Abstract: Aiming at the problem that the anti-interference ability of the traditional deep learning method and the popular target detection model for the passion fruit detection in real orchards are not ideal, this paper proposes a real-time detection of passion fruit in real orchards based on the improved YOLO-V3 network. Firstly, the prediction scale of large objects in the YOLO-V3 model is eliminated, and the 3-scale prediction is reduced to 2-scale prediction to speed up the detection time. Secondly, the DenseNet network is added to the medium object prediction scale to enhance the propagation of network features for improving the detection accuracy of the model. Finally, the improved YOLO-V3 network is used to train the passion fruit dataset several times to obtain the optimal pre-training model.The experimental results show that the improved YOLO-V3 network has a good real-time detection effect.The average detection accuracy of the target is more than 97.5% and the detection speed is 38 frames/s. The new method provides an effective theoretical basis for real-time detection of passion fruit.

Key words: deep learning, improved YOLO-V3 network, real-time detection, DenseNet network, passion fruit