Super 8 Maps

[Martta]

Im using GWT 2.7.0 with GXT 3.1.1 and IntelliJ 15.0.5 and Chrome 50. I debug my application using the Super Dev Mode (with a separate code server and bookmarklets) and up to now it's worked quite well.

However, for no apparent reason, today the Super Dev Mode has stopped working under Chrome (I can get it to work under IE 11). Everything works as before, but my *.java files are no longer visible in the Chrome Dev Tools.

I had a problem today with similar symptoms.Eventually found that in chrome dev tool, i'd blacklisted an important gwt js file (project-0.js). I just didn't want to see it in tracebacks when i was debugging some native js stuff, but with that blacklisted, the sourcemaps weren't getting pulled in! Once I un-blacklisted it, I could once again load and debug gwt java in cdt.

Battle of Halbe

A super map centered around the Battle of Halbe would be super cool. Focusing mainly on German offensive missions, but with a good chunk of Russian offensive missions. Maps here would be open in the Spree forest, as well as some maps in the town of Halbe itself. I think this would be a nice mix of urban combat and open combat.

Battle of Seelow Hights

This one was coming. A sparling map with lots of open combat. Minimal urban combat in small houses. Lots of trench combat I would assume. I would love to love see this battle as a super map!

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

In computer vision, single-image super-resolution (SISR) has been extensively explored using convolutional neural networks (CNNs) on optical images, but images outside this domain, such as those from scientific experiments, are not well investigated. Experimental data is often gathered using non-optical methods, which alters the metrics for image quality. One such example is electron backscatter diffraction (EBSD), a materials characterization technique that maps crystal arrangement in solid materials, which provides insight into processing, structure, and property relationships. We present a broadly adaptable approach for applying state-of-art SISR networks to generate super-resolved EBSD orientation maps. This approach includes quaternion-based orientation recognition, loss functions that consider rotational effects and crystallographic symmetry, and an inference pipeline to convert network output into established visualization formats for EBSD maps. The ability to generate physically accurate, high-resolution EBSD maps with super-resolution enables high-throughput characterization and broadens the capture capabilities for three-dimensional experimental EBSD datasets.

The term image super-resolution is used to describe methods designed to infer high-resolution (HR) image output from low-resolution (LR) input. Since their development, super-resolution methods have been used in applications such as surveillance and security, biometric information identification, remote sensing, astronomy, and medical imaging1. Generally, these algorithms can be categorized into three groups based on the information available during training: (a) supervised, which have paired LR-HR images during training, (b) semi-supervised, where no LR-HR image pairing is available, and (c) unsupervised, where no ground truth HR is available. Recently, because of their superior performance to traditional methods, various supervised deep CNN architectures using recursive residual blocks2,3, residual connections, and attention-based modules4,5,6,7 have seen significant use in super-resolution applications.

The approaches for different super-resolution methods vary, but they all share the common goal of producing high-resolution image output that is a clear representation of the low-resolution input in the context of both image content and visual fidelity. Generally, evaluation metrics are centered around the idea that the output image is the product of intensity-based visible-light photography, where the goal is to represent what is seen with the human eye. However for scientific image applications, this idea is often incorrect, since many experimental methods construct images or image maps using electromagnetic information gathered from outside the visible-light range (e.g., X-rays, infrared information), or even not from light at all (e.g., electrons, neutrons), which means the ideas of sharpness and visual clarity have very different meanings. One such example, electron backscatter diffraction (EBSD), used for characterization of crystalline materials, relies on electron diffraction to build maps of material crystallographic information.

