The nonlinear two-time-scale stochastic approximation is widely studied under conditions of bounded variances in noise. Motivated by recent advances that allow for variability linked to the current state or time, we consider state- and time-dependent noises. We show that the Lyapunov function exhibits polynomial convergence rates in both cases, with the rate of polynomial delay depending on the parameters of state- or time-dependent noises. Notably, if the state noise parameters fully approach their limiting value, the Lyapunov function achieves an exponential convergence rate. We provide two numerical examples to illustrate our theoretical findings in the context of stochastic gradient descent with Polyak-Ruppert averaging and stochastic bilevel optimization.

Calibration-based methods have dominated RAW image denoising under extremely low-light environments. However, these methods suffer from several main deficiencies: 1) the calibration procedure is laborious and time-consuming, 2) denoisers for different cameras are difficult to transfer, and 3) the discrepancy between synthetic noise and real noise is enlarged by high digital gain. To overcome the above shortcomings, we propose a calibration-free pipeline for Lighting Every Drakness (LED), regardless of the digital gain or camera sensor. Instead of calibrating the noise parameters and training repeatedly, our method could adapt to a target camera only with few-shot paired data and fine-tuning. In addition, well-designed structural modification during both stages alleviates the domain gap between synthetic and real noise without any extra computational cost. With 2 pairs for each additional digital gain (in total 6 pairs) and 0.5% iterations, our method achieves superior performance over other calibration-based methods.

Medical segmentation of cochlear images aims to identify the area of the cochlea in a set of CT slices. The shape of cochlea will vary a quite in different CT slicing levels, and the relevant dataset has a higher labeling cost. This will lead to segmentation results with edge discontinuity when we implement supervised algorithm under few samples. In order to solve the problem of a small number of labeled images, this paper proposes a semi-supervised model called GSDNet which is based on GAN, which captures the features of the cochlear image without labels, so as to achieve high performance for processing fewer sampled data. To further improve the generalization of the model, we adopt a training method that allows the model to gradually distinguish between real images and fake images. In addition, in order to solve the problem of local noise interference and discontinuous segmentation results, we introduce a label discrimination network to force the distribution of generated results from segmentation network to align with the true label distribution, so that the edges of the segmentation results are continuous and the shape is more accurate. Finally, we conduct a segmentation experiment of the cochlear region containing 30 slices about cochlea data, and compare different cutting-edge methods. The method proposed in this paper achieves higher performance on the dice index.

Lighting is a determining factor in photography that affects the style, expression of emotion, and even quality of images. Creating or finding satisfying lighting conditions, in reality, is laborious and time-consuming, so it is of great value to develop a technology to manipulate illumination in an image as post-processing. Although previous works have explored techniques based on the physical viewpoint for relighting images, extensive supervisions and prior knowledge are necessary to generate reasonable images, restricting the generalization ability of these works. In contrast, we take the viewpoint of image-to-image translation and implicitly merge ideas of the conventional physical viewpoint. In this paper, we present an Illumination-Aware Network (IAN) which follows the guidance from hierarchical sampling to progressively relight a scene from a single image with high efficiency. In addition, an Illumination-Aware Residual Block (IARB) is designed to approximate the physical rendering process and to extract precise descriptors of light sources for further manipulations. We also introduce a depth-guided geometry encoder for acquiring valuable geometry- and structure-related representations once the depth information is available. Experimental results show that our proposed method produces better quantitative and qualitative relighting results than previous state-of-the-art methods. The code and models are publicly available on https://github.com/NK-CS-ZZL/IAN.

Bounding box regression is a very effective method to improve the localization accuracy of object detection. Recently, the IoU-based regression losses have been widely used in object detection algorithms. However, we observe that they degenerate seriously in the late training period, leading to slow convergence and inaccurate localization. In this paper, we design a Fast Intersection over Union (FIoU) loss, which can not only keep the advantages but also solve the weakness of IoU-based losses. Furthermore, FIoU can be directly applied to Non-Maximum Suppression (NMS) as a criterion to improve the localization performance. Numerous experiments on two popular benchmark datasets show that our method is superior to other the-state-of-art methods.

