Few-shot Aircraft Detection Based on Deep Reinforcement Learning

Data dependence is one of the challenging problems in aircraft object detection tasks in the field of remote sensing. Compared with traditional machine learning methods, deep learning relies heavily on large amounts of training data to understand potential data patterns. Specifically, due to the few number of samples, the generalization performance of the model is limited. To deal with this issue, this paper combines deep learning with perceptive ability and reinforcement learning with decision-making ability to achieve an effective object detection method. Firstly, we adopt an improved LSTM network to generate the strategy of data augmentation; Secondly, the strategy is applied to train and test of the object detection network to get F1-score as reward; Then, the policy gradient in reinforcement learning is employed to update the LSTM network, so that the network can generate profitable augmentation strategies that are more conducive to improving the accuracy of object detection. Finally, these processes are looped until the network reaches convergence. Compared with the existing object detection algorithm based on deep learning, the performance of our deep reinforcement learning (DRL) method is superior to the prevalent algorithms in aircraft detection on the public data sets DIOR and RSOD.