As an important structure in the convolutional neural network, the pooling layer can effectively reduce the generalization of the improved features after convolution. However, the current pooling method is easy to lose the key low-frequency small sample data for the unbalanced hyperspectral image data, which leads to a decrease in the classification accuracy for such data. In this paper, we propose a pooling method, GL-pooling (global-local pooling), for the target category of hyperspectral samples. Calculating the joint probability of samples in the global and local pooling and maximizing the retention of small-sample detail information in the pooling process to improve the detection efficiency of the network for small targets. Concomitantly, the pooling method can be widely combined in many deep learning networks to solve the problem of difficult classification of small sample targets without increasing the network depth and the number of parameters. We achieve the best performance in two hyperspectral data sets.

While deep neural network-based video denoising methods have achieved promising results, it is still hard to deploy them on mobile devices due to their high computational cost and memory demands. This paper aims to develop a lightweight deep video denoising method that is friendly to resource-constrained mobile devices. Inspired by the facts that 1) consecutive video frames usually contain redundant temporal coherency, and 2) neural networks are usually over-parameterized, we propose a multi-input multi-output (MIMO) paradigm to process consecutive video frames within one-forward-pass. The basic idea is concretized to a novel architecture termed Recurrent Multi-output Network (ReMoNet), which consists of recurrent temporal fusion and temporal aggregation blocks and is further reinforced by similarity-based mutual distillation. We conduct extensive experiments on NVIDIA GPU and Qualcomm Snapdragon 888 mobile platform with Gaussian noise and simulated Image-Signal-Processor (ISP) noise. The experimental results show that ReMoNet is both effective and efficient on video denoising. Moreover, we show that ReMoNet is more robust under higher noise level scenarios.

News recommendation aims to alleviate the big explosion of news information and helps users find their interesting news. Existing news recommendation models model users' historical click news as users' interests. Although they have achieved acceptable recommendation accuracy, they suffer from severe data sparse problems because of the limited news clicked by users. Further, the user's historical click sequence information has different effects on the user's interest, and simply combining them can not reflect this difference. Therefore, we propose an attention-based graph neural network news recommendation model. In our model, muti-channel convolutional neural network is used to generate news representations, and recurrent neural network is used to extract the news sequence information that users clicked on. Users, news, and topics are modeled as three types of nodes in a heterogeneous graph, and their relationships are modeled as edges. Graph neural network is used to effectively extract the structural information from heterogeneous graph, and helps to solve the problem of sparse data. Taking into account the different effects of different information on recommendation results, we use the attention mechanism to fuse this information distinctively. Extensive experiments conducted on the real online news datasets show that our model is superior to advanced deep learning-based recommendation methods.