Search

• We proposed a novel Robust Matrix Completion method based on Rank-One matrix approximation (RMCRO), which divides the noisy and incomplete data matrix into two parts: a low-rank matrix and a sparse matrix, and it is robust to sparse noise. • We use a rank-one matrix pursuit method to approximate the low-rank part of the incomplete data matrix. Therefore, the proposed algorithm can recover the missing entries of the incomplete data matrix and estimate the rank simultaneously. • Extensive experimental results on both synthetic datasets and real world datasets demonstrate the effectiveness of the proposed method for matrix completion and rank estimation. Matrix completion aims at estimating the missing entries of a low-rank and incomplete data matrix. It frequently arises in many applications such as computer vision, pattern recognition, recommendation system, and data mining. Most of the existing methods face two problems. Firstly, the data matrix in real world is often disturbed by noise. Noise may change the date structure of the incomplete matrix, thereby degrade the performance of matrix completion algorithms. Secondly, some existing methods need to preset a reasonable rank as input, and the value of rank will affect the performance of the algorithms. Therefore, we proposed a robust rank-one matrix completion method with rank estimation in this paper. To mitigate the influence of noise, we divide the incomplete and noisy data matrix into two parts iteratively: low-rank and sparse parts. Besides, we use a weighted rank-one matrix pursuit algorithm to approximate the low-rank part of the data matrix, and the rank of the matrix can be estimated with the adaptive weight vector. The performance of the proposed method is demonstrated by experiments on both synthetic datasets and image datasets. The experimental results demonstrate the performance of the proposed method with incompleted matrices distrubed by sparse noise. [ABSTRACT FROM AUTHOR]

Realistic action recognition has been one of the most challenging research topics in computer vision. The existing methods are commonly based on non-probabilistic classification, predicting category labels but not providing an estimation of uncertainty. In this paper, we propose a probabilistic framework using Gaussian processes (GPs), which can tackle regression problems with explicit uncertain models, for action recognition. A major challenge for GPs when applied to large-scale realistic data is that a large covariance matrix needs to be inverted during inference. Additionally, from the manifold perspective, the intrinsic structure of the data space is only constrained by a local neighborhood and data relationships with far-distance usually can be ignored. Thus, we design our GPs covariance matrix via the proposed ℓ 1 construction and a local approximation (LA) covariance weight updating method, which are demonstrated to be robust to data noise, automatically sparse and adaptive to the neighborhood. Extensive experiments on four realistic datasets, i.e., UCF YouTube, UCF Sports, Hollywood2 and HMDB51, show the competitive results of ℓ 1 -GPs compared with state-of-the-art methods on action recognition tasks. [ABSTRACT FROM AUTHOR]

Abstract: The technique of visual saliency detection supports video surveillance systems by reducing redundant information and highlighting the critical, visually important regions. It follows that information about the image might be of great importance in depicting the visual saliency. However, the majority of existing methods extract contrast-like features without considering the contribution of information content. Based on the hypothesis that information divergence leads to visual saliency, a two-stage framework for saliency detection, namely information divergence model (IDM), is introduced in this paper. The term “information divergence” is used to express the non-uniform distribution of the visual information in an image. The first stage is constructed to extract sparse features by employing independent component analysis (ICA) and difference of Gaussians (DoG) filter. The second stage improves the Bayesian surprise model to compute information divergence across an image. A visual saliency map is finally obtained from the information divergence. Experiments are conducted on nature image databases, psychological patterns and video surveillance sequences. The results show the effectiveness of the proposed method by comparing it with 13 state-of-the-art visual saliency detection methods. [Copyright &y& Elsevier]

• A GA training pipeline to select key input channels of DenseNet automatically. • A 'Weight inheritance' strategy to improve efficiency whilst maintaining accuracy. • To reduce up to 30% parameters yet achieving a similar accuracy as the DenseNet. • Validated robustness/scalability on several image recognition/classification datasets. Convolutional neural networks (CNNs) have been successfully applied in many computer vision applications [1] , especially in image classification tasks, where most of the structures have been designed manually. With the aid of skip connection and dense connection, the depths of the models are becoming "deeper" and the filters of layers are getting "wider" in order to tackle the challenge of large-scale datasets. However, large-scale models in convolutional layers become inefficient due to the redundant channels from input feature maps. In this paper, we aim to automatically optimize the topology of the DenseNet, in which unnecessary convolutional kernels are reduced. To achieve this, we present a training pipeline that generates the network structure using a genetic algorithm. We first propose two encoding methods that can represent the structure of the model using a fixed-length binary string. A three-step based evolutionary process consisting of selection, crossover, and mutation is proposed to optimize the structure. We also present a pretrained weight inheritance method which can largely reduce the total time consumption of the genetic process. Experimental results have demonstrated that our proposed model can achieve comparable accuracy to the state-of-the-art models, across a wide range of image recognition and classification datasets, whilst significantly reducing the number of parameters. [ABSTRACT FROM AUTHOR]

• This paper proposes a novel stacked Squeeze-and-Excitation Recurrent Residual Network (SER2-Net) for visual-semantic matching. • This paper develops an effective and efficient cross-modal representation learning module, which is capable of generating semantically complementary multi-level features for both modalities. • This paper presents a novel objective function for aligning cross-modal data, which is able to capture the interdependency among multiple semantic levels to alleviate the distribution inconsistency between visual and textual modality. • Extensive experiments on two benchmark datasets demonstrate the superiority of our proposed model compared to the state-of-the-art approaches. In recent years, visual-textual matching has been widely studied in the intersection of computer vision and natural language processing communities. A feasible scheme for learning discriminative representations is leveraging hierarchical features to align both modalities at multiple semantic levels. However, most existing approaches rely on pre-trained object detectors or semantic parsers to generate multi-level representations, whose performance is overly dependent on the extra supervision and thereby leads to its vulnerability. In this paper, we introduce a Stacked Squeeze-and-Excitation Recurrent Residual Network (SER2-Net) for visual-textual matching. Firstly, an efficient multi-level representation module is presented to produce a series of semantically discriminative features without the aid of extra supervision, which is built by stacking the squeeze-and-excitation recurrent residual (SER2) learning components. Specifically, SER2 incorporates the residual learning and inverse recurrent connection into the squeeze-and-excitation learning block, which allows for utilizing complementary current information and residual information to improve the modality-specific representation ability. Besides, to capture the implicit correlations contained among multi-level features, we propose a novel objective namely Cross-modal Semantic Discrepancy (CMSD) loss, which is characterized by exploiting the interdependency among different semantic levels to narrow the cross-modal distribution discrepancy. Extensive experiments on two benchmark datasets validate the superiority of our model, which compares favorably with the state-of-the-art approaches. [ABSTRACT FROM AUTHOR]