Wheat straw target detection algorithm based on improved YOLOv5

Abstract

In agricultural production, the growth and yield of crops have always attracted people's attention. For the detection of wheat planting density, a wheat straw detection model based on improved YOLOv5 is proposed in this paper. Firstly, at the end of the backbone network, the C3 module (C3TR) integrated with Transformer is used to replace the traditional C3 module, so that the model can extract more feature information about wheat straw in the feature extraction stage; Secondly, after the improved C3 module is embedded the location attention module (Coordinate Attention, CA), by capturing the long-distance dependence on the space and the channel, makes the model more focused on the feature extraction of the target area, and further strengthens the feature extraction ability of the backbone network; Finally, for the traditional frame regression loss the function cannot solve the problem of returning gradients when the predicted frame and the real frame intersect. It is proposed to use CIoU instead of the traditional GIoU, and continue to guide the predicted frame while considering the Euclidean distance and aspect ratio of the center point of the predicted frame and the real frame. Moving closer to the ground-truth box, the loss function is further reduced. On the homemade wheat straw dataset, under the same training strategy, the experimental results show that! Compared with the traditional YOLOv5 model, the improved model has a 1.71% increase in mAP, which proves that the improved model is superior to the traditional YOLOv5 model in terms of accuracy, and has a better detection effect on small targets such as wheat straw some practicality.