A LIGHTWEIGHT FEW-SHOT LEARNING OBJECT DETECTION METHOD FOR UAVS’ ONBOARD COMPUTING PLATFORM

Deep learning often needs to be driven by big data, and has certain limitations in the fields where data is scarce, such as military and medical fields. Aiming at scarcity of label sample data in military field, this paper proposes a kind of few-shot object detection method which contains region proposal generate, pretraining model to extract features, and support vector machine (SVM) classification. In order to verify the effectiveness of the algorithm, this paper took UAV to collect the aerial image data of military model targets with practical reference value, and constructed a dataset. Based on this dataset, this paper carried out experiments on RK3399Pro, and achieved the best accuracy of 27.3% and recall rate of 62.4%, as well as considerable real-time performance. On this basis, localization constraints based on prior knowledge were added to further improve the performance of the method.