Large-Scale Target Detection and Classification Based on Improved Candidate Regions

Abstract

Target detection of aerial images has become a frontier subject of concern in the image processing field. Using existing method to detect and classify large-scale building objects in aerial images, the accuracy is still a little low. This is mainly because the current method does not make full use of the prior information of the target to be detected, so there are too much redundant information in the candidate box. In this paper, our own dataset were built and then utilize the Hough transform to filter out the images that may exist in the sequence image. For images with dense lines or circles, it is possible that there is an artificial building target which will be detected, otherwise it is excluded directly. Besides, this paper exploits significance analysis from the filtered image and then extract the area of interest where the potential target is located. The results of the above-mentioned processing lay a good foundation for the subsequent detection and classification which can help improve the accuracy.

Supported by: [1] the National Natural Science Foundation’s project “Research on Multi-source Image Cooperative Detection Method Based on Biological Vision for UAV Groups” (No. 61572405). [2] Major Science and Technology Project of Shaanxi Province “Development and application demonstration of Apple’s quality and safety supervision and traceability system based on the Internet of Things”.