When identifying rice pests, issues such as targets being obscured, similarity to the background color, and proximity of multiple targets due to the rice field environment can lead to reduced identification accuracy. To address this, a rice pest identification method was proposed based on an improved YOLO v5s. The method enhanced the model's ability to capture target location information by replacing ordinary convolution in the backbone network with CoordConv. It introduced the CBAM attention mechanism to increase the model's focus on the target area. The Slim — neck architecture was adopted to enhance feature processing capabilities and reduce computational load. The introduction of the Soft — NMS algorithm optimized the selection of adjacent target bounding boxes, reducing missed detections. Experimental results showed that the improved YOLO v5s model achieved an mAP of 94. 3% on the rice pest dataset, which was an increase of 3. 8 percentage points over the original model and 1.5, 12. 7, 13. 6 and 1. 9 percentage points higher than that of the other mainstream models such as YOLO v3, YOLO R —CSP, YOLO v7, and YOLO v8s, respectively. Ablation experiments further validated the effectiveness of each component in the improved model. Heat map analysis also demonstrated that the improved model can better learn pest features. In summary, the improved YOLO v5s model proposed achieved significant results in improving the accuracy of rice pest detection, providing a more precise identification method for the prevention and control of rice pests. © 2024 Chinese Society of Agricultural Machinery. All rights reserved.
