Fast identification of nematode via coordinate attention mechanism and efficient bounding box regression loss

Abstract: Plant nematode disease is one of the major diseases to threaten agricultural safety in China. The green, high-efficiency, and low-toxic nematicide pesticides can be one of the better means to control plant nematode disease, further to prevent the large-scale spread of nematode disease. However, the manual visual inspection cannot fully meet the large-scale screening and test of the nematicidal pesticide activity, such as time-consuming, low accuracy, and heavy workload. Particularly, it is very necessary to accurately and rapidly count the number of nematodes in the solution, and then to identify the dead and living insects. In this study, a fast identification was proposed for the nematodes using a coordinate attention mechanism and efficient bounding box regression loss YOLOFN (YOLO for nematodes). Firstly, the sawdust samples of tree trunks with the pine wood nematode were collected in the epidemic area of pine wood nematode disease. The solution slides of pine wood nematode were prepared with the different concentrations, after the pine wood nematode was separated in the laboratory. Secondly, the optical microscope and single lens reflex camera were used to collect the pine wood nematode images. The offline data enhancement was combined with the mosaic online data enhancement to expand the training samples of pine wood nematode images, according to the similar characteristics of pine wood nematode images. Thirdly, the feature extraction module of the coordinate attention mechanism was embedded in the backbone network, according to the theoretical framework of YOLOv5s target detection. The position information of the nematode feature map was then integrated into the channel attention, As such, this improvement enabled the model to focus on the target category and target location at the same time. Finally, a tradeoff was made on the overlapping ratio of the nematode target, the center point distance of the nematode target, the width and height of the prediction frame in the nematode target, as well as the proportion of positive and negative samples of the nematode target. The Efficient Intersection over Union (EIoU) and Focal loss functions was utilized to optimize the localization and classification loss function. There was a minimum difference between the width and height of the real frame and the predicted frame. The weight of the easily distinguishable samples was dynamically reduced to quickly focus the beneficial training samples. The analytical ability and regression accuracy of the model were improved to overlap the nematode targets. The experimental results showed that the performance indicators of YOLOFN were improved by 0.2, 4.4, and 3.8 percentage points, in terms of accuracy, recall, and mean Average Precision (mAP). The mAP of the improved model increased by 1.1, 31.7, and 15.1 percentage points, respectively, compared with the classical detection YOLOv3, SSD, and Faster R-CNN3. There was no difference in the inference time, where the mAP was higher than 11.0, 16.3, and 15.0 percentage points, respectively, compared with the lightweight backbone depth-separable convolution-YOLOv5, Mobilenetv2-YOLOv5, GhostNet-YOLOv5. Therefore, the YOLOFN model can be expected to quickly, accurately and efficiently realize the nematode microscopic examination and statistics, fully meeting the actual needs of research and development of plant nematode pesticides. The finding can provide strong technical support to accelerate the development of new drugs for plant nematode disease control. In the future research, the systematic deployment of the model can greatly contribute to the development of green nematicide, and the protection of plant nematode diseases. The high requirements of rapid end-to-side nematode counting and mortality measurement can be fully met using the expansion of the nematode species image data set.