Scene text recognition has been a hot research topic in computer vision due to its various applications. The state of the art is the attention-based encoder-decoder framework that learns the mapping between input images and output sequences in a purely data-driven way. However, we observe that existing attention-based methods perform poorly on complicated and/or low-quality images. One major reason is that existing methods cannot get accurate alignments between feature areas and targets for such images. We call this phenomenon "attention drift". To tackle this problem, in this paper we propose the FAN (the abbreviation of Focusing Attention Network) method that employs a focusing attention mechanism to automatically draw back the drifted attention. FAN consists of two major components: an attention network (AN) that is responsible for recognizing character targets as in the existing methods, and a focusing network (FN) that is responsible for adjusting attention by evaluating whether AN pays attention properly on the target areas in the images. Furthermore, different from the existing methods, we adopt a ResNet-based network to enrich deep representations of scene text images. Extensive experiments on various benchmarks, including the IIIT5k, SVT and ICDAR datasets, show that the FAN method substantially outperforms the existing methods.Comment: Revise the description of IC15 datasets (1811 samples

The India‐Eurasia collision zone is the largest deforming region on the planet; direct measurements of present‐day deformation from Global Positioning System (GPS) have the potential to discriminate between competing models of continental tectonics. But the increasing spatial resolution and accuracy of observations have only led to increasingly complex realizations of competing models. Here we present the most complete, accurate, and up‐to‐date velocity field for India‐Eurasia available, comprising 2576 velocities measured during 1991–2015. The core of our velocity field is from the Crustal Movement Observation Network of China‐I/II: 27 continuous stations observed since 1999; 56 campaign stations observed annually during 1998–2007; 1000 campaign stations observed in 1999, 2001, 2004, and 2007; 260 continuous stations operating since late 2010; and 2000 campaign stations observed in 2009, 2011, 2013, and 2015. We process these data and combine the solutions in a consistent reference frame with stations from the Global Strain Rate Model compilation, then invert for continuous velocity and strain rate fields. We update geodetic slip rates for the major faults (some vary along strike), and find that those along the major Tibetan strike‐slip faults are in good agreement with recent geological estimates. The velocity field shows several large undeforming areas, strain focused around some major faults, areas of diffuse strain, and dilation of the high plateau. We suggest that a new generation of dynamic models incorporating strength variations and strain‐weakening mechanisms is required to explain the key observations. Seismic hazard in much of the region is elevated, not just near the major faults.

With the continuous development of space and sensor technologies during the recent 40 years, ocean remote sensing has entered into the Big Data era with typical Five-V (volume, variety, value, velocity, and veracity) characteristics. Ocean remote sensing data archives reach several tens of petabytes, and massive satellite data are acquired worldwide daily. To precisely, efficiently and intelligently mining the useful information submerged in such ocean remote sensing data sets is a big challenge. Deep learning, a powerful technology recently emerging in the machine-learning field, has demonstrated its more significant superiority over traditional physical- or statistical-based algorithms for image information extraction in many industrial-field applications and starts to draw interest in ocean remote sensing applications. In this review paper, we first systematically reviewed two deep learning frameworks that carry out ocean remote sensing image classifications and then presented eight typical applications in ocean internal wave/eddy/oil spill/coastal inundation/sea-ice/green algae/ship/coral reef mapping from different types of ocean remote sensing imagery to show how effective of these deep learning frameworks. Researchers can also readily modify these existing frameworks for information mining of other kinds of remote sensing imagery.

Forecasting fields of oceanic phenomena has long been dependent on physical equation–based numerical models. The challenge is that many natural processes need to be considered for understanding complicated phenomena. In contrast, rules of the processes are already embedded in the time-series observation itself. Thus, inspired by largely available satellite remote sensing data and the advance of deep learning technology, we developed a purely satellite data–driven deep learning model for forecasting the sea surface temperature evolution associated with a typical phenomenon: a tropical instability wave. During the testing period of 9 years (2010–2019), our model accurately and efficiently forecasts the sea surface temperature field. This study demonstrates the strong potential of the satellite data–driven deep learning model as an alternative to traditional numerical models for forecasting oceanic phenomena.