This work addresses approximate nearest neighbor search applied in the domain of large-scale image retrieval. Within the group testing framework we propose an efficient off-line construction of the search structures. ...
详细信息
ISBN:
(纸本)9783030012168;9783030012151
This work addresses approximate nearest neighbor search applied in the domain of large-scale image retrieval. Within the group testing framework we propose an efficient off-line construction of the search structures. The linear-time complexity orthogonal grouping increases the probability that at most one element from each group is matching to a given query. Non-maxima suppression with each group efficiently reduces the number of false positive results at no extra cost. Unlike in other well-performing approaches, all processing is local, fast, and suitable to process data in batches and in parallel. We experimentally show that the proposed method achieves search accuracy of the exhaustive search with significant reduction in the search complexity. The method can be naturally combined with existing embedding methods.
Nonnegative matrix factorization (NMF) has been successfully applied in different fields, such as text mining, imageprocessing, and video analysis. NMF is the problem of determining two non-negative low rank matrices...
详细信息
ISBN:
(纸本)9781450355810
Nonnegative matrix factorization (NMF) has been successfully applied in different fields, such as text mining, imageprocessing, and video analysis. NMF is the problem of determining two non-negative low rank matrices U and V, for a given input matrix M, such that M approximate to UVT. There is an increasing interest in parallel and distributed NMF algorithms, due to the high cost of centralized NMF on large matrices. In this paper, we propose a distributed sketched alternating nonnegative least squares (DSANLS) framework for NMF, which utilizes a matrix sketching technique to reduce the size of nonnegative least squares subproblems in each iteration for U and V. We design and analyze two different random matrix generation techniques and two subproblem solvers. Our theoretical analysis shows that DSANLS converges to the stationary point of the original NMF problem and it greatly reduces the computational cost in each subproblem as well as the communication cost within the cluster. DSANLS is implemented using MPI for communication, and tested on both dense and sparse real datasets. The results demonstrate the efficiency and scalability of our framework, compared to the state-of-art distributed NMF MPI implementation.
Recently, based on novel convolutional neural net-work architectures proposed, tremendous advances have been achieved in image denoising task. An effective and efficient multi-level network architecture for image deno...
详细信息
ISBN:
(纸本)9781728152103
Recently, based on novel convolutional neural net-work architectures proposed, tremendous advances have been achieved in image denoising task. An effective and efficient multi-level network architecture for image denoising refers to restore the latent clean image from a coarser scale to finer scales and pass features through multiple levels of the model. Unfortunately, the bottleneck of applying multi-level network architecture lies in the multi-scale information from input images is not effectively captured and the fine-to-coarse feature fusion strategy to be ignored in image denoising task. To solve these problems, we propose a multi-scale & multi-level shuffle-CNN Via multi-level attention (DnM3Net), which plugs the multi-scale feature extraction, fine-to-coarse feature fusion strategy and multi-level attention module into the new network architecture in image denoising task. The advantage of this approach are two-fold: (1) It solve the multi-scale information extraction issue of multi-level network architecture, making it more effective and efficient for the image denoising task. (2) It is impressive performance because the better trade-off between denoising and detail preservation. The proposed novel network architecture is validated by applying on synthetic gaussian noise gray and RGB images. Experimental results show that the DnM3Net effectively improve the quantitative metrics and visual quality compared to the state-of-the-art denoising methods.
Visual tracking is a fundamental problem in computer vision. Lots efforts have been made in the past decades and researchers have obtained some achievements, but there are still problems exist. To get a more powerful ...
详细信息
ISBN:
(纸本)9781728152103
Visual tracking is a fundamental problem in computer vision. Lots efforts have been made in the past decades and researchers have obtained some achievements, but there are still problems exist. To get a more powerful feature representation, deep learning based methods appear layer upon layer, and most of them achieve good performance at the price of losing speed. In this paper, we propose a SiamFL network for visual tracking, which introduce the focal loss with the Siamese network. With the focal loss, the network is capable of filtering out the easy examples, leaving the hard samples to train a discriminative network. Moreover, the SiamFL network also can alleviate the class imbalance problem by properly sampling the positive and negative samples at a certain ratio. Experimental results demonstrate the SimaFL has quick convergence and also have a better performance over the baseline tracker.
Aiming at the problem that the existing license plate location methods can't work well in locating the license plate in certain natural scenes such as low luminance, low resolution and inclination scene of vehicle...
详细信息
ISBN:
(纸本)9781728152103
Aiming at the problem that the existing license plate location methods can't work well in locating the license plate in certain natural scenes such as low luminance, low resolution and inclination scene of vehicle, a license plate location method based on cascade classifier and convolution neural network is proposed. Firstly, the cascade classifier is used for license plate coarse location, and the convolution neural network is used for license plate precise location. Experiments show that, compared with cascade classifier, the method in this paper has higher accuracy and recall rate for target license plate location, and the method in this paper has strong adaptability to different environments.
Face recognition have been widely used in different industries due to the advancement of deep convolutional neural networks. Although deep learning has greatly promoted the development of face recognition technology, ...
详细信息
ISBN:
(数字)9781728143286
ISBN:
(纸本)9781728143293
Face recognition have been widely used in different industries due to the advancement of deep convolutional neural networks. Although deep learning has greatly promoted the development of face recognition technology, its computing-intensive and memory-intensive features make it difficult to deploy the model on some embedded devices or mobile computing platforms. Many solutions which include Knowledge Distillation have been proposed to increase the calculation speed of model and reduce the storage space required for calculations, in this paper, we propose a novel Two Stage Knowledge Distillation which enhances the performance of knowledge distillation in face recognition and low resolution face recognition. After experimenting on the several major face datasets, our method turns out to have better results compared to the traditional optimization methods.
