基于空洞空间金字塔池化和多头自注意力的特征提取网络  

Feature extraction network based on atrous spatial pyramid pooling and multi-head self-attention

作　　者：万黎明 张小乾[2] 刘知贵[2] 李理[1] WAN Liming;ZHANG Xiaoqian;LIU Zhigui;LI Li(College of Computer Science and Technology,Southwest University of Science and Technology,Mianyang Sichuan 621000,China;College of Information Engineering,Southwest University of Science and Technology,Mianyang Sichuan 621000,China)

机构地区：[1]西南科技大学计算机科学与技术学院,四川绵阳621000 [2]西南科技大学信息工程学院,四川绵阳621000

出　　处：《计算机应用》2022年第S02期79-85,共7页journal of Computer Applications

基　　金：国家自然科学基金资助项目(61772272,62102331)。

摘　　要：针对深度学习在图像处理领域中多尺度特征提取能力弱、特征内部信息捕获能力差的问题,提出了一种基于空洞空间金字塔池化和多头自注意力的特征提取网络(PPSANet)。首先,引入小扩张率的空洞卷积对空洞空间金字塔池化(ASPP)模型进行改进,提高局部特征信息的感受野;其次,将改进的ASPP模型合并到残差网络(ResNet)的每个残差块中,使网络在多个维度上都具有多尺度特征提取能力;最后,将残差网络的底层残差块替换为多头自注意力(MHSA),增强网络特征学习能力,捕获数据和特征内部的相关性。图像分割实验中,与残差网络相比,在肺结节数据集中DICE相似系数(DICE)提升了5.16个百分点,肝癌数据集中DICE提升了5.22个百分点;目标检测实验中,与残差网络相比,平均精度均值(MAP)提升了2.9个百分点。实验结果表明,PPSANet能够有效解决图像处理中多尺度特征提取能力弱和内部信息捕获能力差的问题,在一定程度上提高了图像处理的能力。To address the problems of weak multi-scale feature extraction ability and poor internal information capture ability of deep learning in the field of image processing,a feature extraction Network based on atrous spatial Pyramid Pooling and multi-head Self-Attention(PPSANet)was proposed.Firstly,the improved Atrous Spatial Pyramid Pooling(ASPP)model was modified by introducing an atrous convolution with small dilation rate to improve the perceptual field of local feature information.Secondly,ASPP model was merged into each residual block of Residual Network(ResNet),so that the network had multi-scale feature extraction ability in multiple dimensions.Finally,the bottom residual block of the residual network was replaced by Multi Head Self-Attention(MHSA)to enhance the learning ability of network features and capture the internal correlation between data and features.In the image segmentation experiment,compared with the residual network,the DICE similarity coefficient(DICE)in the pulmonary nodule dataset increased by 5.16 percentage points,and the DICE in the liver cancer dataset increased by 5.22 percentage points.In the object detection experiments,the Mean Average Precision(MAP)was improved by 2.9 percentage points compared with the residual network.Experimental results show that PPSANet can effectively solve the problems of weak multi-scale feature extraction ability and poor internal information capture ability in image processing,and improve the ability of image processing to a certain extent.

关 键 词：深度学习 特征提取 图像分割 目标检测 自注意力 空洞空间金字塔池化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]