基于多重差异特征网络的街景变化检测

Street Scene Change Detection Based on Multiple Difference Features Network

街景变化检测对于自然灾害破坏和城市发展变化的研究起着重要作用。其主要目标是将成对的输入图片中变化的区域标注出来,其实质是二分类的语义分割问题。不同时间拍摄的街景图片可能受到如光线、天气、背景噪声、视角误差等诸多干扰因素的影响,这给传统的变化检测方法带来挑战。针对该问题,提出了一种新的神经网络模型(Multiple Difference Features Network,MDFNet)。该模型首先使用孪生网络提取成对输入图片的不同深度特征,并使用差异模块对相同深度特征计算差异,以此有效获得不同尺度的变化信息;然后通过JPU模块融合多重差异特征,在不损失细节信息的情况下提取其深层语义信息;最后使用金字塔池化模块结合全局和局部信息生成二分类的变化检测图像。在PCD数据集上的GSV和TSUNAMI部分分别采用5折交叉验证法对模型进行实验,实验结果表明,MDFNet获得了0.787和0.862的F-score,相比排名第二的DOF-CDNet方法,其值提高了约11.9%和2.9%,同时其能够更精准地分割变化细节。因此,所提模型可以有效应对干扰,对于复杂场景也具备优秀的检测能力。

Street scene change detection plays an important role in the study of natural disaster damage and urban development.Its main goal is to mark out the changing areas in the pair of input images,which is essentially a semantic segmentation problem of binary classification.There may be many interference factors such as light,weather,background noise,viewpoints error and so on when taking street view pictures at different times,which challenges traditional change detection methods.To solve this problem,a new neural network model(Multiple Difference Features Network,MDFNet)is proposed.First,siamese networks are used to extract the different depth features of pairs of input images,and the difference modules are used to calculate the difference of the same depth features to effectively obtain the change information of different depth.Then,by using JPU module to fuse multiple difference features,the deep semantic information can be extracted without losing detail information.Finally,the pyramid pooling module is used to generate the change detection image of the binary classification combined with the global and local information.MDFNet has obtained 0.787 and 0.862 F-scores in the GSV and TSUNAMI part on PCD dataset with 5 fold cross-validation,which are 11.9% and 2.9% higher than the second ranked DOF-CDNet,and can segment the change details more accurately.Therefore,the proposed model can effectively deal with interferences and has an excellent detection ability for complex scenes.