Enhanced Military Aircraft Detection Using YOLOv5s with Hyperparameter Evolution in Remote Sensing Images

Aircraft detection in aerial imaging plays an important role in various applications, such as surveillance, reconnaissance, and disaster response. Unmanned aerial vehicles and satellite imagery are valuable resources for remote sensing applications. The detection of aircraft in these images is challenging because of the diverse aircraft types, varying image scales, and complex backgrounds. In this study, a modified YOLOv5s that incorporates the Adam optimizer instead of standard Stochastic Gradient Descent (SGD) is proposed. Furthermore, the impact of genetic evolution on both the base and optimized models was studied. The best hyperparameters obtained after 100 generations were used to train YOLOv5s. Optimizing hyperparameters through genetic evolution and switching the optimizer from SGD to Adam increased the mAP score from the base YOLOv5s model. A mean average precision of 96.57% was achieved for the Adam optimizer, which outperformed the base YOLOv5s model by a margin of 1.8%. The proposed method outperformed YOLOv5m and had a similar mAP score to YOLOv51, making it a feasible option for applications that require high accuracy, low computational cost, and fast inference times. Genetic evolution increased the mAP score of the model using the SGD optimizer but did not improve the results of the model using the Adam optimizer.