Low-level Optimizations for Faster Mobile Deep Learning Inference Frameworks

2020 ◽  
Author(s):  
Mathieu Febvay
Keyword(s):  
Deep Learning ◽  
Low Level

scholarly journals Building Extraction in Very High Resolution Imagery by Dense-Attention Networks

2018 ◽  
Vol 10 (11) ◽  
pp. 1768 ◽  
Author(s):  
Hui Yang ◽  
Penghai Wu ◽  
Xuedong Yao ◽  
Yanlan Wu ◽  
Biao Wang ◽  
...  
Keyword(s):  
Deep Learning ◽  
High Resolution ◽  
Building Extraction ◽  
Learning Networks ◽  
Feature Maps ◽  
Low Level ◽  
High Resolution Imagery ◽  
Very High Resolution Imagery ◽  
High Level ◽  
Very High

Building extraction from very high resolution (VHR) imagery plays an important role in urban planning, disaster management, navigation, updating geographic databases, and several other geospatial applications. Compared with the traditional building extraction approaches, deep learning networks have recently shown outstanding performance in this task by using both high-level and low-level feature maps. However, it is difficult to utilize different level features rationally with the present deep learning networks. To tackle this problem, a novel network based on DenseNets and the attention mechanism was proposed, called the dense-attention network (DAN). The DAN contains an encoder part and a decoder part which are separately composed of lightweight DenseNets and a spatial attention fusion module. The proposed encoder–decoder architecture can strengthen feature propagation and effectively bring higher-level feature information to suppress the low-level feature and noises. Experimental results based on public international society for photogrammetry and remote sensing (ISPRS) datasets with only red–green–blue (RGB) images demonstrated that the proposed DAN achieved a higher score (96.16% overall accuracy (OA), 92.56% F1 score, 90.56% mean intersection over union (MIOU), less training and response time and higher-quality value) when compared with other deep learning methods.

scholarly journals General recipe to form input space for deep learning analysis of HEP scattering processes

2020 ◽  
Vol 35 (21) ◽  
pp. 2050119
Author(s):  
Lev Dudko ◽  
Georgi Vorotnikov ◽  
Petr Volkov ◽  
Maxim Perfilov ◽  
Andrei Chernoded ◽  
...  
Keyword(s):  
Deep Learning ◽  
Multivariate Data Analysis ◽  
Hard Scattering ◽  
Input Space ◽  
Low Level ◽  
Multivariate Technique ◽  
Deep Learning Neural Network ◽  
Learning Analysis ◽  
Sophisticated Analysis ◽  
Kinematic Properties

Deep learning neural network (DNN) technique is one of the most efficient and general approach of multivariate data analysis of the collider experiments. The important step of the analysis is the optimization of the input space for multivariate technique. In the paper we propose the general recipe how to form the set of low-level observables sensitive to the differences in hard scattering processes at the colliders. It is shown in the paper that without any sophisticated analysis of the kinematic properties one can achieve close to optimal performance of DNN with the proposed general set of low-level observables.

Photodamage Reduction on Harmonic Generation Microscopy at Low-Level Optical Power based on Deep Learning

2021 ◽  
Author(s):  
Yi-Jiun Shen ◽  
En-Yu Liao ◽  
Tsung-Ming Tai ◽  
Yi-Hua Liao ◽  
Chi-Kuang Sun ◽  
...  
Keyword(s):  
Deep Learning ◽  
Harmonic Generation ◽  
Optical Power ◽  
Low Level

Saliency detection via integrating deep learning architecture and low-level features

