无人机在复杂战场环境下, 因敌我双方无人机外形、颜色等特征较为相似, 如何准确地对敌方无人机识别是实现其自主导航及作战任务执行的关键。由于受敌方无人机飞行速度、形状、尺寸、姿态等的改变及气象环境因素的影响, 无法准确地对其进行识别与分类。针对这一问题, 提出基于迁移学习卷积稀疏自动编码器(Sparse Auto-Encoder, SAE)实现对航拍多帧图像中敌方目标对象的识别与分类。算法首先借助SAE对源领域数据集中大量无标记样本进行无监督学习, 获取其局部特征; 然后, 采用池化层卷积神经网络(CNN)算法提取目标图像全局特征; 最后, 送入Softmax回归模型实现目标对象的识别与分类。实验结果表明: 与传统非迁移学习的SAE算法及基于底层视觉特征学习的识别算法相比, 该算法具有更高的准确性。

UAV in complex battlefield environment, because the two sides of the UAV shape, color and other characteristics are more similar, how to identity enemy UAV accurately is the key to realize the autonomous navigation and combat mission execution. However, due to changes in the speed, shape, size, attitude of enemy UAV and the impact of meteorological and environmental factors, they can not be accurately identified and classified. Aiming at this problem, a kind of sparse auto-encoder(SAE) based on the transfer learning was proposed, and the target objects in the multi-frame aerial images were identified. The algorithm firstly used SAE to study the unsupervised learning of a large number of unmarked samples in the data concentration of source domain, and obtained its local characteristics. Then, the global feature response of the aerial image in the target domain was extracted by the convolution neural network (CNN) algorithm. Finally, the different categories of target objects were identified and classified by the Softmax regression model. The experimental results show that new algorithm proposed in this paper for multiple target objects in aerial multi-frame images is better than more traditional non-transfer learning SAE algorithm, and underlying visual feature recognition transfer learning algorithm, which has higher recognition rate.