Few-shot object detection model based on meta-learning for UAV
junli liu · Fifth International Conference on Mechatronics and Computer Technology Engineering (MCTE 2022) · 2022
In image-based object detection, traditional deep learning algorithms have made some achievements. However, the algorithm has a large amount of interaction with the environment and high sample complexity when solving tasks, so the convergence of the algorithm is difficult. At present, it is difficult to collect UAV aerial photography target data. To solve this problem, a few-shot object detection model based on meta-learning is proposed by referring to the idea of meta-learning. Based on the Faster R-CNN algorithm, the MAML algorithm of the meta-learning is introduced, and its RPN structure is improved. The proposed model includes training phase and testing phase. In the training stage, the object detection network is trained with the training dataset formed by the combination of public data sets, so that the network can learn the meta-knowledge of object detection. In the test phase, the model is fine-tuned with the test dataset to enable the network to detect new classes of targets. The results show that the effect of the model based on the few-shot meta-learning algorithm is better than that of the classical algorithm such as Faster R-CNN, which can realize the accuracy of the UAV in the case of few-shot data training and verify the effectiveness and advanced nature of the model.