Currently, recurrence plot (RP) is a good tool to analyze the non-stationary sea clutter time series for small target detection in the existing detectors. To further improve detection performance, a dual-channel detector is proposed to fully exploit information from complex time series. First, amplitude time series and phase time series form two channels and generate two RPs. Second, two well-designed features are, respectively, extracted from the two RPs to construct a 2-D feature vector, called relative Shannon entropy (RSE) and relative truncate diagonal length (RTDL). Third, the decision threshold is determined by an improved extreme learning machine (IELM) classifier with false-alarm control, where target-guided prior information is utilized in parameter optimization. Finally, experimental results using IPIX datasets show that the proposed detector can attain good and robust performance, compared with the existing detectors.

In RGB-infrared aerial image object detection, fully utilizing the advantages of both RGB and infrared images for effective detection is a key challenge in this field. In response to the challenges outlined, an RGB-infrared multibranch progressive feature fusion method for object detection in aerial images is proposed. Specifically, since accurate detection of objects in aerial images usually requires extensive context information, considering the connection between objects and backgrounds, the global-local synergistic attention (GLSA) is constructed. To mitigate the impact of feature domain differences between RGB and infrared images during the complementary information filtering process, a multimodal complementary information filter (MCIF) for RGB and infrared images is designed on the basis of GLSA. To fully exploit the advantages of the two types of model images, a multibranch progressive feature fusion network is designed. The multibranch progressive feature fusion network progressively fuses the enhanced features from the GLSA with the filtered complementary information by the MCIF, resulting in a final fused feature that integrates the rich visual details of RGB images with the distinctive physical characteristics of infrared images. Results against publicly released datasets demonstrate that the proposed method achieves state-of-the-art detection performance.

Multimodal remote sensing image object detection enhances detection accuracy by fusing complementary information from multimodal image data. However, complex environments significantly affect the reliability and complementarity of multimodal data, and traditional methods struggle to dynamically adapt to environmental changes, leading to degraded detection performance. To address this challenge, in this paper, we propose a multimodal remote sensing object detection method based on a prior information-enhanced mixture-of-experts fusion network. Specifically, we first introduce a prior information-enhanced mixture-of-experts fusion network framework to achieve environment-adaptive multimodal image feature fusion. Secondly, we propose a dynamic gating network that combines prior information and multimodal image features to endow the system with environmental perception capabilities. This network is employed to dynamically allocate weights to sub-fusion experts optimized for different environmental conditions within the mixture-of-experts fusion network framework. Furthermore, to fully exploit the complementary information present in multimodal image features, we propose a frequency-decoupled feature fusion network as a sub-fusion expert within the mixture-of-experts fusion network framework. This utilizes wavelet transform to decouple the features of each modality and then develops personalized fusion strategies for each frequency subband. In addition, to enhance detection efficiency, we introduce a cross-scale feature channel interleaved fusion strategy, which significantly reduces computational cost while ensuring stable detection performance. Experimental results on the DroneVehicle and RGBT-Tiny datasets demonstrate that our method achieves competitive performance compared to state-of-the-art approaches. Code will be available at: https://github.com/LiuKewei0110/MDPMFN.

To improve target detection performance in SAR images, the YOLOV5 network is modified at four different parts, including the input, backbone, neck, and head modules. At the input end, the local binary patterns (LBP) feature is extracted and concatenated with the original SAR image to enhance low-level texture features. In the backbone network, the deformable convolution network (DCN) is utilized to modify the C3 module to enhance geometrical invariance. The content-aware reassembly of features (CARAFE) upsampling module is used in the neck network to improve the feature extraction performance. Finally, the adaptively spatial feature fusion (ASFF) module is adopted to emphasize key features, thus the prediction feature maps at different levels are further integrated. The experiment against the FaradSAR dataset validates the performance of the proposed method.

The modeling of channel and temporal information is of crucial importance for action recognition tasks. To build a high-performance action recognition network by effectively capturing channel and temporal information, we propose CLS-Net: an action recognition algorithm based on channel-temporal information modeling. The proposed CLS-Net characterizes channel and temporal information by inserting multiple modules to an end-to-end backbone network, including a channel attention module (CA module) for modeling channel information, a long-term temporal module (LT module) and a short-term temporal module (ST module) for modeling temporal information. Specifically, the CA module extracts the correlation between feature channels so the network can learn to selectively strengthen the features containing useful information and suppress the useless features through global information. The LT module moves some channels in the temporal dimension to realize information interaction across time domains and model global temporal information. The ST module enhances the motion-sensitive features by calculating the feature-level frame difference information and realizes the representation of local motion information. Since the multi-module insertion mode directly affects the whole model’s final performance, we propose a novel multi-module insertion mode instead of a simple series or parallel connection to ensure that the multiple modules can complement one another and cooperate with each other more efficiently. CLS-Net achieves SOTA performance on the EgoGesture and Jester dataset in the same type of network and achieves competitive results on the Something-Something V2 dataset.