2019 ◽  
Vol 352 ◽  
pp. 75-92 ◽  
Author(s):  
Jianning Chi ◽  
Chengdong Wu ◽  
Xiaosheng Yu ◽  
Hao Chu ◽  
Peng Ji
Keyword(s):  
Deep Learning ◽  
Saliency Detection ◽  
Low Level

scholarly journals DEEP LEARNING FOR 3D RECONSTRUCTION OF THE MARTIAN SURFACE USING MONOCULAR IMAGES: A FIRST GLANCE

2020 ◽  
Vol XLIII-B3-2020 ◽  
pp. 1111-1116
Author(s):  
Z. Chen ◽  
B. Wu ◽  
W. C. Liu
Keyword(s):  
Deep Learning ◽  
3D Reconstruction ◽  
Experimental Results ◽  
Training Data ◽  
Network System ◽  
Stereo Images ◽  
Martian Surface ◽  
Preliminary Results ◽  
Low Level ◽  
Traditional Technology

Abstract. The paper presents our efforts on CNN-based 3D reconstruction of the Martian surface using monocular images. The Viking colorized global mosaic and Mar Express HRSC blended DEM are used as training data. An encoder-decoder network system is employed in the framework. The encoder section extracts features from the images, which includes convolution layers and reduction layers. The decoder section consists of deconvolution layers and is to integrate features and convert the images to desired DEMs. In addition, skip connection between encoder and decoder section is applied, which offers more low-level features for the decoder section to improve its performance. Monocular Context Camera (CTX) images are used to test and verify the performance of the proposed CNN-based approach. Experimental results show promising performances of the proposed approach. Features in images are well utilized, and topographical details in images are successfully recovered in the DEMs. In most cases, the geometric accuracies of the generated DEMs are comparable to those generated by the traditional technology of photogrammetry using stereo images. The preliminary results show that the proposed CNN-based approach has great potential for 3D reconstruction of the Martian surface.

Improved YOLOv3 with duplex FPN for object detection based on deep learning

2021 ◽  
pp. 002072092098352
Author(s):  
Seokyong Shin ◽  
Hyunho Han ◽  
Sang Hun Lee
Keyword(s):  
Deep Learning ◽  
Object Detection ◽  
Autonomous Vehicles ◽  
Detection Accuracy ◽  
Small Object ◽  
Feature Maps ◽  
Low Level ◽  
Small Object Detection ◽  
High Level ◽  
Networks Structure

YOLOv3 is a deep learning-based real-time object detector and is mainly used in applications such as video surveillance and autonomous vehicles. In this paper, we proposed an improved YOLOv3 (You Only Look Once version 3) applied Duplex FPN, which enhanced large object detection by utilizing low-level feature information. The conventional YOLOv3 improved the small object detection performance by applying FPN (Feature Pyramid Networks) structure to YOLOv2. However, YOLOv3 with an FPN structure specialized in detecting small objects, so it is difficult to detect large objects. Therefore, this paper proposed an improved YOLOv3 applied Duplex FPN, which can utilize low-level location information in high-level feature maps instead of the existing FPN structure of YOLOv3. This improved the detection accuracy of large objects. Also, an extra detection layer was added to the top-level feature map to prevent failure of detection of parts of large objects. Further, dimension clusters of each detection layer were reassigned to learn quickly how to accurately detect objects. The proposed method was compared and analyzed in the PASCAL VOC dataset. The experimental results showed that the bounding box accuracy of large objects improved owing to the Duplex FPN and extra detection layer, and the proposed method succeeded in detecting large objects that the existing YOLOv3 did not.

scholarly journals CRISPR-VAE: A Method for Explaining CRISPR/Cas12a Predictions, and an Efficiency-aware gRNA Sequence Generator

2021 ◽  
Author(s):  
Ahmad Shaker Obeid ◽  
Hasan Al Marzouqi
Keyword(s):  
Deep Learning ◽  
Genome Editing ◽  
Real Data ◽  
Black Box ◽  
Accurate Prediction ◽  
Great Promise ◽  
Adequate Explanation ◽  
Low Level ◽  
Sequence Generator ◽  
Target Effects

