Ruiqiu Wang

Due to the limited number of SAR samples in the dataset, current networks for SAR automatic target recognition (SAR ATR) are prone to overfitting the environmental information, which diminishes their generalization ability under cross-background conditions. However, acquiring sufficient measured data to cover the entire environmental space remains a significant challenge. This paper proposes a novel feature disentanglement network, named FDSANet. The network is designed to decouple and distinguish the features of the target from the background before classification, thereby improving its adaptability to background changes. Specifically, the network consists of two sub-networks. The first is an autoencoder sub-network based on dual-mask-guided slot attention. This sub-network utilizes target mask to guide the encoder to distinguish between target and background features. It then outputs these features as independent representations, respectively, achieving feature disentanglement. The second is a classification sub-network. It includes an encoder and a classifier, which work together to perform the classification based on the extracted target features. This network enhances the causal relationship between the target and the classification result, while mitigating the background’s interference on the classification. Moreover, the network, trained under a fixed background, demonstrates strong adaptability when applied to a new background. Experiments conducted on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset, as well as the OpenSARShip dataset, demonstrate the superior performance of FDSANet.

Optical-SAR image registration has wide applications in change detection and disaster early warning. However, the significant differences in noise characteristics and radiometric properties between the two types of images remain major challenges for achieving accurate registration. In this article, we propose an automatic registration algorithm for optical and SAR images based on phase congruency and multimoment features. First, a multiangle bandwidth log-Gabor filter bank is designed to enhance structural details and extract phase congruency information. Based on the enhanced phase congruency map, we compute the maximum and minimum moment maps (MaxMM and MinMM) from the phase congruency responses, which, respectively, capture salient edge and corner features that are robust to radiometric and geometric variations. A novel feature detection strategy is applied on these moment maps, followed by a voting mechanism to select highly stable keypoints. For feature description, we introduce the multimoment orientation histogram, which concatenates histograms computed from MaxMM and MinMM, significantly improving both robustness and distinctiveness in heterogeneous image matching. Comprehensive experiments on real-world multisource datasets demonstrate that the proposed OS-PCMM algorithm achieves superior accuracy and robustness compared to state-of-the-art methods.

The presence of disparities in radiation characteristics and diverse noise interferences is ubiquitous in multi-source remote imagery (MSRI), which is captured across various platforms, at different times, and from multiple angles. Unfortunately, these problems significantly hinder the performance of gradient-based MSRI registration algorithms. To mitigate these challenges, we introduce an novel methodology for deriving feature descriptors, leveraging the mapped anisotropic diffusion coefficient (ADC) as the cornerstone. Initially, the images are processed by a series of anisotropic diffusion filters to construct an anisotropic scale space (ASS) and facilitate the subsequent computation of the Mapped ADC. Subsequently, within the Harris scale space, rooted in the ASS, we meticulously extract features. Finally, within the feature circular neighborhood under the log-polar coordinate system, the mapped ADC values are statistically summarized into histograms, resulting in the feature descriptor. Comparing with the state-of-the-art descriptors: scale invariant feature transform (SIFT), a SIFT-like algorithm named OS-SIFT, the frequency-domain descriptor named RIFT and I-KAZE based on anisotropic filtering, the ADCOH emerges as a superior choice, demonstrating remarkable distinctiveness and robustness.

Neural networks for synthetic aperture radar (SAR) automatic target recognition often encounter overfitting challenges owing to limited training samples. Moreover, the azimuth angle of SAR, a vital parameter for improving network generalization, is frequently disregarded in most models. In response, we propose MIGA-Net, a classification neural network that effectively perceives azimuthal information using multi-view images to improve classification performance. Specifically, we quantize low-dimensional azimuthal values for sample-limited scenarios. Then, we utilize encoded image sequences as training data because they encompass spatial context information compared to individual images. After extracting features of the sequence samples through convolutional layers, we design a two-layer output module. One layer converts these sequence features into graph data. Then the dense graph attention network (GAT) extracts contextual features from the graph data for angle estimation. Simultaneously, another layer combines these features for target classification. During the network training, the GAT module can extract image azimuth features with powerful information aggregation capabilities. It supervises the convolutional layers to learn azimuth features, which are fused with class features from another layer to obtain a more structured feature domain. This feature domain significantly enhances the classification performance of the network. Experiments conducted on the moving and stationary target acquisition and recognition (MSTAR) dataset have proven the superior performance of the proposed method, achieving at least 1% higher accuracy compared to other state-of-the-art algorithms.