Today, drones are used in various fields such as agriculture, military, construction and transportation. Due to the increasing expansion of drones, object detection is one of the subfields of computer vision, which is very widely used to make them intelligent. Considering the constraints of drones, single-board computers are the best processing hardware for implementing and using artificial intelligence networks in drones. The YOLOv5n network is the best choice for object detection with a speed of at least 30 frames per second using a single-board Nvidia Jetson Nano computer. However, this network exhibits a considerable decrease in accuracy in hazy images compared to non-hazy images. The primary aim of this research is to enhance the accuracy of YOLOv5n for detecting objects in hazy images collected by small drones. In the proposed method, a network for de-hazing is added to the YOLOv5n network and removed at the end of the three-step training process. In the first stage of training, the backbone and haze remover network, in the second stage, the neck and head, and in the third stage, the entire network has been trained with the data that haze added to them in three different ways. The eva‎luation criterion in this research is the priority of inference speed and accuracy. Other related methods are slower than the proposed method of this research. In this research, the accuracy of YOLOv5n in hazy aerial images has been improved by about 13%, without reducing the inference speed.

محمدعلي خسروي فرد

رسول امير فتاحي ورنوسفادراني , محمدرضا احمدزاده