Research On New Building Detection Based On Improved DeepLabV3+ Algorithm

The detection of new buildings is the process of analyzing the bi-temporal image images taken at different times in the same area and determining the new buildings.In recent years,convolutional neural networks have shown amazing performance in image classification tasks,and a large number of researchers have combined them with remote sensing detection tasks.However,the detection of new buildings is easily affected by interference such as light intensity and multi-scale targets.Carrying out relevant work for the above problems,and the main research contents of the paper are as follows:(1)Interfering factors such as seasonal changes and illumination changes in the bi-temporal remote sensing image will affect the generalization ability of the model.The data enhancement method of color gamut transformation and random splicing is added to enrich the background information of the image,reduce the interference caused by external factors,and improve the generalization ability of the model.(2)Extracting new buildings from bitemporal image images can be regarded as a binary classification semantic segmentation problem.The SN-DeepLabV3+model based on the Siamese Network improved deeplabv3+model,the PSN-DeepLabV3+ model based on the Pseudo-Siamese Network improved DeepLabV3+model and the 2c-DeepLabV3+model based on the 2-Channels network improved deeplabv3+model solve the problem that the amount of input data in the new building detection task is inconsistent with that in the building detection task.(3)Aiming at the problem of segmentation accuracy in the DeepLabV3+algorithm,the DeepLabV3+algorithm is improved and the O-DeepLabV3+model is proposed.On the basis of the ASPP module,the SKNet attention mechanism is used to weigh the multi-scale features and improve the multi-scale expression ability of the model.At the same time,SCSE is added in the up-sampling process.The attention mechanism weighs shallow features and deep features,improves the contextual expression ability of the model,and then improves the accuracy of the model.In view of the problem of poor edge effect of segmentation,firstly,focal loss is used to replace the standard cross entropy loss function,and then the edge information of buildings is obtained by edge detection using the label map of buildings,and finally the edge information is combined on the basis of focal loss to increase Loss weights at the edges of buildings to optimize the detection performance of edge parts.And based on Siamese Network,Pseudo-Siamese Network,2-channels network to improve O-DeepLabV3+algorithm,proposed SN-O-DeepLabV3+,PSN-O-DeepLabV3+,2c-O-DeepLabV3+detection model.In order to verify the performance of the improved network,the superiority of the algorithm is verified on the LEVIR-CD dataset,and relevant ablation experiments are carried out.The experimental results show that the algorithm proposed in this paper has higher detection accuracy,and the edges of buildings are also optimized.