Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2022-05-18 Cooperative journals: 《计算机应用研究》
Abstract： Video super-resolution (VSR) , which aims to exploit information from multiple adjacent frames to generate a high-resolution version of a reference frame. Many existing VSR works focus on how to effectively align adjacent frames to better fuse adjacent frame information, and little research has been done on the important step of adjacent frame information fusion. To solve this problem, This paper propose a video super-resolution model based on group feedback fusion mechanism (GFFMVSR) . Specifically, after adjacent frames are aligned, the aligned video sequences are fed into the first temporal attention module. Then, the sequence is divided into several groups, and each group achieves preliminary fusion through the intra-group fusion module in turn. Next, the fusion results of different groups go through a second temporal attention module. Then, each group inputs the feedback fusion module group by group, and uses the feedback mechanism to feedback and fuse the information of different groups. Finally, the fusion result output is reconstructed. It has been verified that the model has strong information fusion ability, and is superior to the existing models in both objective evaluation indicators and subjective visual effects.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2022-04-07 Cooperative journals: 《计算机应用研究》
Abstract： Recommendation systems are widely used everywhere and have a great influence on daily life. Aiming to train an ideal recommendation system, a massive of ‘use-item’ interactive pairs should be provided. However, the dataset obtained is usually sparse, which might result in an over-fitting model and be hard to obtain the satisfying performance. In order to address this problem, the cross-domain recommendation is raised. Most of the existing methods on cross-domain recommendation systems borrow the ideas of the conventional unsupervised domain adaptation, which employ the feature alignment or adversarial training methods. Hence they can transfer the domain-invariant interests of users from the source to the target domains, e. g. , from the Douban Movies to the Douban Books. However, since the network structures vary with different recommendation platforms, the existing methods on cross-domain recommendation systems straightforwardly extract the domain-invariant representation may entangle the structure information, which may result in the false alignment phenomenon. Furthermore, the previous efforts ignore the noise information behind the graph data, which further degenerate the experimental performance. In order to address the aforementioned problems, this paper brings the causal generation process of graph data into the cross-domain recommendation systems, this paper use the semantic latent variables of each node to calculate the relationships between users and items via disentangling the semantic latent variables, domain latent variables and noise latent variables. Experiments studies testify that the proposed method yields state-of-the-art performance on several cross-domain recommendation system benchmark datasets.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2019-01-28 Cooperative journals: 《计算机应用研究》
Abstract： The existing mainstream methods use the pre-trained convolutional neural networks to extract image features and usually have the following limitations: a）Only using a single layer of pre-trained features to represent image; b）Inconsistency between the pre-trained task and the actual research task. These limitations result in that the existing methods of image-text matching cannot make full use of image features and is easily influenced by the noises. To solve the above limitations, this paper used multi-layer features from a pre-trained network and proposed a fusion algorithm of multi-level image features accordingly. Under the guidance of the image-text matching objective function, the proposed algorithm fused the multi-level pre-trained image features and reduced their dimensionality using a multi-layer perceptron to generate fusion features. It is able to make full use of pre-trained features and successfully reduce the influences of noises. The experimental results show that the proposed fusion algorithm makes better use of pre-trained image features and outperforms the methods using single-level features in the image-text matching task.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2018-11-29 Cooperative journals: 《计算机应用研究》
Abstract： The method based on recurrent neural network combined with syntactic structure is widely used in relation classification, and the neural network is used to automatically acquire features and realize relation classification. However, the existing methods are mainly based on a single specific syntactic structure model, and the model of a specific syntactic structure cannot be transferred to other types of syntactic structures. Aiming at this problem, a hierarchical recurrent neural network model with multi-syntactic structure is proposed. The hierarchical recurrent neural network is divided into two layers for network construction. Firstly, entity pre-training is performed in the sequence layer. The Bi-LSTM-CRF fusion Attention mechanism is used to improve the model's attention to the entity information on the text sequence, thereby obtaining more accurate. The more accurate entity feature information promotes better classification in the relation layer stage. Secondly, in the relation layer, the Bi-Tree-LSTM is nested above the sequence layer, and the hidden state and entity feature information of the sequence layer is passed into the relation layer, then three different syntax structures are weighted learned using the shared parameters and classify the semantic relation finally. The experimental results show that the model has a marco-F1 value of 85.9% on the SemEval-2010 Task8 corpus, and further improves the robustness of the model.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2018-11-29 Cooperative journals: 《计算机应用研究》
Abstract： Convolutional neural network's high computational complexity often obstructs its widespread adhibition in real-time and low-power applications. The existing software implementation solution cannot meet the demands of the convolutional neural network for computing performance and power consumption. The traditional FPGA-oriented convolutional neural network construction method has problems such as complicated process, long cycle and small optimization space. For these problems, according to the characteristics of convolutional neural network calculation pattern, this paper proposed a design and scheduling mechanism of convolutional neural network accelerator for cloud FPGAs. By using for reference the design which based HLS technology, imported the cyclic cutting parameters and rearranged the convolution layer circularly. Then constructed the network in a modular way, and extended parameters to further optimize the accelerator processing process. Summarized the scheduling scheme by analyzing the characteristics of system tasks and resources, and optimized its design from two aspects of control and data flow. In comparison with other existing work, the proposed design provided a solution with flexibility, low energy consumption, high energy efficiency and performance. The design also discussed the efficient universal scheduling scheme of the accelerator. Experimental results show that compared with the CPU implementation, this design achieves 8.84 times speedup of AlexNet, while the power consumption of Cifar implementation is only 24.96% of it. Compared with the CPU+GPU to achieve 6.90 times speedup of Cifar, although the performance of large-scale network is inferior to the GPU, but the minimum power consumption is only 14.98%. This design achieves the maximum acceleration of 6.29 times in comparison with the existing research results. Compared to the accelerators generated for large platforms, even if it only has comparable performance but with a lower clock frequency.
Hits
Hits
Downloads
Comment