关键词: 机载视觉感知, 鲁棒性设计, 深度集成学习, K折交叉验证方法, 深度可分离卷积

Abstract: The visual perception function based on machine learning is crucial for enhancing situational awareness or autonomous flight capa-bilities of aircraft in complex environments, with its performance significantly impacting flight safety. However, the inherent proba-bilistic nature of machine learning techniques poses significant challenges to meeting airworthiness safety objectives, hindering their application in airborne systems. Therefore, this paper proposes a robust optimization design for airborne visual perception based on deep ensemble learning.First, a highly representative dataset is generated based on the operational design domain, and a K-fold cross-validation method based on CW-SSIM is proposed to improve the independence between the training and validation sets with limited data. Second, based on the YOLO architecture, depthwise separable convolution is introduced, and three optimized base learners are designed to address different detection needs through multi-scale feature fusion, enhanced focus on small object detec-tion, and fine-grained feature extraction. Finally, an ensemble learning method is designed, employing a weighted adaptive fusion strategy to dynamically adjust the weights of base learners, improving the model's accuracy and robustness.Experimental results show that the ensemble learning model outperforms detection box fusion algorithms such as NMS and WBF. When the IoU is not less than 0.7, the ensemble model improves the average P-value, R-value, and F1 score by at least 11.36%, 2.06%, and 6.78%, re-spectively, compared to a single model. When the IoU is not less than 0.75, the AP value increases by at least approximately 3%. The proposed method significantly enhances target detection accuracy and robustness in complex environments, effectively reducing false positives and missed detections, providing technical assurance for the safe flight of aircraft.

Key words: airborne visual perception, safety design, deep ensemble learning, k-fold cross-validation method, depthwise separable convo-lutions