Lightweight fungal spore detection based on improved YOLOv5 in natural scenes

Plant disease diagnosis plays a crucial role in precision agriculture, and detecting fungal spores in complex scenes is essential for disease diagnosis. The spores of downy mildew obtained in the natural scene carry a lot of impurities such as dust, and bacteria, and the small size of downy mildew spores themselves also pose challenges for accurate detection. At the same time, it is necessary to improve the real-time detection of a large number of images of downy mildew spores collected by the microscope. This study addresses the challenges of low accuracy and complex detection algorithms in fungal spore detection under complex backgrounds. We propose a high-precision and lightweight detection method based on YOLOv5. Our approach incorporates the Ghost module and its optimization module GhostC3 to reduce model complexity and improve detection efficiency. Additionally, the Normalization-based Attention module (NAM) is embedded in the backbone network to enhance the model's ability to extract fungal spore features at different depths, mitigating the negative impact of complex backgrounds. Meanwhile, we introduce the Receptive Field Block with dilated convolution (RFB-s) to improve the detection accuracy of tiny target spores by increasing the receptive field and extracting detailed information on fungal spore location. Experimental evaluations using spores of downy mildew fungi collected in natural scenes demonstrate that our proposed method achieves an Average Precision (AP) of 0.956, with a model size of 2.8 M. Compared to the baseline model, our approach reduces parameters by 34.9% while improving AP by 1.4%. Notably, our model outperforms mainstream detection algorithms such as YOLOv3, SSD, YOLOv4, YOLOv5, and YOLOv7 regarding AP. This method offers significant technical support for automating the detection of downy mildew spores and early prediction of downy mildew outbreaks in complex field environments.