Image super-resolution reconstruction based on deep learning improves the image reconstruction performance by deepening the network. However, its application on resource-limited devices is limited due to the sharp increase in the number of parameters caused by complex networks. To solve this problem, an image super-resolution reconstruction method based on feature aggregation and propagation network is proposed, enriching internal information of images by extracting and fusing features step by step. Firstly, a contextual interaction attention block is proposed to enable the network to learn the rich contextual information of feature maps as well as improve the utilization of features. Then, a multi-dimensional attention enhancement block is designed to improve the network's ability to discriminate the key features and extract high-frequency information in channel dimension and spatial dimension, respectively. Finally, a feature aggregation and propagation block is proposed to effectively aggregate deep detail information, remove redundant information and promote the propagation of effective information in the network. Experimental results on Set5,Set14,BSD100 and Urban100 datasets demonstrate the superiority of the proposed method with clearer details of reconstructed images.

High dynamic range(HDR) imaging based on multi-exposure fusion aims to generate high-quality HDR images by integrating the information from multiple low dynamic range(LDR) images. However, HDR imaging is faced with two major challenges, ghosting artifact suppression in motion regions and lost information restoration in over-saturated areas. To comprehensively address the challenges of restoring missing content from reference images and suppressing ghosting artifacts in motion regions, a missing content restoration and ghosting suppression network for high dynamic range imaging is proposed in this paper. In terms of content restoration, a predictive filtering-based content restoration block is introduced. The filtering kernel predicted by the content restoration block is employed to filter reference image features, integrating key information from both reference images and non-reference images to provide richer information for effective reconstruction of missing content. To suppress ghosting artifacts in motion regions and fully exploit complementary information from non-reference images, deformable convolutions are introduced to align features from non-reference images with those from the reference image. Additionally, to enhance the HDR image reconstruction capability of the network, a three-branch image reconstruction module is constructed, including a main branch and two auxiliary branches. The auxiliary branches assist the main branch with better preserved details during the generation of HDR results. Experimental results demonstrate superior performance of the proposed network.

Methodssuch as stimulative training and group knowledge based training are employed to collect group knowledge from shallow subnets in residual networks for self-distillation, thereby enhancing network performance. However, the group knowledge acquired by these methods suffers from issues such as slow updating and difficulties in combining with DataMix techniques. To address these issues, enhanced residual networks via mixed knowledge fraction(MKF) are proposed. The mixed knowledge is decomposed and modeled as quadratic programming by minimizing the fraction loss, and thus high-quality group knowledge is obtained from the mixed knowledge. To improve the robustness and diversity of the knowledge, a compound DataMix technique is proposed to construct a composite data augmentation method. Different from high-precision optimization algorithms with poor efficiency, a simple and efficient linear knowledge fraction technique is designed. The previous group knowledge is taken as knowledge bases, and the mixed knowledge is decomposed based on the knowledge bases. The enhanced group knowledge is then adopted to distill sampled subnetworks. Experiments on mainstream residual networks and classification datasets verify the effectiveness of MKF.

Tabular datasets with limited sample size lack invariance structure and enough samples, making traditional generative data augmentation methods difficult to obtain diverse data that conforms to the original data distribution. To address this issue, a vicinal distribution-based denoising diffusion probabilistic model(VD-DDPM) and its learning algorithm based on the characteristics of tabular data and the principle of vicinal risk minimization are proposed. Firstly, features of the tabular data with limited sample size are analyzed. Weakly correlated features are selected via priori knowledge, and the vicinal distribution of the training sample is constructed. Then, the VD-DDPM is built on the data sampled from vicinal distribution. A diverse dataset that conforms to the original data distribution is generated via VD-DDPM generation algorithm. Experiments on multiple datasets verify the effectiveness of the proposed algorithm in terms of the quality of the generated data and the performance of the downstream model.

Multi-scale decision system is one of hot issues in the field of data mining, and cost factors appear frequently in data mining. A method for simultaneous selection of attributes and scales can effectively solve the knowledge reduction problem of multi-scale decision systems involving cost factors. However, in the existing research, there are few studies on the simultaneous selection of attributes and scales based on costs, and most of the algorithms only focus on consistent or inconsistent multi-scale decision systems. To address this issue, a test cost sensitive method for simultaneously selecting attributes and scales is proposed with the goal of minimizing the total test cost of data processing. The method is applicable to both consistent and inconsistent multi-scale decision systems. Firstly, a theoretical model is constructed based on rough set. In the model, both the attribute factor and the scale factor are taken intoaccount by concepts and properties. Secondly, a heuristic algorithm is designed based on the theoretical model. By the proposed algorithm, attribute reduction and scale selection can be simultaneously performed in the multi-scale decision systems based on test costs, and different attributes can choose different scales. Finally, the experiments verify the effectiveness, practicality and superiority of the proposed algorithm.