Authors

Abstract

This paper proposes a lightweight spore detection method for millet downy mildew based on an improved YOLOv8s, aiming to enhance the accuracy and efficiency of spore detection. First, the backbone network of the YOLOv8s model was modified by replacing the original backbone with EfficientViT. The substitution of the EfficientViT backbone enables global receptive field and multi-scale learning, which helps to reduce computational costs. While maintaining high performance, this modification significantly improves computational efficiency. Second, a Frequency-Adaptive Dilated Convolution (FADC) module was added to the neck of the model. By adaptively adjusting the receptive field of dilated convolution, the FADC module optimizes the detection of different frequency information. It improves the detection of small objects without adding extra computational burden. Finally, the detection head was optimized to better adapt to the task of detecting millet downy mildew spores, resulting in enhanced detection speed and accuracy. The improved algorithm, named EFP-YOLOv8s, maintains the same mAP50 as the original YOLOv8s model while reducing the number of parameters by 37.8% and computational cost by 58.5%. By balancing high performance with reduced computational resource demands, the model achieves lightweight design, making it more deployable and scalable in practical applications.

Abstract in Chinese

谷子白发病是一种严重影响谷子产量的病害，传统的白发病监测方法通常依赖人工识别，容易出现漏检和误判的情况。本文提出了一种基于改进YOLOv8s的轻量化谷子白发病孢子检测方法，旨在提高孢子检测的准确性和效率。首先，针对YOLOv8s模型的主干网络进行了改进，将原有的主干网络替换为EfficientViT。通过替换主干网络EfficientViT实现全局感受野和多尺度学习，有助于降低计算成本。在保持高性能的同时，显著提高了计算效率。其次，在模型的颈部添加了FADC（Frequency-Adaptive Dilated Convolution）模块，通过自适应地调整膨胀卷积的感受野，针对不同频率的信息进行优化。在不增加额外计算量的同时改善小目标的检测效果。最后，对检测头进行了优化，使其更加适应谷子白发病孢子的检测任务，提高了检测速度和准确性。改进后的算法EFP-YOLOv8s相比原YOLOv8s模型mAP50保持不变，达到97.6%，参数量下降37.8%，计算量下降58.5%。该模型在保持高性能的同时，也降低了计算资源的需求，实现轻量化，更易于在实际应用中部署和推广。