Motivation: Sizeable research has been conducted to facilitate the usage of CRISPR-Cas systems in genome editing, in which deep learning-based methods among others have shown great promise in the prediction of the gRNA efficiency. An accurate prediction of gRNA efficiency helps practitioners optimize their engineered gRNAs, maximizing the on-target efficiency, and minimizing the off-target effects. However, the black box prediction of deep learning-based methods does not provide adequate explanation to the factors that make a sequence efficient; rectifying this issue can help promote the usage of CRISPR-Cas systems in numerous domains. Results: We put forward a framework for interpreting gRNA efficiency prediction, dubbed CRISPR-VAE, and apply it to CRISPR/Cpf1. We thus help open the door to a better interpretability of the factors that make a certain gRNA efficient. We further lay out a semantic articulation of such factors into position-wise k-mer rules. The paradigm consists of building an efficiency-aware gRNA sequence generator trained on available real data, and using it to generate a large amount of synthetic sequences with favorable traits, upon which the explanation of the gRNA prediction is based. CRISPR-VAE can further be used as a standalone sequence generator, where the user has access to a low-level editing control. The framework can be readily integrated with different CRISPR-Cas tools and datasets, and its efficacy is confirmed in this paper.

scholarly journals Combining Low-Level Image Features with Features from A Simple Convolutional Neural Network

2019 ◽  
pp. 136-139
Author(s):  
Ozge Oztimur Karadag ◽  
Ozlem Erdas
Keyword(s):  
Image Processing ◽  
Deep Learning ◽  
Image Data ◽  
Image Features ◽  
Data Set ◽  
Low Level ◽  
Learning Techniques ◽  
Processing Techniques ◽  
Layered Architectures ◽  
Traditional Image

In the traditional image processing approaches, first low-level image features are extracted and then they are sent to a classifier or a recognizer for further processing. While the traditional image processing techniques employ this step-by-step approach, majority of the recent studies prefer layered architectures which both extract features and do the classification or recognition tasks. These architectures are referred as deep learning techniques and they are applicable if sufficient amount of labeled data is available and the minimum system requirements are met. Nevertheless, most of the time either the data is insufficient or the system sources are not enough. In this study, we experimented how it is still possible to obtain an effective visual representation by combining low-level visual features with features from a simple deep learning model. As a result, combinational features gave rise to 0.80 accuracy on the image data set while the performance of low-level features and deep learning features were 0.70 and 0.74 respectively.

scholarly journals A Multi-Scale Deep Neural Network for Water Detection from SAR Images in the Mountainous Areas

2020 ◽  
Vol 12 (19) ◽  
pp. 3205
Author(s):  
Lifu Chen ◽  
Peng Zhang ◽  
Jin Xing ◽  
Zhenhong Li ◽  
Xuemin Xing ◽  
...  
Keyword(s):  
Deep Learning ◽  
Domain Knowledge ◽  
Contextual Information ◽  
Improvement Strategy ◽  
Sar Images ◽  
Low Level ◽  
Multi Scale ◽  
Water Detection ◽  
Multi Level ◽  
End To End

Water detection from Synthetic Aperture Radar (SAR) images has been widely utilized in various applications. However, it remains an open challenge due to the high similarity between water and shadow in SAR images. To address this challenge, a new end-to-end framework based on deep learning has been proposed to automatically classify water and shadow areas in SAR images. This end-to-end framework is mainly composed of three parts, namely, Multi-scale Spatial Feature (MSF) extraction, Multi-Level Selective Attention Network (MLSAN) and the Improvement Strategy (IS). Firstly, the dataset is input to MSF for multi-scale low-level feature extraction via three different methods. Then, these low-level features are fed into the MLSAN network, which contains the Encoder and Decoder. The Encoder aims to generate different levels of features using residual network of 101 layers. The Decoder extracts geospatial contextual information and fuses the multi-level features to generate high-level features that are further optimized by the IS. Finally, the classification is implemented with the Softmax function. We name the proposed framework as MSF-MLSAN, which is trained and tested using millimeter wave SAR datasets. The classification accuracy reaches 0.8382 and 0.9278 for water and shadow, respectively; while the overall Intersection over Union (IoU) is 0.9076. MSF-MLSAN demonstrates the success of integrating SAR domain knowledge and state-of-the-art deep learning techniques.

Sign in / Sign up

Export Citation Format

Share Document