Abstract:In recent years, the optimization of network architecture plays an increasingly important role in the performance improvement of neural networks. We introduce an interactive dual-branch attention mechanism and three different lightweight-oriented strategies to build an accurate and compact residual network model in this work. The channel attention and spatial attention are fused to construct a novel bottleneck to enhance the feature representation ability for accurate performance. Asymmetric convolutions with spatial factorization, channel splitting, depthwise separable convolution with width multiplier adjustment are further combined to compress the parameter size of the attention-driven model for a lightweight and compact residual network named ALResNet. The experimental results of 92.1% top-1 testing accuracy at the inference speed of 14.90 fps on Animals-10 and 89.4% top-1 testing accuracy at the inference speed of 16.21 fps on CIFAR-10, as well as 4.77M parameters and 736.82 MFLOPs, demonstrate that the proposed ALResNet achieves a decent tradeoff between accuracy and computing efficiency for fast inference on resource-limited mobile devices for vision-based tasks. ∗Corresponding author: Aiwen Luo (faith.awluo@gmail.com). Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. MLMI’21, September 17–19, 2021, Hangzhou, China © 2021 Association for Computing Machinery. ACM ISBN 978-1-4503-8424-7/21/09. . . $15.00 https://doi.org/10.1145/3490725.3490729 CCS CONCEPTS • Computing methodologies; • Object recognition; • Neural networks; • Model verification and validation;

RA-Net: reverse attention for generalizing residual learning

A Multi-Scale Residual Attention Network for Retinal Vessel Segmentation

Look Closer to See Better: Recurrent Attention Convolutional Neural Network for Fine-Grained Image Recognition

RAANet: A Residual ASPP with Attention Framework for Semantic Segmentation of High-Resolution Remote Sensing Images

Recurrent Deep Attention Network for Person Re-Identification.

GRA-Net: Group Response Attention for Deep Learning

Residual Attention Unit for Action Recognition.

Residual Attention Encoding Neural Network for Terrain Texture Classification.

Rega-Net: Retina Gabor Attention for Deep Convolutional Neural Networks

Cerebrovascular Segmentation in MRA Via Reverse Edge Attention Network

Residual Attribute Attention Network for Face Image Super-Resolution.

Cross-Layer Retrospective Retrieving via Layer Attention

Research on image classification based on residual group multi-scale enhanced attention network

ALResNet: Attention-Driven Lightweight Residual Network for Fast and Accurate Image Recognition

Residual Alignment: Uncovering the Mechanisms of Residual Networks

Recurrent Attention Model for Pedestrian Attribute Recognition.

Attention Based Residual Network For High-Resolution Remote Sensing Imagery Scene Classification

Recurrent Attention Unit

Parallel Residual Attention Network for Image Super-Resolution

RNA-Net: Residual Nonlocal Attention Network for Retinal Vessel Segmentation.

Residual Inception: A New Module Combining Modified Residual with Inception to Improve Network Performance