Lightweight architecture optimization of YOLOv12n for improved cotton verticillium wilt detection.

Abstract

To address the significant morphological variability of cotton Verticillium wilt lesions and the complex background interference present in field environments, existing detection models often struggle to achieve an effective balance between detection accuracy and model complexity. In this study, a precise and lightweight detection model, YOLO-SCOD, is proposed based on the YOLOv12n framework to enhance lesion recognition performance. During the feature extraction stage, YOLO-SCOD adopts the StarNet architecture as the backbone network. Its efficient feature mapping mechanism enhances the ability to interact with multi-scale features, thereby optimizing the model's capacity to represent multi-scale lesion information. Meanwhile, a channel aggregation block is integrated into the C3k module of the neck network. Through adaptive channel reallocation and enhancement of key lesion features, the perception capability of the C3k2 and A2C2f modules for discriminative lesion features is improved. In the detection head, depthwise convolution is replaced with omni-dimensional dynamic convolution, which dynamically and adaptively adjusts convolutional weights through multi-dimensional attention collaboration, further improving the model's localization accuracy and recognition capability. Experimental results demonstrate that the YOLO-SCOD model achieves improved performance improvements in the task of cotton Verticillium wilt detection. Compared with YOLOv12n, its precision and recall increase to 0.960 and 0.911, respectively, while mAP50-95 improves by 6.436%. In addition, the number of model parameters, FLOPs, and model size are reduced by 13.728%, 20.635%, and 12.727%, respectively, and inference speed increased by 4.167%. While maintaining high detection accuracy, YOLO-SCOD exhibits favorable lightweight characteristics, providing a viable solution for efficient automatic identification and intelligent detection of cotton Verticillium wilt.

Plant phenotyping relevance

綿花の萎凋病病斑という植物の病害状態を画像検出するモデルを開発し、精度・計算量・推論速度を比較検証しており、病害表現型の取得手法が中心的です。

Code and data availability

The paper's cotton Verticillium wilt image dataset (471 raw images, 2355 augmented images, 5785 annotations) and YOLO-SCOD model/code are not publicly deposited. The Data Availability Statement says raw data will be made available by the authors upon request; no public URL or repository is provided anywhere in the text

No evidence-backed public reproduction asset is currently recorded.