Abstract:
Automatic detection of wheat spike is of great scientific value in wheat yield estimation and breeding, however there are still problems in current wheat spike detection such as high model complexity and low accuracy.Deep learning technology was applied to wheat spike detection, and an improved YOLOv5 model for accurate detection of wheat spikes was proposed.The model replaced convolution module in the YOLOv5 backbone network with Ghost convolution to achieve a lightweight model; used ACON activation function to replace default SiLU activation function, thus making activation function more flexible to expand design space; used the Alpha-IoU Loss, which increased alpha power for all IoU Losses, to replace default CIoU Loss of YOLOv5.Finally, by adding a weighted bi-directional feature pyramid(BiFPN)to the network, improved model achieved a 63.3% reduction in number of parameters and a 66.8% reduction in computation with only a 2.17% reduction in
mAP, met requirements of practical applications and mobile deployments.Use of Decouple Head was proposed to replace default YOLO Head, which improved
mAP by 1.83% over official YOLOv5, demonstrated that decoupling head ccould improve model accuracy.