Journal of the Korea Society of Computer and Information (한국컴퓨터정보학회논문지)

Korean Society of Computer Information (한국컴퓨터정보학회)
권호 표지
DOI QR코드

DOI QR Code

Effective Automatic Weed Detection With Improved YOLOv10

  • Hyeon-Jae Kwon (Dept. of Information & Telecommunication Engineering, Gangneung-Wonju National University) ;
  • Sangmin Suh (Dept. of Information & Telecommunication Engineering, Gangneung-Wonju National University, Scalable-AI)
  • Received : 2024.10.17
  • Accepted : 2024.11.08
  • Published : 2024.11.29
Download PDF
⟨ Previous Next ⟩

Abstract

In this paper, we design an improved weed detection model using YOLOv10, a deep learning-based object detection algorithm. YOLOv10 improves its performance compared to previous versions by adding an attention module, the PSA module. PSA is strong at recognising complex patterns in large areas because it uses some features of its own attention to reduce computation and learn global information. However, it may be inefficient for certain problems, such as weeds, which are generally small objects. Therefore, in this paper, we propose an improved YOLOv10 by applying another attention module, SENet, instead of the PSA module. Since, SENet learns the importance between channels, it can learn the features of weeds in more detail than the PSA module. In addition, SENet is lighter, less computationally intensive, and faster than the PSA module, so we conducted experiments by replacing the PSA module with SENet, which is suitable for weed detection. The experiment consisted of 200 training runs with a total of 14 classes, and we compared the performance through various performance evaluations. The experimental results showed that the FPS increased from 476.19 to 526.32, which is about 9.52% processing speed improvement. The mAP50-95 value increased from 88.7% to 88.3%, which shows that the proposed model is lighter than the existing model and performs similarly to the existing model.

이 논문은 딥러닝 기반 객체 탐지 알고리즘인 YOLOv10을 활용하여 개선된 잡초 탐지 모델을 설계한다. 기존 YOLOv10에서는 Attention 모듈인 PSA 모듈을 추가하여 이전 버전들보다 성능을 개선하였다. PSA는 Self-Attention의 강력한 성능을 일부만 적용하여 연산량을 줄이고, 전역적 정보를 학습할 수 있어 큰 영역의 객체가 복잡한 패턴 인식에 강하다. 하지만 대체로 작은 크기의 객체인 잡초 같은 특정 문제에서는 비효율적일 수 있다. 따라서, 이 논문은 PSA 모듈 대신 다른 Attention 모듈인 SENet을 적용하여 개선된 YOLOv10을 제안한다. SENet은 채널 간 중요도를 학습하기 때문에 PSA 모듈보다 더 세밀하게 잡초의 특징을 학습할 수 있다. 또한, SENet은 PSA 모듈보다 더 가벼워 더 적은 연산을 수행하고, 더 빠른 속도로 탐지가 가능하여 잡초 탐지에 적합한 SENet으로 대체하여 실험을 진행했다. 실험은 총 14가지의 클래스로 200회 훈련을 수행했고, 다양한 성능평가를 통해 성능을 비교하였다. 실험 결과에 따르면, FPS는 476.19에서 526.32로 처리속도가 약 9.52%정도 향상되었다. mAP50-95값은 88.7%에서 88.3%로, 제안된 모델이 기존 모델보다 더 경량화된 모델임에도 불구하고 유사한 성능을 보인다.

Keywords

Acknowledgement

This work was supported by Innovative Human Resource Development for Local Intellectualization program through the Institute of Information & Communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT)(IITP-2024-RS-2023-00260267)

References

  1. Campaigns surveys https://campaigns.do/surveys/329 
  2. Yoo Kwang Hyun, Vo-Hoang Trong, Kim Jin Young, "Deep Learning Research Trends For Weed Identification", Journal of The Korean Institute of Communication Sciences, No.8, 2021. 
  3. S. Suh, "Improved Motion Recognition Based On Convolution Neural Network For High Accuracy", journal of Knowledge Information Technology and Systems(JKITS), No. 3, 2021. 
  4. S. Suh, "Human Activity Recognition Systems Based On Maximum Color Difference And Deep Learning", Journal of Knowledge Information Technology and Systems(JKITS), No.4, 2021. 
  5. Gwanghyun Yu, Jaewon Lee, Vo Hoang Trong, Dang Thanh Vu, Huy Toan Nguyen, HooHwan Lee, Dosung Shin, Jinyoung Kim, "Alien Weed Classification: A Hierarchical Structure Based On Convolutional Neural Networks", Journal of Korean Institute of Information Technology, No. 12, Dec, 2019. 
  6. Harshita Shri Panati, Gopika P, Diana Andrushia A and Mary Neebha T, "Weeds And Crop Image Classificationo Using Deep Learning Technique", International Conference on Advanced Compution and Communication Systems(ICACCS), pp 117-122, DOI:10.1109/ICACCS57279.2023.1011958. 
  7. Y. Dandekar, K. Shinde, J. Gangan, S. Firdausi and S. Bharne, "Weed Plant Detection From Agricultural Filed Images Using YOLOv3 Algorithm", International Conference On Computing, Communication, Control And Automation(ICCUBEA), pp 1-4, DOI: 10.1109/ICCUBEA54992.2022.10011010. 
  8. Liu, W. et al. "SSD: Single Shot MultiBox Detector", ECCV Lecture notes in Computer Science, vol 99095. DOI: Https://doi.org/10.1007/978-3-319-46448-0_2. 
  9. J. Redmon, S.Divvala, R. Girshick and A. Farhadi, "YOU Only Look Once: Unified, Real-Time Object Detection", IEEE Conference on Computer Vision and Pattern Recognition(CVPR), pp. 779-788, DOI: 10.1109/CVPR.2016.91. 
  10. S. Tariyal, R. Chauhan, Y. Bijalwan, R. Rawat and R. Gupta, "A comparitive study of MTCNN, Viola-Jones, SSD and YOLO face detection algorithms," International Conference on Intelligent and Innovative Technologies in Computing, Electrical and Electronics (IITCEE), pp. 1-7, doi: 10.1109/IITCEE59897.2024.10467445. 
  11. YOLOv10 object detection Better, Faster and Smaller, visionplatform, https://visionplatform.ai 
  12. YOLOv10-docs https://docs.ultralytics.com/ko/models/yolov10/
  13. Dataset, https://www.aihub.or.kr/aihubdata/data/view.do?currMenu=115&topMenu=100&aihubDataSe=data&dataSetSn=527