Detection of Rice Pests Based on Self-Attention Mechanism and Multi-Scale Feature Fusion

• Published in: Insects (Basel, Switzerland) Vol. 14; no. 3; p. 280
• Main Authors: Hu, Yuqi, Deng, Xiaoling, Lan, Yubin, Chen, Xin, Long, Yongbing, Liu, Cunjia
• Format: Journal Article
• Language: English
• Published: Switzerland MDPI AG 13.03.2023; MDPI
• Subjects: Accuracy; Agricultural pests; Agriculture; Algorithms; Artificial intelligence; Artificial neural networks; BiFPN; Crop production; Crop yield; Datasets; Deep learning; Digital imaging; Diseases and pests; Growth; Identification; Identification and classification; Image classification; Insects; Methods; Neural networks; Object recognition; Optimization; Pests; Rice; rice pest detection; self-attention; Swin Transformer; YOLOv5; China; self-attention; YOLOv5; BiFPN; Swin Transformer; rice pest detection
• ISSN: 2075-4450; 2075-4450
• DOI: 10.3390/insects14030280

In recent years, the occurrence of rice pests has been increasing, which has greatly affected the yield of rice in many parts of the world. The prevention and cure of rice pests is urgent. Aiming at the problems of the small appearance difference and large size change of various pests, a deep neural network named YOLO-GBS is proposed in this paper for detecting and classifying pests from digital images. Based on YOLOv5s, one more detection head is added to expand the detection scale range, the global context (GC) attention mechanism is integrated to find targets in complex backgrounds, PANet is replaced by BiFPN network to improve the feature fusion effect, and Swin Transformer is introduced to take full advantage of the self-attention mechanism of global contextual information. Results from experiments on our insect dataset containing Crambidae, Noctuidae, Ephydridae, and Delphacidae showed that the average mAP of the proposed model is up to 79.8%, which is 5.4% higher than that of YOLOv5s, and the detection effect of various complex scenes is significantly improved. In addition, the paper analyzes and discusses the generalization ability of YOLO-GBS model on a larger-scale pest data set. This research provides a more accurate and efficient intelligent detection method for rice pests and others crop pests.