A survey of few-shot object detection and its application in remote sensing image interpretation
The effective construction of deep learning object detection model usually relies on large-scale training sample sets,but the limited labeled samples in remote sensing image interpretation are often difficult to meet the training needs of neural networks.Few-shot learning(FSL)is an effective method for building a deep learning model under the condition that training samples are scarce.Few-shot object detection(FSOD)methods are widely used for natural images,but lacking for remote sensing images at present.The principle,advantages and disadvantages of FSOD method for natural image is summarized from four aspects:metric learning,meta-learning,transfer learning and data enhancement.Then,the research status of FSOD in remote sensing images is described,and three trends of FSOD in remote sensing images are proposed:the establishment of recognized evaluation indicators,three-dimensional object detection,and multiple feature fusion.
万方数据
维普
