Robust Attention Deraining Network for Synchronous Rain Streaks and Raindrops Removal

<div>Synchronous Rain streaks and Raindrops Removal (SR3) is a very hard and challenging task, since rain streaks and raindrops are two wildly divergent real-scenario phenomena with different optical properties and mathematical distributions. As such, most of existing deep learning-based Singe Image Deraining (SID) methods only focus on one of them or the other. To solve this issue, we propose a new, robust and hybrid SID model, termed Robust Attention Deraining Network (RadNet) with strong robustenss and generalztion ability. The robustness of RadNet has two implications :(1) it can restore different degenerations, including raindrops, rain streaks, or both; (2) it can adapt to different data strategies, including single-type, superimposed-type and blended-type. Specifically, we first design a lightweight robust attention module (RAM) with a universal attention mechanism for coarse rain removal, and then present a new deep refining module (DRM) with multi-scales blocks for precise rain removal. The whole process is unified in a network to ensure sufficient robustness and strong generalization ability. We measure the performance of several SID methods on the SR3 task under a variety of data strategies, and extensive experiments demonstrate that our RadNet can outperform other state-of-the-art SID methods.</div>

Robust Attention Deraining Network for Synchronous Rain Streaks and Raindrops Removal

<div>Synchronous Rain streaks and Raindrops Removal (SR3) is a very hard and challenging task, since rain streaks and raindrops are two wildly divergent real-scenario phenomena with different optical properties and mathematical distributions. As such, most of existing deep learning-based Singe Image Deraining (SID) methods only focus on one of them or the other. To solve this issue, we propose a new, robust and hybrid SID model, termed Robust Attention Deraining Network (RadNet) with strong robustenss and generalztion ability. The robustness of RadNet has two implications :(1) it can restore different degenerations, including raindrops, rain streaks, or both; (2) it can adapt to different data strategies, including single-type, superimposed-type and blended-type. Specifically, we first design a lightweight robust attention module (RAM) with a universal attention mechanism for coarse rain removal, and then present a new deep refining module (DRM) with multi-scales blocks for precise rain removal. The whole process is unified in a network to ensure sufficient robustness and strong generalization ability. We measure the performance of several SID methods on the SR3 task under a variety of data strategies, and extensive experiments demonstrate that our RadNet can outperform other state-of-the-art SID methods.</div>

The Development of Multi-scale Data Management for CityGML-based 3D Buildings

The CityGML model is now the norm for smart city or digital twin city development for better planning, management, risk-related modelling and other applications. CityGML comes with five levels of detail (LoD), mainly constructed from point cloud measurements and images of several systems, resulting in a variety of accuracies and detailed models. The LoDs, also known as pre-defined multi-scale models, require large storage-memory-graphic consumption compared to single scale models. Furthermore, these multi-scales have redundancy in geometries, attributes, are costly in terms of time and workload in updating tasks, and are difficult to view in a single viewer. It is essential for data owners to engage with a suitable multi-scale spatial management solution in minimizes the drawbacks of the current implementation. The proper construction, control and management of multi-scale models are needed to encourage and expedite data sharing among data owners, agencies, stakeholders and public users for efficient information retrieval and analyses. This paper discusses the construction of the CityGML model with different LoDs using several datasets. A scale unique ID is introduced to connect all respective LoDs for cross-LoD information queries within a single viewer. The paper also highlights the benefits of intermediate outputs and limitations of the proposed solution, as well as suggestions for the future.

Research on clothing patterns generation based on multi-scales self-attention improved generative adversarial network

PurposeClothing patterns play a dominant role in costume design and have become an important link in the perception of costume art. Conventional clothing patterns design relies on experienced designers. Although the quality of clothing patterns is very high on conventional design, the input time and output amount ratio is relative low for conventional design. In order to break through the bottleneck of conventional clothing patterns design, this paper proposes a novel way based on generative adversarial network (GAN) model for automatic clothing patterns generation, which not only reduces the dependence of experienced designer, but also improve the input-output ratio.Design/methodology/approachIn view of the fact that clothing patterns have high requirements for global artistic perception and local texture details, this paper improves the conventional GAN model from two aspects: a multi-scales discriminators strategy is introduced to deal with the local texture details; and the self-attention mechanism is introduced to improve the global artistic perception. Therefore, the improved GAN called multi-scales self-attention improved generative adversarial network (MS-SA-GAN) model, which is used for high resolution clothing patterns generation.FindingsTo verify the feasibility and effectiveness of the proposed MS-SA-GAN model, a crawler is designed to acquire standard clothing patterns dataset from Baidu pictures, and a comparative experiment is conducted on our designed clothing patterns dataset. In experiments, we have adjusted different parameters of the proposed MS-SA-GAN model, and compared the global artistic perception and local texture details of the generated clothing patterns.Originality/valueExperimental results have shown that the clothing patterns generated by the proposed MS-SA-GAN model are superior to the conventional algorithms in some local texture detail indexes. In addition, a group of clothing design professionals is invited to evaluate the global artistic perception through a valence-arousal scale. The scale results have shown that the proposed MS-SA-GAN model achieves a better global art perception.

Advanced Visualization Polarimetric Imaging: Removal of Water Spray Effect Utilizing Circular Polarization

Circular polarization (CP) memory is a well-known phenomenon whereby natural light becomes partially circularly polarized after scattering by water spray several times, and the circularly polarized state can be well preserved within a certain propagation distance. In this study, a CP imaging method combined with the multi-scale analysis in the frequency domain is proposed to enhance the vision in rainy conditions. The images were first decomposed into multi-scales. CP characteristics of light were employed in the low-frequency parts to improve the quality of images in rainy conditions, and the high-frequency parts compensated specific structure information. Experimental results demonstrate that the proposed method can remove the water spray effect thereby improving the vision of degraded rainy-day images.

MCA-Unet: Multi-class Attention-aware U-net for Seismic Phase Picking

&lt;p&gt;This study presents a deep learning based algorithm for seismic event detection and simultaneous phase picking in seismic waveforms. U-net structure-based solutions which consists of a contracting path (encoder) to capture feature information and a symmetric expanding path (decoder) that enables precise localization, have proven to be effective in phase picking. The network architecture of these U-net models mainly comprise of 1D CNN, Bi- &amp; Uni-directional LSTM, transformers and self-attentive layers. Althought, these networks have proven to be a good solution, they may not fully harness the information extracted from multi-scales.&lt;/p&gt;&lt;p&gt;&amp;#160;In this study, we propose a simple yet powerful deep learning architecture by combining multi-class with attention mechanism, named MCA-Unet, for phase picking. &amp;#160;Specially, we treat the phase picking as an image segmentation problem, and incorporate the attention mechanism into the U-net structure to efficiently deal with the features extracted at different levels with the goal to improve the performance on the seismic phase picking. Our neural network is based on an encoder-decoder architecture composed of 1D convolutions, pooling layers, deconvolutions and multi-attention layers. This architecture is applied and tested to a field seismic dataset (e.g. Wenchuan Earthquake Aftershocks Classification Dataset) to check its performance.&lt;/p&gt;