Non-locally Enhanced Encoder-Decoder Network for Single Image De-raining

Abstract

Traffic flow prediction is crucial for urban traffic management and public safety. Its key challenges lie in how to adaptively integrate the various factors that affect the flow changes. In this paper, we propose a unified neural network module to address this problem, called Attentive Crowd Flow Machine (ACFM), which is able to infer the evolution of the crowd flow by learning dynamic representations of temporally-varying data with an attention mechanism. Specifically, the ACFM is composed of two progressive ConvL-STM units connected with a convolutional layer for spatial weight prediction. The first LSTM takes the sequential flow density representation as input and generates a hidden state at each time-step for attention map inference, while the second LSTM aims at learning the effective spatial-temporal feature expression from attentionally weighted crowd flow features. Based on the ACFM, we further build a deep architecture with the application to citywide crowd flow prediction, which naturally incorporates the sequential and periodic data as well as other external influences. Extensive experiments on two standard benchmarks (i.e., crowd flow in Beijing and New York City) show that the proposed method achieves significant improvements over the state-of-the-art methods.

Dual Residual Networks Leveraging the Potential of Paired Operations for Image Restoration

“…Table 8 shows the results. Those for the previous methods except RESCAN [25] are imported from [23]. It is seen that the proposed network achieves the best performance.…”

“…Li et al [25] regards a heavy rainy image as a clear image added by an accumulation of multiple rainstreak layers and proposed a RNN-based method to restore the clear image. Li et al [23] proposed an non-locally enhanced version of DenseBlock [16] for this task, their network outperforms previous approaches by a good margin. Figure 3: Four different implementations of the DuRB; c is a convolutional layer with 3×3 kernels; ct l 1 and ct l 2 are convolutional layers, each with kernels of a specified size and dilation rate; up is up-sampling (we implemented it using PixelShuffle [38]); se is SE-ResNet Module [15] that is in fact a channel-wise attention mechanism.…”

“…REDNet [26] 27.95 / 0.8019 25.75 / 0.7167 24.37 / 0.6551 MemNet [41] 28.04 / 0.8053 25.86 / 0.7202 24.53 / 0.6608 E-CAE [39] 28. 23 More details of the experiments are provided in the supplementary material. We show the quantitative results in Table 2 and qualitative results in Fig.…”

Attention-Based Adaptive Selection of Operations for Image Restoration in the Presence of Unknown Combined Distortions

“…Specifically, we created a training set by randomly sampling the standard deviations of Gaussian noise and the quality of JPEG compression from [0, 20] and [60, 100], respectively and sampling the max length of trajectories for motion blur from [10,40]. We then apply trained models to a test set with a different range of distortion parameters; the test set is created by sampling the standard deviations of Gaussian noise, the quality of JPEG compression, and the max trajectory length from [20,40], [15,60], and [40,80], respectively. We used the same image set used in the ablation study for the base image set.…”

“…Then we added Gaussian noise to them, yielding a training set and a testing set which consist of 50, 000 and 1, 000 patches, respectively. The standard deviation of the Gaussian noise is randomly chosen from the range of [10,20]. Using these datasets, we compare our method with DnCNN [51], FFDNet [52], and E-CAE [37], which are the state-of-the-art dedicated models for this task.…”