Research on Feature Fusion Algorithms Based on the YOLO Model

Hu Li; Lei Dong; Siyuan Zhao

doi:10.54097/k6g1d334

Authors

Hu Li
Lei Dong
Siyuan Zhao

DOI:

https://doi.org/10.54097/k6g1d334

Keywords:

Poppy Identification, YOLOv9, HSV Color Filter, Small Object Detection, Feature Fusion.

Abstract

As a crop of both medicinal value and illicit abuse potential, the rapid and accurate identification of illegally cultivated poppies is critical for global counter-narcotics operations. However, in complex natural environments, poppy plants exhibit high morphological similarity to other vegetation, resulting in substantial false positives and false negatives in existing detection models. To address this, we propose a feature fusion algorithm that integrates YOLOv9 with HSV color space filtering. The method employs empirically defined HSV thresholds to pre-screen poppy-specific color features, generating color attention masks. These masks are channel-aligned and fused pointwise with multi-scale deep features extracted by YOLOv9, achieving complementary enhancement between low-level color priors and high-level shape semantics. Evaluated on a custom drone aerial poppy dataset, the proposed model reduces the false positive rate from 18.5% to 3.2% while maintaining high detection efficiency. This work presents an effective feature fusion strategy to mitigate misclassification in small-target aerial detection within complex scenes, offering practical value for enhancing the automation and precision of anti-drug surveillance.

References

[1] Rominger K, Meyer S E. Application of UAV-based methodology for census of an endangered plant species in a fragile habitat[J]. Remote Sensing, 2019, 11(6): 719.

[2] Zhang X, Luo D, Zhou L. Opium Poppy Image Detection Based on Improved YOLOV5[C]//2024 International Symposium on Digital Home (ISDH). IEEE, 2024: 127-132.

[3] Wang Q, Wang C, Wu H, et al. A two-stage low-altitude remote sensing papaver somniferum image detection system based on YOLOv5s+ DenseNet121[J]. Remote Sensing, 2022, 14(8): 1834.

[4] Radovic M, Adarkwa O, Wang Q. Object recognition in aerial images using convolutional neural networks[J]. Journal of Imaging, 2017, 3(2): 21.

[5] Zhang Z, Xia W, Xie G, et al. Fast opium poppy detection in unmanned aerial vehicle (UAV) imagery based on deep neural network[J]. Drones, 2023, 7(9): 559.

[6] Mieites V, Gutiérrez-Gutiérrez J A et al. Single-Image Multi-Parametric Representation of Optical Properties through Encodings to the HSV Color Space[J]. Applied Sciences, 2023, 14(1): 155.

[7] Zhang Z, Xia W, Xie G, et al. Fast opium poppy detection in Unmanned Aerial Vehicle (UAV) imagery based on deep neural network[J]. Drones, 2023, 7(9): 559.

[8] Lahoti A, Karp S, Winston E, et al. Role of Locality and Weight Sharing in Image-Based Tasks: A Sample Complexity Separation between CNNs, LCNs, and FCNs[J]. arXiv preprint arXiv:2403.15707, 2024.

[9] Rahman A, He L, Wang H. Activation function optimization scheme for image classification[J]. Knowledge-Based Systems, 2024, 305: 112502.

[10] Galanis N I, Vafiadis P, Mirzaev K G, et al. Convolutional neural networks: A roundup and benchmark of their pooling layer variants[J]. Algorithms, 2022, 15(11): 391.

[11] Babaiee Z, Kiasari P M, Rus D, et al. Unveiling the unseen: Identifiable clusters in trained depthwise convolutional kernels[J]. arXiv preprint arXiv:2401.14469, 2024.

[12] Pham M T, Courtrai L, Friguet C, et al. YOLO-Fine: One-stage detector of small objects under various backgrounds in remote sensing images[J]. Remote Sensing, 2020, 12(15): 2501.

[13] Hu J, Shi C J R, Zhang J. Saliency-based YOLO for single target detection[J]. Knowledge and Information Systems, 2021, 63(3): 717-732.

[14] Lei L, Fang Z, Ren J, et al. Two-click based Fast Small Object Annotation in Remote Sensing Images[J]. IEEE Transactions on Geoscience and Remote Sensing, 2024.

[15] Dendorfer P, Ošep A, Milan A, et al. Motchallenge: A benchmark for single-camera multiple target tracking[J]. arXiv preprint arXiv:2010.07548, 2020.