Hyperspectral anomaly detection (HAD) is a challenging task since it identifies the anomaly targets without prior knowledge. In recent years, deep learning methods have emerged as one of the most popular algorithms in the HAD. These methods operate on the assumption that the background is well reconstructed while anomalies cannot, and the degree of anomaly for each pixel is represented by reconstruction errors. However, most approaches treat all background pixels of a hyperspectral image (HSI) as one type of ground object. This assumption does not always hold in practical scenes, making it difficult to distinguish between backgrounds and anomalies effectively. To address this issue, a novel deep feature aggregation network (DFAN) is proposed in this paper, and it develops a new paradigm for HAD to represent multiple patterns of backgrounds. The DFAN adopts an adaptive aggregation model, which combines the orthogonal spectral attention module with the background-anomaly category statistics module. This allows effective utilization of spectral and spatial information to capture the distribution of the background and anomaly. To optimize the proposed DFAN better, a novel multiple aggregation separation loss is designed, and it is based on the intra-similarity and inter-difference from the background and anomaly. The constraint function reduces the potential anomaly representation and strengthens the potential background representation. Additionally, the extensive experiments on the six real hyperspectral datasets demonstrate that the proposed DFAN achieves superior performance for HAD. The code is available at https://github.com/ChengXi-1217/DFAN-HAD.

Convolutional neural networks (CNNs) have achieved milestones in object detection of synthetic aperture radar (SAR) images. Recently, vision transformers and their variants have shown great promise in detection tasks. However, ship detection in SAR images remains a substantial challenge because of the characteristics of strong scattering, multi-scale, and complex backgrounds of ship objects in SAR images. This paper proposes an enhancement Swin transformer detection network, named ESTDNet, to complete the ship detection in SAR images to solve the above problems. We adopt the Swin transformer of Cascade-R-CNN (Cascade R-CNN Swin) as a benchmark model in ESTDNet. Based on this, we built two modules in ESTDNet: the feature enhancement Swin transformer (FESwin) module for improving feature extraction capability and the adjacent feature fusion (AFF) module for optimizing feature pyramids. Firstly, the FESwin module is employed as the backbone network, aggregating contextual information about perceptions before and after the Swin transformer model using CNN. It uses single-point channel information interaction as the primary and local spatial information interaction as the secondary for scale fusion based on capturing visual dependence through self-attention, which improves spatial-to-channel feature expression and increases the utilization of ship information from SAR images. Secondly, the AFF module is a weighted selection fusion of each high-level feature in the feature pyramid with its adjacent shallow-level features using learnable adaptive weights, allowing the ship information of SAR images to be focused on the feature maps at more scales and improving the recognition and localization capability for ships in SAR images. Finally, the ablation study conducted on the SSDD dataset validates the effectiveness of the two components proposed in the ESTDNet detector. Moreover, the experiments executed on two public datasets consisting of SSDD and SARShip demonstrate that the ESTDNet detector outperforms the state-of-the-art methods, which provides a new idea for ship detection in SAR images.

Recently, the autoencoder (AE)-based methods in hyperspectral anomaly detection (HAD) have attracted a lot of attention from scholars and researchers, and they acquire satisfying detection accuracy. However, most existing AE-based methods ignore good prior knowledge which is the low-rank and sparse property in a hyperspectral image (HSI). To this end, a novel deep self-representation learning framework (DSLF) is presented in this article. The DSLF combines the robust principal component analysis with the AE model to achieve an alternating optimization, and it is adaptive to remove anomalies and optimize the background reconstruction. To further characterize the HSI background, a novel subspace recovery AE (SRAE) is put forward and a joint loss function is introduced to construct the background subspace, which makes anomalies farther away from the background subspace. Notably, the SRAE is different from the previous AE-based model. The superpixel segmentation divides an HSI into some locally homogeneous regions, and the training and testing of each SRAE is based on one of the local regions. To fully exploit the spatial information, the multi-scale strategy is proposed to generate several reconstruction results and these results are fused to further optimize detection performance. Additionally, the experiment results on the six HSI datasets verify that the proposed DSLF is superior to the state-of-the-art methods.

