In view of the loss of agricultural production caused by bird pecking in recent years, a target detection algorithm based on YOLOv3 bird detection was proposed to optimize the switching strategy of traditional bird repellent machine by identifying birds accurately in real time. This method improves the feature fusion in YOLOv3 network, and brings SE module embedded in the Darknet53 network of the trunk network, focusing on the importance of different channel characteristics. Using adaptive spatial feature fusion (ASFF) enhanced network feature pyramid network (FPN) feature fusion to enhance the detection capability of all scales. The loss function of CIOU boundary box regression is used to take into account the overlap or even inclusion of the prediction-box and the target-box, so that the target-box regression becomes more accurate and stable. The improved YOLOv3 model was used on the self-made bird data set. The average precision (AP) reached 96.65%, and the detection time of single image was only 0.058 s, which increased by the 2.54 percentage point compared with the original YOLOv3 model under the condition of little change in detection speed. The improved method can achieve good real-time performance and better detection accuracy, and provide a basis for optimizing the switching strategy of bird repellent for farmland bird prevention and control.