YOLOv5 Spore Detection Algorithm for Wheat Powdery Mildew Based on Attentional Feature Fusion

Chuangang Chong,Botao Wang,Haozhe Yuan
DOI: https://doi.org/10.1117/12.2686372
2023-01-01
Abstract:Wheat powdery mildew spore detection is one of the most common scenarios in the field of agricultural pest and disease detection. Since wheat powdery mildew spore images possess the characteristics of small spore targets, easy adhesion, shallow feature dimension, and many background clutter, it leads to the increased difficulty of wheat powdery mildew spore detection. By analyzing the above detection difficulties, a YOLOv5 wheat powdery mildew spore detection algorithm based on attentional feature fusion is proposed in this paper. Improvements are made on the basis of YOLOv5s. Firstly, an attentional mechanism-based feature fusion module AM-FF (Attentional Mechanisms-Feature Fusion) is proposed to strengthen the fusion of features at different scales, which enhances the learning ability of network contextual association by autonomously learning and selecting the optimal features through the attentional mechanism. Secondly, in order to enhance the shallow features and small target detection accuracy, we modify the original network structure to enhance the acquisition of shallow features and add fused reinforced shallow features in the neck network FPN (Feature Pyramid Networks). Finally, the SIoU (SCYLLA-IoU) loss function is introduced in order to increase the consideration of vector angle in the boundary box regression calculation, which effectively improves the detection accuracy. Experiments were conducted on the expert-labeled wheat powdery mildew spore dataset, and the final result of the algorithm experiment AP (Average Precision) was 94.23%, which was a large improvement compared with the original YOLOv5 detection algorithm vertically and also compared with the algorithm made by previous authors horizontally, and the experimental results proved the superiority of the proposed method in this paper.
What problem does this paper attempt to address?