【作者】 陈诚；

【导师】 王卫亚；

【作者基本信息】 长安大学 ， 计算机技术（专业学位）， 2021， 硕士

【摘要】 随着现代社会的进步,航空运输领域也迎来了飞速发展,对于使用计算机技术来加强机场的安全措施越来越重视。目前鸟撞是威胁航空安全的重要因素之一。如果可以进一步提高飞鸟目标识别的准确性,就能及时采取相应的驱鸟措施,从而有效降低鸟撞带来的危害性。但是在目标识别领域针对机场驱鸟应用的算法较少,识别效果欠佳。本文对基于深度学习的机场飞鸟目标识别模型展开研究。对卷积神经网络模型中的双阶段和单阶段算法进行对比与分析,选取对机场飞鸟目标识别适用性更好的Faster R-CNN目标识别算法进行研究与改进。采用基于密集特征融合的金字塔网络解决了卷积神经网络的加深导致特征减少的问题,提高了小目标飞鸟的识别精度;采用基于K-Means聚类算法对锚框生成方法进行改进,增强了锚框对于飞鸟目标的适用性;采用双线性插值算法对感兴趣区域池化进行改进,提升了鸟群中飞鸟目标的识别精度;采用改进的非极大值抑制算法有效解决了飞鸟目标之间出现重叠的问题。为了验证本文提出的四种改进方法之间不会相互冲突,同时直观地展现这几种方法分别有多大的贡献值,将其集成到Faster R-CNN基础算法上进行综合消融实验。实验结果表明,改进后的模型对于小目标飞鸟及鸟群的识别精度显著提高,并有效解决了飞鸟目标之间重叠导致的漏识别问题。更多还原

【Abstract】 With the progress of modern society,the field of air transportation has also ushered in rapid development,and more and more attention is paid to the use of computer technology to strengthen airport security measures.At present,bird strike is one of the important factors threatening aviation safety.If the accuracy of bird target recognition can be further improved,corresponding bird repellent measures can be taken in time,thereby effectively reducing the harm caused by bird strikes.However,in the field of target recognition,there are fewer algorithms for airport bird repelling,and the recognition effect is not good.In this paper,the target recognition model of airport birds based on deep learning is studied.The two-stage and single-stage algorithms in the convolutional neural network model are compared and analyzed,and the Faster R-CNN target recognition algorithm,which is more applicable to airport bird target recognition,is selected for research and improvement.The use of a pyramid network based on dense feature fusion solves the problem of reduced features caused by the deepening of convolutional neural networks,and improves the recognition accuracy of small target birds.The K-Means clustering algorithm is used to improve the anchor frame generation method,which enhances the applicability of the anchor frame to the bird target.The bilinear interpolation algorithm is used to improve the pooling of the region of interest,and the recognition accuracy of the bird target in the bird flock is improved.An improved non-maximum suppression algorithm is used to effectively solve the overlap problem among birds.In order to verify that the four improved methods proposed in this paper will not conflict with each other,and at the same time visually show how much contribution these methods have,they are integrated into the Faster R-CNN basic algorithm for comprehensive ablation experiments.The experimental results show that the improved model has significantly improved the recognition accuracy of small targets flying birds and bird groups,and effectively solves the problem of missing recognition caused by overlapping bird targets.更多还原