基于MoblieNet v2的图像语义分割网络

Image semantic segmentation network based on MobileNet v2

语义分割是计算机视觉重要的一环,其核心是对图像中的每个像素进行分类与定位,需要耗费巨量计算资源.针对经典网络参数多且分割速度慢的问题,本文基于MobileNet v2提出一种轻量级的图像语义分割网络.首先,将轻量级网络MobileNet v2与空洞卷积相结合作为语义分割的特征提取网络.其次,提出条形注意力模块以捕获更多上下文信息,并利用池化来降低计算成本.最后,以类内与类间关系作为约束条件设计了新的辅助损失函数,提升网络的判别能力.建议的网络能够权衡计算量与分割精度之间的关系,并能以较小的计算成本捕获更多的全局信息.在公共数据集PASCAL VOC 2012和遥感数据集DLRSD、WHDLD上进行了大量的测试表明,建议的方法能够有效提升分割效果,在三个数据集上分别取得了71.7%、60.8%和59.3%的平均交并比.

Semantic segmentation is an important part of computer vision,and its core is to classify and locate each pixel in the image,which requires a huge amount of computing resources.Aiming at the problem of the classical network with many parameters and slow segmentation speed,a lightweight image semantic segmentation network based on MobileNet v2 is proposed.Firstly,combine the lightweight network MobileNet v2 with hole convolution as a feature extraction network.Secondly,strip position attention modules are introduced to capture more contextual information,and pooling is used to reduce computing costs.Finally,with the relationship between inner-classes and inter-classes as constraints,a new auxiliary loss function is proposed,which improves the discriminative ability of the network.The proposed network can trade off the relationship between computation amount and segmentation accuracy.The PASCAL VOC 2012 and the remote sensing datasets DLRSD and WHDLD have been subjected to numerous tests.The experimental results show that the proposed method can effectively improve the segmentation effect,and the proposed method obtained mIoU(Mean Intersection over Union)71.7%,60.8%,and 59.3%respectively.