A deep learning approach incorporating YOLO v5 and attention mechanisms for field real-time detection of the invasive weed Solanum rostratum Dunal seedlings

Abstract

Solanum rostratum Dunal is a common invasive alien weed that can damage native ecosystems and biodiversity. Detecting Solanum rostratum Dunal at an early stage of growth will make it possible to treat it before it causes serious damage. Therefore, a convolution neural network model YOLO-CBAM is constructed in this paper for the detection of the Solanum rostratum Dunal seedings, which is incorporating YOLO v5 and attention mechanism. A method is designed for slicing the high-resolution images by calculating the overlap rate to construct datasets that reduce the possibility of detail loss due to compressing high-resolution images during the training process. Multiscale training methods have been used to improve training performance. The comparison tests show that the Precision and Recall of the proposed YOLO_CBAM are both higher than that of YOLO v5. The performance of the network is further improved after multi-scale training, and the Average Precision (AP) of YOLO_CBAM increased from 0.9017 to 0.9272. The trained network model was deployed to Jetson AGX Xavier for field trials. The network model achieved a Precision of 0.9465 and a Recall of 0.9017 for real-time recognition. The detection speed and the detection effectiveness can be applied to field real-time detection of the invasive weed Solanum rostratum Dunal seedlings.