We introduce an extremely computation-efficient CNN architecture named ShuffleNet, which is designed specially for mobile devices with very limited computing power (e.g., 10-150 MFLOPs). The new utilizes two operations, pointwise group convolution and channel shuffle, to greatly reduce computation cost while maintaining accuracy. Experiments on ImageNet classification MS COCO object detection demonstrate the superior performance of ShuffleNet over other structures, e.g. lower top-1 error...

We propose DoReFa-Net, a method to train convolutional neural networks that have low bitwidth weights and activations using parameter gradients. In particular, during backward pass, gradients are stochastically quantized numbers before being propagated layers. As convolutions forward/backward passes can now operate on activations/gradients respectively, DoReFa-Net use bit convolution kernels accelerate both training inference. Moreover, as be efficiently implemented CPU, FPGA, ASIC GPU,...

Previous approaches for scene text detection have already achieved promising performances across various benchmarks. However, they usually fall short when dealing with challenging scenarios, even equipped deep neural network models, because the overall performance is determined by interplay of multiple stages and components in pipelines. In this work, we propose a simple yet powerful pipeline that yields fast accurate natural scenes. The directly predicts words or lines arbitrary...

We introduce an extremely computation-efficient CNN architecture named ShuffleNet, which is designed specially for mobile devices with very limited computing power (e.g., 10-150 MFLOPs). The new utilizes two operations, pointwise group convolution and channel shuffle, to greatly reduce computation cost while maintaining accuracy. Experiments on ImageNet classification MS COCO object detection demonstrate the superior performance of ShuffleNet over other structures, e.g. lower top-1 error...

Most graph-network-based meta-learning approaches model instance-level relation of examples. We extend this idea further to explicitly the distribution-level one example all other examples in a 1-vs-N manner. propose novel approach named distribution propagation graph network (DPGN) for few-shot learning. It conveys both relations and each learning task. To combine examples, we construct dual complete which consists point with node standing an example. Equipped architecture, DPGN propagates...

Recently, scene text detection has become an active research topic in computer vision and document analysis, because of its great importance significant challenge. However, vast majority the existing methods detect within local regions, typically through extracting character, word or line level candidates followed by candidate aggregation false positive elimination, which potentially exclude effect wide-scope long-range contextual cues scene. To take full advantage rich information available...

Previous approaches for scene text detection have already achieved promising performances across various benchmarks. However, they usually fall short when dealing with challenging scenarios, even equipped deep neural network models, because the overall performance is determined by interplay of multiple stages and components in pipelines. In this work, we propose a simple yet powerful pipeline that yields fast accurate natural scenes. The directly predicts words or lines arbitrary...

Reducing bit-widths of weights, activations, and gradients a Neural Network can shrink its storage size memory usage, also allow for faster training inference by exploiting bitwise operations. However, previous attempts quantization RNNs show considerable performance degradation when using low bit-width weights activations. In this paper, we propose methods to quantize the structure gates interlinks in LSTM GRU cells. addition, balanced further reduce degradation. Experiments on PTB IMDB...

In this paper, we propose and study a technique to reduce the number of parameters computation time in convolutional neural networks. We use Kronecker product exploit local structures within convolution fully-connected layers, by replacing large weight matrices combinations multiple products smaller matrices. Just as is generalization outer from vectors matrices, our method low rank approximation for also introduce different shapes increase modeling capacity. Experiments on SVHN, scene text...

Out-of-Distribution (OoD) detection is important for building safe artificial intelligence systems. However, current OoD methods still cannot meet the performance requirements practical deployment. In this paper, we propose a simple yet effective algorithm based on novel observation: in trained neural network, samples with bounded norms well concentrate feature space. We call center of features Feature Space Singularity (FSS), and denote distance sample to FSS as FSSD. Then, can be...

Semantic communication is a new paradigm that aims at providing more efficient for the next-generation wireless network. It focuses on transmitting extracted, meaningful information instead of raw data. However, deep learning-enabled image semantic models often require significant amount time and energy training, which unacceptable, especially mobile devices. To solve this challenge, our paper first introduces distributed system where base station local devices will collaboratively train...

Small and dim target detection is a critical challenge in hyperspectral remote sensing, particularly complex, large-scale scenes where spectral variability across diverse land cover types complicates the process. In this paper, we propose novel reasoning algorithm named Adaptive Global Dense Nested Reasoning Network (AGDNR). This integrates spatial, spectral, domain knowledge to enhance accuracy of small targets environments simultaneously enables about categories. The proposed method...

Conventional ship detection using synthetic aperture radar (SAR) is typically limited to fully focused spatial features of the target in SAR images. In this paper, we propose a multi-stage feature transfer (MFT)-based reasoning RCNN (MFT-Reasoning RCNN) detect ships This algorithm can MFT strategy and adaptive global module over all object regions by exploiting diverse knowledge between its surrounding elements. Specifically, first calculate probability simultaneous occurrence environmental...

Detection of infrared dim and small targets with diverse cluttered background plays a significant role in many applications. In this paper, we propose deep low-rank sparse patch-image network, termed as Deep-LSP-Net, to effectively detect single image. Specifically, by using the local patch construction scheme, first transform original image into patch-image, which can be decomposed superposition component target component. The detection is thus formulated an optimization problem...

Recently, text detection and recognition in natural scenes are becoming increasing popular the computer vision community as well document analysis community. However, majority of existing ideas, algorithms systems specifically designed for English. This technical report presents final results ICDAR 2015 Text Reading Wild (TRW 2015) competition, which aims at establishing a benchmark assessing devised both Chinese English scripts providing playground researchers from In this article, we...

In this paper, we propose a novel method called Residual Steps Network (RSN). RSN aggregates features with the same spatial size (Intra-level features) efficiently to obtain delicate local representations, which retain rich low-level information and result in precise keypoint localization. Additionally, observe output contribute differently final performance. To tackle problem, an efficient attention mechanism - Pose Refine Machine (PRM) make trade-off between global representations further...