Innovative methods to interact with electronic devices are attracting researchers attentions, and many applications of Human Computer Interaction have been developed in recent years. Hand detection under different ill...
详细信息
ISBN:
(数字)9781728143286
ISBN:
(纸本)9781728143293
Innovative methods to interact with electronic devices are attracting researchers attentions, and many applications of Human Computer Interaction have been developed in recent years. Hand detection under different illuminations and backgrounds is crucial for tasks such as, actions recognition, gestures recognition and hand-based interaction with wearable devices. Many existing methods are able to detect the human hand precisely, but the task still remains challenging due to environment complexity, different invariants, processing time, and accuracy. It is required for an application to recognize the human hand precisely for proper interaction with the wearable devices. This paper proposed a robust and a compact method for hand detection by employing you only look once (YOLO) and convolutional pos machines (CPM). This paper consists of two parts, network training and testing. In the first part, Oxford hand dataset is used to train a neural network. This dataset has various hand images under different invariants, and backgrounds. Afterwards, images are captured using a digital camera in real-time and then human hands are detected using a trained neural network. In addition, detected hands, positions and orientations are calculated accordance with defined active zone parameters and a depth sensor. Our trained neural network is dubbed as HHDNet, an abbreviation for human hand detection network. The HHDNet is tested under different environment conditions, and experimental results showed that the accuracy and the frame rate of the proposed model are superior than others state of the art methods.
Digital data of patients can aid a pathologist along the diagnostic process In Medical devices generate data sets that are processed by specialized computing applications, which often run on a single computer. The res...
详细信息
ISBN:
(纸本)9781538650356
Digital data of patients can aid a pathologist along the diagnostic process In Medical devices generate data sets that are processed by specialized computing applications, which often run on a single computer. The resolution power of the devices is increasing steadily and, consequently, the volumes of the data sets are also growing and can no longer be analyzed in a reasonable amount of time. Big Data tools like Apache Spark [2] provide methods for analyzing data, however, they are not directly applicable and need considerable implementation efforts, in general. Usually, well established analysis tools for medical data are designed to run on single workstations. These tools are not designed to meet current and future challenges. Migrating processing tools from single nodes to distributed environments is nontrivial. Moreover, partitioning data sets for a parallelprocessing is a further challenge [3]. In this work, we continue our efforts for improving the speedup of a bio-medical big data application further by partitioning the images of a Whole Slide image (WSI) [4] into sub tiles and by analyzing these sub tiles on a cluster of computer nodes. The idea is to benefit from the divide and conquer strategy. However, it is shown that the score parameter is determined incorrectly, when the software package is applied to each sub tile and the score parameters of all sub tiles are combined in an apparently natural manner. The cause of this anomaly is determined and a solution suggested. The original software is based on implicit assumptions. For example, the size of the tiles is assumed to be 1024 x 1024 px(2). The anomaly shows up when this constraint is reduced.
In this paper, we propose an identification framework to determine copyright infringement in the form of illegally distributed print-scan books in a large database. The framework contains following main stages: image ...
详细信息
In this paper, we propose an identification framework to determine copyright infringement in the form of illegally distributed print-scan books in a large database. The framework contains following main stages: image pre-processing, feature vector extraction, clustering, and indexing, and hierarchical search. The image pre-processing stage provides methods for alleviating the distortions induced by a scanner or digital camera. From the preprocessed image, we propose to generate feature vectors that are robust against distortion. To enhance the clustering performance in a large database, we use a clustering method based on the parallel-distributed computing of Hadobp MapReduce. In addition, to store the clustered feature vectors efficiently and minimize the searching time, we investigate an inverted index for fedture vectors. Finally, we implement a two-step hierarchical search to achieve fast and accurate on-line identification. In a simulation, the proposed identification framework shows accurate and robust in the presence of print-scan distortions. The processing time analysis in a parallel computing environment gives extensibility of the proposed framework to massive data. In the matching performance analysis, we empirically and theoretically find that in terms of query time, the optimal number of clusters scales with O(root N) for N print-scan books. (C) 2017 Elsevier Inc. All rights reserved.
Deep clustering attempts to capture the feature representation that benefits the clustering issue for inputs. Although the existing deep clustering methods have achieved encouraging performance in many research fields...
详细信息
ISBN:
(数字)9781728143286
ISBN:
(纸本)9781728143293
Deep clustering attempts to capture the feature representation that benefits the clustering issue for inputs. Although the existing deep clustering methods have achieved encouraging performance in many research fields, there still presents some shortcomings, such as the lack of consideration of local structure preservation and sparse characteristics of input data. To this end, we propose the deep-stacked sparse embedded clustering method in this paper, which considers both the preservation of local structure and sparse property of inputs. The proposed network is trained to capture the feature representation for input data by the guidance of clustering loss and reconstruction loss, where the reconstruction loss prevents the corruption of feature space and guarantees the local structure preservation. Besides, some sparse parameters are added to the encoder to avoid learning of meaningless features. Through simultaneously minimizing the reconstruction loss and cluster loss, the proposed network can jointly learn the clustering oriented feature and optimize the assignment of cluster labels. Then we conduct amounts of comparative experiments, which consist of six clustering methods and four publicly available data sets. Eventually, the clustering accuracy, adjusted rand index and normalized mutual information are utilized as three evaluation metrics to provide a comparison. Comprehensive experiments validate the effectiveness of introducing sparse property and preserving local structure in our method.
暂无评论