Abstract:  In order to address the problem that traditional deep learning algorithms are difficult to identify rice diseases accurately and efficiently in complex environments, an improved YOLOv5 algorithm was proposed to detect the disease spots of common rice blight, rice blast, tungro disease and brown spot disease. A hybrid domain attention mechanism was combined with the original YOLOv5 algorithm for feature correction, which improved the model's ability to determine the position of rice leaves and disease spot location. The original CIoU_loss (Complete Intersection over Union) was replaced by SIoU_loss (SCYLLA Intersection over Union) in the loss function part to compensate for the problem that CIoU_loss did not focus on the angular offset of the bounding box and the Ground truth box . Soft-NMS (Soft Non Maximum Suppression) was chosen to filter the prediction boxes to alleviate the leakage caused by the overlapping area of different lesions in the conventional NMS. In the ablation experiment, the mAP (mean Average Presicion) of the improved algorithm reached 0.884 in the rice disease identification task, which was 2.9 percentage points higher than the original YOLOv5 algorithm, and the improvement was greater in the identification of brown spot. It proved the effectiveness of the improved YOLOv5 algorithm in the rice disease identification task.

Key words: rice disease, YOLOv5 algorithm, attention mechanism, object detection

摘要： 针对传统深度学习算法难以在复杂环境下准确且高效地识别水稻病害问题，提出一种改进的YOLOv5算法，对水稻常见的白叶枯病、稻瘟病、东格鲁病和褐斑病的病斑进行检测。在原YOLOv5算法上结合混合域注意力机制进行特征校正，提高模型对水稻叶片和病斑位置信息的定位能力。在损失函数部分将原CIoU_loss更换为SIoU_loss，弥补CIoU_loss未关注边界框和真实框角度偏移的问题。选用Soft-NMS筛选预测框，缓和传统NMS因不同病斑重叠区域过大而发生预测框误删造成的漏检情况。在消融试验中，改进算法在水稻病害识别任务中mAP达到0.884，比原YOLOv5算法提升2.9个百分点，在针对褐斑病病斑的识别上提升较大。证明改进的YOLOv5算法在水稻病害识别任务中的有效性。

关键词: 水稻病害, YOLOv5算法, 注意力机制, 目标检测