Due to the time-consuming nature and low accuracy of traditional point cloud-based defect detection methods, as well as the scarcity of reconstructed surface workpieces, the excessive amount of surface point cloud data, and the small proportion of defects, it is difficult to effectively segment the defects. To solve these problems, we propose a fusion prediction model named FPP-PreNet (Fusion of Projection and Point Network) to quantitatively estimate the proportion of defects. Based on the PointNet classification model, it treats defect proportion detection as a regression task. Considering the difference between classification and regression, it adjusts the activation and loss functions, replaces the maximum pooling operation, and integrates the orientation information for multi-feature learning and fusion. The global feature descriptors obtained by a projection network will be aggregated, and then the defect proportion prediction value of the point cloud slice is obtained through the fully connected layer and the activation function. We also propose a dataset construction strategy based on point cloud slices and merged orientation information to reduce the amount of calculation and expand the dataset when there is non-zero curvature. Finally, the effectiveness of the dataset and prediction model is verified through experiments.

Few-shot semantic segmentation methods aim for predicting the regions of different object categories with only a few labeled samples. It is difficult to produce segmentation results with high accuracy when a new category appears. In this paper, we propose a Multi-scale Discriminative Location-aware (MDL) network to tackle the few-shot semantic segmentation problem. In order to use information from different levels, we first keep the last three convolutional layers of FCN, and then use the VGG-16 network to extract features from the support image-label pair, which adjusts the weight of the query image segmentation branch. Discriminative location-aware architecture can improve the efficiency of few-shot segmentation, and therefore the global average pooling layer is added to produce location feature information. Finally, we evaluate our MDL model on the Pascal VOC 2012 challenge, and show that it achieves competitive mIoU score compared to methods in recent years.

Convolutional neural networks (CNNs) have recently achieved impressive performances in image processing tasks such as image classification and object recognition. However, CNNs only process images in the spatial domain whereas spectral analysis operates in the frequency domain. In this paper, we propose the 2D wavelet transform layer. The learned features from images are viewed as two-directional sequential data, and we use two LSTM layers that sweep both horizontally and vertically across the image to compress feature matrices. Based on this, the 2D wavelet transform decomposes the above feature matrices as a learned mixing of different harmonic functions, and therefore integrating the spectral analysis into CNNs. We also select 3 × 3 convolutional mixing style with Gaussian+LSM filter to mix the output of the 2D wavelet transform to generate new output features. Finally, we combine the sequential and spectral features to build our CNN-RNN architecture with skip layers and apply it to image classification. Our proposed network is evaluated on three widely-used benchmark datasets: CIFAR-10, CIFAR-100 and Tiny ImageNet. Experiments show that our CNN-RNN hybrid model achieves better accuracy in image classification tasks.

The combination of Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) have played important roles in deep learning in recent years to improve the prediction performance, especially in the context of temporal data analysis. Previous research has shown that certain time series could have common time-dependent characteristics. Therefore, in order to make good prediction, it is necessary to take into account the correlation between different temporal data in modeling. However, general RNN models have serious limitation to achieve this goal. In this paper, a new architecture, Deep and Wide Neural Networks (DWNN), is proposed, where CNN's convolution layer is added to the RNN's hidden state transfer process. CNN is combined with RNN to extract the correlation characteristics of different RNN models while RNNs running along the time steps. This new architecture not only has the depth of RNN in the time dimension, but also has the width of the number of temporal data. The intuition behind the DWNN model, as well as different kinds of DWNN model structures are discussed in this paper. We use stock data from the sandstorm sector of Shanghai Stock Exchange for our experiment. As shown in the result, our proposed DWNN model can reduce the prediction mean squared error by 30% compared with the general RNN model.

General recommender and sequential recommender are two commonly applied modeling paradigms for recommendation tasks. General recommender focuses on modeling the general user preferences, ignoring the sequential patterns in user behaviors; whereas sequential recommender focuses on exploring the item-to-item sequential relations, failing to model the global user preferences. In addition, better recommendation performance has recently been achieved by adopting an approach to combine them. However, previous approaches are unable to solve both tasks in a unified way and cannot capture the whole historical sequential information. In this paper, we propose a recommendation model named Recurrent Collaborative Filtering (RCF), which unifies both paradigms within a single model.Specifically, we combine recurrent neural network (the sequential recommender part) and matrix factorization model (the general recommender part) in a multi-task learning framework, where we perform joint optimization with shared model parameters enforcing the two parts to regularize each other. Furthermore, we empirically demonstrate on MovieLens and Netflix datasets that our model outperforms the state-of-the-art methods across the tasks of both sequential and general recommender.