Transfer Learning With Fully Pretrained Deep Convolution Networks for Land-Use Classification

摘要: In recent years, transfer learning with pretrained convolutional networks (CNets) has been successfully applied to land-use classification high spatial resolution (HSR) imagery. The commonly used CNets partially use the feature descriptor part of pretained CNets, and replace classifier in old task a new one. This causes separation asynchrony between transferred during process, which reduces effectiveness training process. To overcome this weakness, method fully is proposed letter for HSR images. method, multilayer perceptron (MLP) quickly using high-level features extracted by CNets. Fully can be generated concatenating MLP. Because both are pretrained, two parts avoided final then obtained fine-tuning random cropping mirroring strategy. experiments show that accelerate convergence process no loss accuracy classification, its performance comparable other latest methods.