Research on building extraction method based on surveillance images

Abstract

To solve the problems such as large workload, easy omission, low timeliness and low degree of automation in the method of visual identification of buildings in surveillance images, this paper studies the building extraction method based on surveillance images. In this paper, first of all, datasets of relevant scenes are collected and annotated. Then, we fine-tuned the Deeplabv3plus model to improve the accuracy of building extraction. Specifically, replace the backbone network with the resnet, the dilation rate is reduced to improve the detection accuracy of small objects, the output of the res net is combined with the output of the ASPP module through the way of skip connection, and the spatial details of the lower level and the semantic information of the higher level are fused. Besides, the multiple loss strategy is adopted. we also compared the fine-tuned model combined with different deep-level feature extraction networks with other classical semantic segmentation models on the open source CAMVID dataset, and the experiment showed that the combination of fine-tuned deeplabv3plus model and resnet50 reached the optimal IoU, F1 score and precision. In addition, we conducted an experimental comparison between the two training methods of only using the collected data training and the joint training of CAMVID dataset. The experiment shows that the model segmentation effect obtained by the joint training of data set is better. It significantly improves the details of the edge of the building, which can achieve robust extraction of the building.