YOLOv10-LGDA: An Improved Algorithm for Defect Detection in Citrus Fruits Across Diverse Backgrounds.

Abstract

Citrus diseases can lead to surface defects on citrus fruits, adversely affecting their quality. This study aims to accurately identify citrus defects against varying backgrounds by focusing on four types of diseases: citrus black spot, citrus canker, citrus greening, and citrus melanose. We propose an improved YOLOv10-based disease detection method that replaces the traditional convolutional layers in the Backbone network with LDConv to enhance feature extraction capabilities. Additionally, we introduce the GFPN module to strengthen multi-scale information interaction through cross-scale feature fusion, thereby improving detection accuracy for small-target diseases. The incorporation of the DAT mechanism is designed to achieve higher efficiency and accuracy in handling complex visual tasks. Furthermore, we integrate the AFPN module to enhance the model's detection capability for targets of varying scales. Lastly, we employ the Slide Loss function to adaptively adjust sample weights, focusing on hard-to-detect samples such as blurred features and subtle lesions in citrus disease images, effectively alleviating issues related to sample imbalance. The experimental results indicate that the enhanced model YOLOv10-LGDA achieves impressive performance metrics in citrus disease detection, with accuracy, recall, mAP@50, and mAP@50:95 rates of 98.7%, 95.9%, 97.7%, and 94%, respectively. These results represent improvements of 4.2%, 3.8%, 4.5%, and 2.4% compared to the original YOLOv10 model. Furthermore, when compared to various other object detection algorithms, YOLOv10-LGDA demonstrates superior recognition accuracy, facilitating precise identification of citrus diseases. This advancement provides substantial technical support for enhancing the quality of citrus fruit and ensuring the sustainable development of the industry.

Plant phenotyping relevance

柑橘果実の病斑・表面欠陥を画像から検出する改良YOLO手法の開発と性能評価が中心であり、植物器官の病害状態を推定する画像ベース表現型計測に該当する。

Code and data availability

The paper describes a citrus fruit disease image dataset (partly from Kaggle, partly self-collected) and an improved YOLOv10-LGDA model, but the supplied blocks contain no public deposit or availability statement for the dataset, images, code, or trained model. The Kaggle source is mentioned without any URL or access/`

No evidence-backed public reproduction asset is currently recorded.