EBSD has two types of resolution: the accuracy with which the EBSD pattern collected at each pixel can be indexed into a crystallographic orientation, and the spacing between pixels in a given mapping. During EBSD mapping, the electron beam must dwell at each point long enough to form a high-quality Kikuchi pattern, which is then indexed into a specific crystallographic orientation9,10,11. The indexing problem has many salient issues associated with it, among them are the crystallographic differentiation of matrix and precipitate phases, the identification of local strain effects, and the decoupling of overlapping diffraction patterns at grain boundaries. While accurate indexing is critical to EBSD, it is an independent challenge unrelated to super-resolution, as each indexing problem is treated as having no correlation with its neighbors. The lack of assumed spatial correlation separates indexing from other pixel-based problems and makes it ill-suited for SISR. Therefore, the issues associated with indexing accuracy described above are not addressed here. Instead, we consider the improvement of spatial resolution, which, in experiment, equates to collecting a higher density of data points during mapping. This requirement can lead to long mapping times or force the choice to use a coarser resolution mapping grid for expediency. The necessity to reduce EBSD collection time becomes even more critical when performing many scans during serial-sectioning 3D EBSD measurements12, where material is removed layer-by-layer with 2D maps collected at each slice, and then stacked into a 3D dataset. In almost all serial-sectioning experiments, the minimum slicing thickness/resolution is much lower than the achievable in-plane imaging resolution, creating anisotropic voxels. Furthermore, poor electrically conductive materials become electrostatically charged by the beam or degrade from beam exposure (e.g., bio materials, polymers), requiring extremely short exposure times and resulting in Kikuchi patterns with weak contrast.

As the demand for greater volume and more detailed resolution material information grows, so too does the demand and expense of EBSD mapping. Spatial resolution in EBSD is of particular interest in the characterization of deformed materials and additive manufacturing13, where subgrain misorientation gradients are used to quantify local plastic deformation effects and geometrically necessary dislocation densities14,15,16. In efforts to improve EBSD resolution and quality, simulations and experimental studies17,18,19,20,21 have shown that lowering the electron beam accelerating voltage can significantly improve the spatial resolution of EBSD maps, but map quality and achievable resolution vary with differing materials and imaging conditions. To improve indexing accuracy, multiple algorithmic approaches have been developed for better Kikuchi pattern mapping22,23, which improves both the precision and accuracy of the orientations shown at each pixel. Machine learning approaches have also been used to accelerate several tasks in the EBSD map construction process, including Kikuchi pattern indexing24, classification25, and crystal identification26. Recently, a residual-based neural network with traditional L1 loss (ResNet) was used to produce super-resolved EBSD maps from inverse pole figure (IPF) color and Euler angles as an image input27. The desire to accelerate and improve the EBSD mapping process has motivated a wide array of machine learning approaches, but many challenges still exist. One of the most prominent of these is that orientation space is discontinuous and repeating, and the fundamental shape of orientation space changes with the symmetries of the crystal being observed. This makes brute-force network learning with traditional methods highly dependent on the available data for training, and, depending on the orientation, small variations in accuracy can produce dramatically incorrect results.

Given these challenges, we present an adaptable framework for neural-network-based super-resolution of EBSD maps, where all network learning is built around the physics of crystal orientation symmetry. We define a physics-based loss that accounts for crystallographic symmetries, which is used alongside either a traditional L1 loss metric or a loss based directly on rotational arc lengths, which correspond to conventional misorientation measurements in crystallography. All super-resolution is done on crystal orientation data expressed as quaternions, meaning each pixel in a given map contains four channels. Using quaternion space allows for complete representation of orientation space, enabling a training approach that is translatable across all 230 crystallographic space groups. As a proof of concept, four state-of-the-art residual and channel-attention networks are used to generate high-resolution EBSD maps from low-resolution input using this physics-based approach. We demonstrate that regardless of network choice, physics-based approaches outperform traditional approaches both qualitatively and quantitatively. This approach has direct application to experimental EBSD measurements of electron beam-sensitive or low-conductivity materials where charge buildup and extended beam exposure are limiting factors, and for 3D EBSD data collection where high out-of-plane imaging resolution is costly. We expect SR-EBSD to accelerate EBSD mapping for defect detection and fast screening of microstructure configurations that limit material properties.

3a8082e126