Small Object Detection in High-Resolution Images Based on Multiscale Detection and Re-training

GUANG-MIN SUN, JIA-YANG CHEN, BING LI, DONG YAN, YU LI, GANG XIE

Abstract


Most of the current small object detection algorithms are designed for low-resolution images. They can neither directly process high-resolution images nor make full use of the information contained. In this paper, an algorithm is proposed to detect small objects in highresolution images directly. The process of the algorithm is as follows: Firstly, an original detection task is split logically into several relevant detection sub-tasks at different scales. Secondly, a corresponding low-resolution object detector is trained for each sub-task. Thirdly, the detectors are deployed to get detection results at different scales. Finally, the multi-scale detection results are logically combined to derive the final detection results of the small objects. Besides, the detector is re-trained at small scale by introducing negative samples. The algorithm proposed in this paper was tested in the task of defect detection on building wall surface. Experimental results show the reliability and efficiency of our approach.

Keywords


High resolution image, Small object detection, Multiscale, Negative feedback mechanism, Defect detection of building wall surface.Text


DOI
10.12783/dtcse/cmso2019/33598

Full Text:

PDF

Refbacks

  • There are currently no refbacks.