基于Deeplab V3 Plus的自适应注意力机制图像分割算法  

Image segmentation algorithm with adaptive attention mechanism based on Deeplab V3 Plus

作　　者：杨贞 彭小宝 朱强强 殷志坚 YANG Zhen;PENG Xiaobao;ZHU Qiangqiang;YIN Zhijian(College of Communication and Electronics,Jiangxi Science and Technology Normal University,Nanchang Jiangxi 330013,China)

机构地区：[1]江西科技师范大学通信与电子学院,南昌330013

出　　处：《计算机应用》2022年第1期230-238,共9页journal of Computer Applications

基　　金：国家自然科学基金资助项目(61866016,62061019);江西省自然科学基金面上项目(20202BABL202014);江西科技师范大学青年拔尖项目(2018QNBJRC002)。

摘　　要：针对Deeplab V3 Plus在下采样操作中图像细节信息和小目标信息过早丢失的问题,提出了一种基于Deeplab V3 Plus网络架构的自适应注意力机制图像语义分割算法。首先,在Deeplab V3 Plus主干网络的输入层、中间层和输出层均嵌入注意力机制模块,并且引入一个权重值与每个注意力机制模块相乘,以达到约束注意力机制模块的目的;其次,在PASCAL VOC2012公共分割数据集上训练嵌入注意力模块的Deeplab V3 Plus,以此手动获取注意力机制模块权重值(经验值);然后,探索输入层、中间层和输出层中注意力机制模块的多种融合方式;最后,将注意力机制模块的权重值更改为反向传播自动更新,从而得到注意力机制模块的最优权值和最优分割模型。实验结果表明,与原始Deeplab V3 Plus网络结构相比,引入自适应注意力机制的Deeplab V3 Plus网络结构在PASCAL VOC2012公共分割据集和植物虫害数据集上的平均交并比(MIOU)分别提高了1.4个百分点和0.7个百分点。In order to solve the problem that image details and small target information are lost prematurely in the subsampling operations of Deeplab V3 Plus,an adaptive attention mechanism image semantic segmentation algorithm based on Deeplab V3 Plus network architecture was proposed.Firstly,attention mechanism modules were embedded in the input layer,middle layer and output layer of Deeplab V3 Plus backbone network,and a weight value was introduced to be multiplied with each attention mechanism module to achieve the purpose of constraining the attention mechanism modules.Secondly,the Deeplab V3 Plus embedded with the attention modules was trained on the PASCAL VOC2012 common segmentation dataset to obtain the weight values(empirical values)of the attention mechanism modules manually.Then,various fusion methods of attention mechanism modules in the input layer,the middle layer and the output layer were explored.Finally,the weight value of the attention mechanism module was automatically updated by back propagation and the optimal weight value and optimal segmentation model of the attention mechanism module were obtained.Experimental results show that,compared with the original Deeplab V3 Plus network structure,the Deeplab V3 Plus network structure with adaptive attention mechanism has the Mean Intersection over Union(MIOU)increased by 1.4 percentage points and 0.7 percentage points on the PASCAL VOC2012 common segmentation dataset and the plant pest dataset,respectively.

关 键 词：语义分割 下采样操作 自适应注意力机制 注意力机制模块权重值 DeeplabV3 Plus 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]