一种用于街景影像窗户提取的神经网络

A neural network for window extraction of street view imagery

针对现有方法受类内多样性以及窗户间距较近的影响,造成漏提取和分割不足等问题。该文提出一种面向窗户提取的WBSNet模型。根据窗户在影像上的密集分布特点,加入CEB模块,用于扩大感受野,减少了目标的漏提取问题。针对相邻窗户间距较近引起的分割不足现象,引入了一种特征提取模块,在保证细节信息编码的基础上,加强网络的特征提取能力。该文在自制的街景数据集上进行实验,实验结果表明,该文方法精确率、召回率、F1-Score和交并比分别达到了76.42%、91.34%、81.82%和70.46%,验证了该文方法的有效性和可行性。

In view of the existing methods,due to the diversity of classes and the close spacing of windows,resulting in problems such as leakage extraction and insufficient segmentation.In this paper,a WBSNet model for window extraction is proposed.According to the dense distribution characteristics of windows on the image,the Context Embedding Block(CEB)module is added in this paper to expand the receptive field and reduce the leakage extraction problem of the target.Aiming at the insufficient segmentation phenomenon caused by the close spacing of adjacent windows,this paper introduces a feature extraction module(RFB)to strengthen the feature extraction capability of the network on the basis of ensuring the encoding of detailed information.The experimental results show that the Precision,Recall,F1-Score and Intersection-over-Union ratio of the proposed method reach 76.42%,91.34%,81.82%and 70.46%,respectively,which are significantly better than other methods,and verify the effectiveness and feasibility of the proposed method.