One open issue of target detection for synthetic aperture radar (SAR) images is the capture effect from the clutter edge and the interfering outliers, including surrounding targets in the multitarget environment, sidelobes, and ghosts. To address this issue, a superpixel-level constant false-alarm rate (CFAR) detector is proposed based on the truncated Gamma statistics for the multilook intensity SAR data. Superpixel segmentation serves as a preprocessing procedure to divide the SAR image into meaningful patches. By automatic clutter truncation in the superpixel-level background clutter window, the real clutter samples are preserved. The experimental results with real SAR images demonstrate that the proposed method achieves better goodness-of-fit performance for the real clutter background with outlier exclusion, yielding a higher target detection rate in the multitarget environments.

To reduce the influence of speckle noise on the pixel-level target detection method, a novel ship detection method is proposed for polarimetric synthetic aperture radar (PoISAR) images based on superpixelwise local contrast measurement. In the proposed method, superpixels are firstly generated based on the improved simple linear iterative clustering (SLIC) method for polarimetric SAR image. Then the local contrast measurement is calculated between a certain superpixel and its neighborhood superpixels to enhance the discriminability between ship targets and clutter background. Thus, better target detection performance can be obtained. The effectiveness of the proposed method can be demonstrated by experiments on both quad-polarized and dual-polarized real SAR images.

In this letter, we propose a superpixel-level target detection approach for synthetic aperture radar (SAR) images. With superpixel segmentation, SAR image is divided into meaningful patches and more statistical information can be provided in superpixels compared with single pixels. The statistical difference between target and clutter superpixels can be measured with the intensity distributions of pixels in them. With the assumption of SAR data obeying Gamma distribution, the superpixel dissimilarity is defined. With this basis, the global and local contrast can be obtained and integrated to enhance target and suppress clutter simultaneously. Thus, better target detection performance can be achieved. Different from traditional target detection schemes based on backscattering difference between target and clutter pixels, the proposed method relies on the statistical difference of superpixels. The effectiveness of the proposed method can be demonstrated with experimental results on real SAR images.

This paper addresses an autofocusing technique for highly squinted airborne synthetic aperture radar (SAR) imaging. In highly squinted mode, the phase errors resulting by trajectory deviations usually exhibit both range and cross-range variance. To circumvent this problem, a 2-D space-variant phase error model is proposed. Then an autofocusing approach is presented to realize precise SAR reconstruction. This approach mainly consists of three processing steps. First, multiple local images are automatically selected based on sliding window operation in both range and cross-range dimensions. For each local image, weighted squint phase gradient autofocus kernel is applied to estimate the local phase error function. Second, the derived multiple local phase error functions are combined to resolve the 2-D space-variant phase error model using a weighted total least square method. Third, the fast factorized back-projection algorithm with pixelwise phase error correction is utilized to obtain focused image eventually. Experiments on both simulated and real-measured SAR data sets validate the focusing performance of the proposed approach.

Due to the accuracy limitation of the navigation system, deviations between the real trajectory and the measured one appear inevitably in airborne synthetic aperture radar (SAR), which degrades the image quality dramatically. To improve the focusing performance, these trajectory deviations should be well estimated and compensated. In this paper, a data-based autofocus approach is proposed to correct the residual 3-D trajectory deviations. This new approach mainly contains two processing stages. The first stage is the local phase error estimation procedure involving small images autofocusing. A gradient function considering smoothness regularization is developed to efficiently achieve the sharpness-maximizing local phase error functions. In the second stage, the local phase error functions are combined to retrieve the residual 3-D trajectory deviations by a proposed weighted total least square method. This approach has been applied on highly squinted and large-swath airborne SAR raw data, respectively. Both real data experiments generate well-focused SAR images by the estimated trajectory parameters, and thus, validate the effectiveness of the proposed autofocus approach.

The back-projection (BP) algorithm is an ideal solution for large-swath airborne SAR imaging. However, space-variant phase errors induced by trajectory deviations dramatically degrade the BP-focusing performance in a large-swath mode. In this letter, we propose an autofocus method that is compatible with the BP imagery, in which the phase-error function is constructed for individual pixels. In the new method, multiple local areas at different illumination directions within the radar beam are synthesized to trace the local phase gradient. From these phase gradient estimates, the accurate pixel-wise phase-error correction for all the contaminated pulses is achievable. This approach is capable of correcting space-variant phase errors with high precision and efficiency for large-swath SAR imaging. Experiments based on the real data that are recorded by a highly squinted SAR system validates the effectiveness of the proposed autofocus method.