Abstract: As the city expands, the river shoreline is constantly being encroached upon, and the river's flow capacity and water environment are damaged severely. Therefore, efficient methods for monitoring complicated river courses are urgently needed. This paper precents a new method for collecting images and data of a river and its banks using drones and enhancing data with the Generative Adversarial Network. We identify and locate five kinds of typical foreign bodies on the river, based on the YOLOv5 algorithm and the coordinate transformation localization algorithm. The target recognition algorithm of this model introduces the attention mechanism into the backbone network and uses EIOU-Focal Loss as its loss function to improve YOLOv5 in detection accuracy and convergence speed. The results show that data enhancement improves the model’s target recognition and raises the mean Average Precision (mAP) by 9.9%. The ablation experiment results verify that this model has the highest detection accuracy, with its maximum mAP of 0.96 or an increase of 11.6% relative to the one before improvement. Its positioning results show the real average error of the algorithm target object is not greater than 3m, which means a high accuracy. Application of the improved model to the Minjiang River section in Fujian has verified its higher accuracy in detecting target objects and its significance for related research.

Key words: river detection, foreign body recognition, deep learning, YOLOv5, DCGAN