from raw outcrop images. Interpreted fracture maps provide abundant resources to investigate fracture intensities, lengths,orientations, and generations. For large scale faults, we develop a method to generate fault segments from a rough fault trace on a seismic map. Accurate fault geometries have significant impacts on damage zones and fault-related flow problems. For small scale fractures, we consider the impact of fracture sealing on the percolation state of orthogonal fracture networks. We emphasize the importance of non-critically stressed and partially sealed fractures, which are usually neglected because usually they are nonconductive. However, with significant stress perturbations, those noncritically stressed and partially sealed fractures can also contribute to the production by enlarging the stimulated reservoir volume.

No description available.

Deep neural networks (DNNs), particularly convolutional neural networks (CNNs), have been widely applied in the many fields, such as computer vision, natural language processing, speech recognition and etc. Although DNNs achieve dramatic accuracy improvements in these real-world tasks, they require significant amounts of resources (e.g., memory, energy, storage, bandwidth and computation resources). This limits the application of these networks on resource-constrained systems, such as mobile and edge devices. A large body of literature has been proposed to addresses this problem from the perspective of compressing DNNs while preserving their performance. In this thesis, we focus on compressing deep CNNs based on vector quantization techniques. The first part of this thesis summarizes some basic concepts in machine learning and popular techniques on model compression, including pruning, quantization, low-rank factorization and knowledge distillation approaches. Our main interest is quantization techniques, which compress networks by reducing the precision of parameters. Full-precision weights, activations and even gradients in networks can be quantized to 16-bit floating point numbers, 8-bit integers, or even binary numbers. Despite a possible performance degradation, quantization can greatly reduce the model size while maintaining model accuracy. In the second part of this thesis, we propose a novel vector quantization approach, which we refer to as Vector Quantization with Learned Codebook, or VQLC, for CNNs. Rather than performing scalar quantization, we choose vector quantization that can simultaneously quantize multiple weights at once. Instead of taking a pretraining/clustering approach as in most works, in VQLC, the codebook for quantization are learned together with neural network training from scratch. For the forward pass, the traditional convolutional filters are replaced by the convex combinations of a set of learnable codewords. During inference, the compressed model will be represented by a small-sized codebook and a set of indices, resulting in a significant reduction of model size while preserving the network's performance. Lastly, we validate our approach by quantizing multiple modern CNNs on several popular image classification benchmarks and compare with state-of-the-art quantization techniques. Our experimental results show that VQLC demonstrates at least comparable and often superior performance to the existing schemes. In particular, VQLC demonstrates significant advantages over the existing approaches on wide networks at the high rate of compression.

This paper analyzes three techniques attempting to detect strawberries at various stages in its growth cycle. Histogram of Oriented Gradients (HOG), Local Binary Patterns (LBP) and Convolutional Neural Networks (CNN) were implemented on a limited custom-built dataset. The methodologies were compared in terms of accuracy and computational efficiency. Computational efficiency is defined in terms of image resolution as testing on a smaller dimensional image is much quicker than larger dimensions. The CNN based implementation obtained the best results with an 88% accuracy at the highest level of efficiency as well (600x800). LBP generated moderate results with a 74% detection accuracy at an inefficient rate (5000x4000). Finally, HOG’s results were inconclusive as it performed poorly early on, generating too many misclassifications.