Object detection has developed rapidly with the help of deep learning technologies recent years. However, object detection on drone view remains challenging due to two main reasons: (1) It is difficult to detect small-scale objects lacking detailed information. (2) The diversity of camera angles of drones brings dramatic differences in object scale. Although feature pyramid network (FPN) alleviates the problem caused by scale difference to some extent, it also retains some worthless features, which wastes resources and slows down the speed. In this work, we propose a novel High-Resolution Feature Pyramid Network (HR-FPN) to improve the detection accuracy of small-scale objects and avoid feature redundancy. The key components of HR-FPN include a high-resolution feature alignment module (HRFA), a high-resolution feature fusion module (HRFF) and a multi-scale decoupled head (MSDH). HRFA feeds multi-scale features from backbone into parallel resampling channels to obtain high-resolution features at the same scale. HRFF establishes a bottom-up path to distribute context-rich low-level semantic information to all layers that are then aggregated into classification feature and localization feature. MSDH cope with the scale difference of objects by predicting the categories and locations corresponding to different scales of objects separately. Moreover, we train model by scale-weighted loss to focus more on small-scale objects. Extensive experiments and comprehensive evaluations demonstrate the effectiveness and advancement of our approach.

Synthetic aperture radar (SAR) image interpretation has long been an important but challenging task in SAR imaging processing. Generally, SAR image interpretation comprises complex procedures including filtering, feature extraction, image segmentation, and target recognition, which greatly reduce the efficiency of data processing. In an era of deep learning, numerous automatic target recognition methods have been proposed based on convolutional neural networks (CNNs) due to their strong capabilities for data abstraction and mining. In contrast to general methods, CNNs own an end-to-end structure where complex data preprocessing is not needed, thus the efficiency can be improved dramatically once a CNN is well trained. However, the recognition mechanism of a CNN is unclear, which hinders its application in many scenarios. In this paper, Self-Matching class activation mapping (CAM) is proposed to visualize what a CNN learns from SAR images to make a decision. Self-Matching CAM assigns a pixel-wise weight matrix to feature maps of different channels by matching them with the input SAR image. By using Self-Matching CAM, the detailed information of the target can be well preserved in an accurate visual explanation heatmap of a CNN for SAR image interpretation. Numerous experiments on a benchmark dataset (MSTAR) verify the validity of Self-Matching CAM.

This paper investigates the problem of specific emitter identification (SEI), i.e. radar emitter fingerprint or individual emitter identification, which first measures the emitter-specific differences caused by radar's nonlinearities, e.g. mixer, power amplifiers, transmitter, and then makes a decision. In this paper, the SEI problem is considered in the single-modal, dual-modal and multi-modal scenarios, respectively. First, a multi-modal subspace interactive mutual unit is proposed to perform information interaction between radar signal and its multiple transformations. Based on this, a data-driven multi-modal subspace interactive mutual network (MSIMN) is then built to solve the SEI problem. Extensive experiment results demonstrate that the proposed algorithm achieves superior identification performance on the airplane measured data.

Tracking-by-detection algorithms have considerably enhanced tracking performance with the introduction of recent convolutional neural networks (CNNs). However, most trackers directly exploit standard scalar-output CNN features, which may not capture enough feature encoding information, instead of aggregated CNN features of vector-output form. In this paper, we propose an end-to-end feature aggregation capsule framework. First, based on the existing CNN network, we aggregate a certain number of similar position-aware CNN features into a capsule to model the feature similarity. The acquired vector-level feature capsules (rather than previous scalar-level pointwise features) are utilized for differentiation learning. We then propose a group attention module to better model the entity representation between different capsule groups thus optimizes total discriminative capability. Third, to reduce the prediction interference resulted by the side effect of dimension rising within capsules, we propose a penalty attention module. Such strategy could dynamically adjust values of neurons by estimating whether they are beneficial or harmful to tracking. Experimental results on five representative benchmarks (UAVDT, DTB70, UAV123, VOT2016 and VOT2018) demonstrate the excellent tracking performance of our dual attention based capsule tracker (DACapT). Specially, it exceeds the previous top tracker by 4.6%/1.9% in precision/success evaluations on UAVDT.