Ship Detection for Satellite Images based on Classifier Transfer Learning Combined with Feature Transfer Learning

Transfer learning (TL) is a powerful tool to transfer deep learning models from a large source dataset to a small target dataset, but the upper-layers of deep learning models are less transferable for lacking universality and possessing specificity to certain tasks. Most researches have focused on feature-oriented transfer learning base on the feature space, however, both the classifieroriented transfer learning and the label space haven’t been considered. Faced with these issues, a generalized classifier-oriented transfer learning, termed as classifier-TL, is proposed in this paper, which investigates the correlation between source label space and target label space to transfer and refine the generalized classifier. More specifically, for a given task, a label space descriptor is proposed to depict the label space, and a label space similarity is introduced to measure the correlation between source label space and target label space. Then, the target label space is focused through the proposed label driven posteriori optimization, trying to exploit similar label spaces of the closest category. In this procedure, the classifier can be refined from a set of generalized classifiers to a specific classifier. Furthermore, this classifier-TL can be combined with the traditional feature-oriented transfer learning, to form an integrative secondary transfer learning, for further boosting the performance of transfer learning. Experimental results for the task of ship detection, have demonstrated the effectiveness of our proposed method.