Abstract: Aiming at the problem that traditional village aerial images are difficult to accurately identify national buildings,this paper analyzes the recognition results of three deep learning pixel classification algorithms. The recognition accuracy of national buildings of DeepLabV3+,U-Net and PSP-Net algorithms are 0.957,0.929 and 0.943 respectively. The DeepLabV3+ algorithm is used to test six typical areas of traditional villages. The main problems in the test results include: some roads and site areas in traditional villages are marked as national buildings,and the edges of the marked national buildings area are serrated. In order to accurately identify the national buildings in traditional villages,the national building identification results of DeepLabV3+ algorithm are improved based on ArcGIS software platform. The improved post-processing contents include marked area filtering and marked area boundary cleaning. The improved post-processing results can optimize the smoothness of the edge of the national architecture marked area of the traditional village aerial image on the basis of ensuring the accuracy of national architecture recognition.

Key words: traditional villages, national architecture, target identification, deep learning, DeepLabV3+