본문 바로가기
  • Home

Design and Implementation of an Illegal Crop Object Detection System using Image Slicing and Transformer Techniques to High-Resolution Drone Data

  • Journal of Software Assessment and Valuation
  • Abbr : JSAV
  • 2023, 19(1), pp.69-76
  • DOI : 10.29056/jsav.2023.3.09
  • Publisher : Korea Software Assessment and Valuation Society
  • Research Area : Engineering > Computer Science
  • Received : March 10, 2023
  • Accepted : March 20, 2023
  • Published : March 31, 2023

Hyun-Soo Kim 1 Lee Ye-Seul 1 Shin DongMyung 2 Chan-jae Lee 3 Myung Ho Kim 4

1엘에스웨어
2엘에스웨어 (주)
3오케스트로
4숭실대학교

Accredited

ABSTRACT

In this paper, we studied how to apply slicing techniques to 4K and 8K images using high-resolution drone data, and used them to detect and de-identify personal information objects (car, people), and designed and implemented a system that detects and visualizes illegal crop objects in de-identified images. For privacy object detection and de-identification, single-stage techniques such as Yolov5 and Gaussian Blurring were applied, SwinTransformer, Soft-Teacher, and Fast-RCNN techniques were applied to detect illegal crop objects, and SAHI open source framework was used as image slicing techniques. The illegal crop object detection model used an ensemble Soft-Teacher model using SwinTransformer as a Backbone network and Fast-RCNN as a detector. Experiments by applying image slicing techniques to this model showed that the mAP was 0.663, which is improved from 0.456, which is the mAP of the model without applying the conventional image slicing techniques.

Citation status

* References for papers published after 2022 are currently being built.