关键词: 深度学习, 目标检测, YOLOv5, 黄瓜霜霉病, 孢子囊

Abstract: Aiming at the characteristics of dense sporangia distribution, cohesive stacking and complex background in the images collected by the greenhouse spore capture equipment, an improved YOLOv5s-based sporangia detection algorithm for cucumber downy mildew was proposed. Firstly, a Ghost convolution with CBAM (Convolutional Block Attention Module) attention mechanism was used to replace the CSP (Cross Stage Partial) module in the original network, which suppressed impurities in the background and ensured the generation of rich feature maps while reducing the number of model parameters and improving computational speed. Secondly, the connection method of the feature fusion network was modified, the original branch responsible for large object detection was deleted and a finergrained branch was added to strengthen the detection of small objects and dense, stacked objects. Finally, different weights were assigned to the loss values generated by different prediction heads, and the original NMS method was replaced by the DIOU_NMS nonmaximum value suppression method that considered the center point distance. The average accuracy and FPS of the improved YOLOv5s algorithm were 91.18% and 65.4 frames/s respectively, which were 4.88% and 7.1 frames/s higher than the original YOLOv5s algorithm. This study can provide data support for monitoring the occurrence and development of cucumber downy mildew, and is of great significance for ensuring the yield and quality of cucumbers.

Key words: deep learning, object detection, YOLOv5, cucumber downy mildew, sporangia