Linear-based low-rank and sparse models (LRSM) and nonlinear-based deep autoencoder (DAE) models have been proven to be effective for the task of anomaly detection (AD) in hyperspectral images (HSIs). The linear-based LRSM is self-explainable, while it may not characterize the complex scenes well. In contrast, the nonlinear-based DAE is able to extract the discriminative features between the background and anomaly for the complex scenes, whereas it is not self-explainable. To effectively combine the advantages of both, a dynamic low-rank and sparse priors-constrained DAEs (DLRSPs-DAEs) for hyperspectral AD (HAD), in this article, is proposed. In order to utilize the low-rank prior existing in an HSI, a low-rank prior-based DAE (DAE_LR) is designed to generate an excellent background reconstruction effect and terrible anomaly reconstruction performance. Further, to consider the sparsity reflecting the anomalies in the HSI, a DAE that is constrained by the sparse prior obtained by the decomposition of the HSI (DAE_S) is developed. Notably, to make the model more compact, the DAE_LR and DAE_S share a common encoder. To achieve global optimal performance, an end-to-end joint optimization strategy with the consideration of the interaction between the learning of the DAEs and the decomposition of the HSI is proposed. Additionally, to yield better detection performance, a nonlinear fusion strategy is exploited to comprehensively combine the detection results obtained from both the DAE_LR and DAE_S. Extensive experiments conducted on several datasets show that the proposed DLRSPs-DAEs detector achieves tremendous performance with respect to the classical and state-of-the-art detectors.

Hyperspectral anomaly detection (HAD) is a challenging task in hyperspectral image processing, which is to capture the anomaly by spectral and spatial information without prior knowledge. Recently, some isolation forest (IF) methods in the HAD are proposed to achieve good accuracy. However, these methods build isolation trees by the global pixels and single band partition, and the way limits the utilization of spectral-spatial information, resulting in suffering from poor performance in detecting hard anomalies. To this end, a novel two-stream isolation forest based on deep features is proposed for HAD in this article, and it fully exploits the spectral and spatial information to enhance the sensitivity of anomalies. Specifically, the IF model combined with the neural network is first introduced for the HAD, by which deep spectral features are utilized to better guide the construction of IFs. Meanwhile, based on spatial characteristics, the model of two-stream IF is proposed, which is global and local weighted IF. Moreover, to further mine spatial information, the morphological attribute filter and Gaussian filter are applied to generate the spatial detection map, and the spatial result and the two-steam IF spectral result are fused through the nonlinear transformation to obtain the final result. Finally, the experimental results on five hyperspectral data sets demonstrate that the proposed method is superior to other state-of-the-art methods.

Hyperspectral anomaly detection (HAD) aims to identify targets that are significantly different from their surrounding background, employing an unsupervised paradigm. Recently, detectors based on autoencoder (AE) have become predominant methods and demonstrated satisfactory performance. However, there are still two problems that need to be solved. First, the hypothesis that the AE-based models can effectively reconstruct background samples (BSs) while anomalies cannot, may not always be true in practice, due to their powerful capability for feature extraction. Second, the AE-based models primarily concentrate on the quality of sample reconstruction, regardless of whether the encoded features signify the anomalies or background, which is not conducive to the separation of anomalies from the background. To handle the above-mentioned problems, a novel memory-augmented AE (MAAE) model is developed to better reconstruct the background and suppress anomalies reconstruction. Specifically, for the first problem, a novel superpixel-guided adaptive weight calculation (SAWC) module is devised to generate adaptive weights (AWs) by taking into account contextual information in the error map, and then the AWs are incorporated into the reconstruction loss, where the potential BSs are endowed with larger AWs than anomalies during training. For the second problem, a novel sample attribution mining (SAM) module is developed to mine sample attribution (i.e., explore whether a certain sample belongs to the background or anomaly), and the mined background and anomaly samples (ASs) are employed to train different modules for better separating the anomalies and background. In addition, an entropy-based sparse addressing (ESA) module is further designed to weaken the reconstruction ability of ASs by designing a learnable sparse addressing weight for the memory module. The ablation study validates the effectiveness of the proposed SAWC, SAM, and ESA. Extensive comparison experiments on six hyperspectral image (HSI) datasets demonstrate the superiority in terms of comprehensive detection performance and background suppression of our method.