Tremendous progress has been made in object recognition with deep convolutional neural networks (CNNs), thanks to the availability of large-scale annotated dataset. With the ability of learning highly hierarchical image feature extractors, deep CNNs are also expected to solve the Synthetic Aperture Radar (SAR) target classification problems. However, the limited labeled SAR target data becomes a handicap to train a deep CNN. To solve this problem, we propose a transfer learning based method, making knowledge learned from sufficient unlabeled SAR scene images transferrable to labeled SAR target data. We design an assembled CNN architecture consisting of a classification pathway and a reconstruction pathway, together with a feedback bypass additionally. Instead of training a deep network with limited dataset from scratch, a large number of unlabeled SAR scene images are used to train the reconstruction pathway with stacked convolutional auto-encoders (SCAE) at first. Then, these pre-trained convolutional layers are reused to transfer knowledge to SAR target classification tasks, with feedback bypass introducing the reconstruction loss simultaneously. The experimental results demonstrate that transfer learning leads to a better performance in the case of scarce labeled training data and the additional feedback bypass with reconstruction loss helps to boost the capability of classification pathway.

Deep convolutional neural networks (DCNNs) have attracted much attention in remote sensing recently. Compared with the large-scale annotated data set in natural images, the lack of labeled data in remote sensing becomes an obstacle to train a deep network very well, especially in synthetic aperture radar (SAR) image interpretation. Transfer learning provides an effective way to solve this problem by borrowing knowledge from the source task to the target task. In optical remote sensing application, a prevalent mechanism is to fine-tune on an existing model pretrained with a large-scale natural image data set, such as ImageNet. However, this scheme does not achieve satisfactory performance for SAR applications because of the prominent discrepancy between SAR and optical images. In this article, we attempt to discuss three issues that are seldom studied before in detail: 1) what network and source tasks are better to transfer to SAR targets; 2) in which layer are transferred features more generic to SAR targets; and 3) how to transfer effectively to SAR targets recognition. Based on the analysis, a transitive transfer method via multisource data with domain adaptation is proposed in this article to decrease the discrepancy between the source data and SAR targets. Several experiments are conducted on OpenSARShip. The results indicate that the universal conclusions about transfer learning in natural images cannot be completely applied to SAR targets, and the analysis of what and where to transfer in SAR target recognition is helpful to decide how to transfer more effectively.

Convolutional neural network (CNN)-based methods have been successfully applied to SAR target detection. Different from prevalently used detection approaches with rectangle bounding box, rotatable bounding box (RBox)-based methods, such as DRBox-v1, can effectively reduce the interference of background pixels and locate the targets more finely for geospatial object detection. Although DRBox-v1 has achieved impressive detected performance, there still exist some remaining problems and room for improvement. In this paper, an improved RBox-based target detection framework is proposed to boost precision and recall rates of detection, and we refer to the method as DRBox-v2 and apply it to target detection in SAR images. The main improvements of DRBox-v2 as well as the contributions of this paper are fourfold. First, a multi-layer prior box generation strategy is designed for detecting small-scale targets. Since shallow layers lack strong sematic information, the feature pyramid network (FPN) module is applied. Second, a modified encoding scheme for RBox is proposed for more precisely estimating the position of RBox and orientation of targets. Third, a focal loss (FL) combined with hard negative mining (HNM) technique is proposed to mitigate the issue of the imbalance between positive and negative samples, which produces better results than solely employing either one. Fourth, comprehensive ablation studies are conducted to reveal the effect of each improvement on detected results. The results of the target detection on three data sets are illustrated and our method obtains 0.135, 0.081, 0.115 gains in average precision compared with three state-of-the-art methods, respectively.

Deep learning (DL) has been successfully applied to single image super-resolution (SISR), which aims at reconstructing a high-resolution (HR) image from its low-resolution (LR) counterpart. Different from most current DL-based methods, which perform reconstruction in the spatial domain, we use a scheme based in the frequency domain to reconstruct the HR image at various frequency bands. Further, we propose a method that incorporates the wavelet transform (WT) and the recursive Res-Net. The WT is applied to the LR image to divide it into various frequency components. Then, an elaborately designed network with recursive residual blocks is used to predict high-frequency components. Finally, the reconstructed image is obtained via the inverse WT. This paper has three main contributions: 1) an SISR scheme based on the frequency domain is proposed under a DL framework to fully exploit the potential to depict images at different frequency bands; 2) recursive block and residual learning in global and local manners are adopted to ease the training of the deep network, and the batch normalization layer is removed to increase the flexibility of the network, save memory, and promote speed; and 3) the low-frequency wavelet component is replaced by an LR image with more details to further improve performance. To validate the effectiveness of the proposed method, extensive experiments are performed using the NWPU-RESISC45 data set, and the results demonstrate that the proposed method outperforms several state-of-the-art methods in terms of both objective evaluation and subjective perspective.