Scene Coordinate Regression (SCR) is a learning-based visual localization method that estimates camera pose by directly regressing 3D scene coordinates corresponding to given 2D pixel locations. While effective in small-scale settings, existing methods struggle with illumination and viewpoint variations in large-scale environments. To address this, we present AeroSCR, a novel SCR framework for UAV localization. First, we design a structure-aware dual-stream encoder with an Adaptive Structure-Aware Fusion (ASAF) module, jointly learning semantic and structural representations to enhance robustness to appearance variations. Second, we introduce a region partitioning strategy that decomposes the global regression problem into multiple local subtasks, significantly improving localization performance in large and ambiguous environments. Finally, we construct UAV-SceneLoc2025, a large-scale UAV dataset covering diverse altitudes, viewpoints, lighting, and weather conditions, providing a comprehensive benchmark for SCR evaluation under challenging real-world scenarios. Extensive experiments demonstrate that AeroSCR achieves state-of-the-art performance over existing SCR methods.

The goal of image fusion is to generate a new image that incorporates all the high-quality features from source images, such as the salient features of infrared images and the texture details of visible images. Existing image fusion methods primarily focus on deep feature extraction, often neglecting the importance of shallow features. Furthermore, fusion strategies that rely heavily on human intervention can lead to the loss of feature information and limit deep learning performance. To address these issues, we propose an adaptive multi-scale fusion network based on attention mechanisms, named AMSFusion. The autoencoder component of our method is based on the U-Net architecture, with different modules designed to process shallow and deep multi-scale features, thereby improving feature processing quality while maintaining computational efficiency. The fusion network component employs a channel-spatial attention block (CSA) to assign appropriate weights to heterogeneous features and applies a shifted window attention and convolution mix transformer (SWACmixT) to fuse them, enhancing the network’s ability to adaptively fuse features at different scales. In our implementation, the autoencoder and fusion networks are trained separately, with distinct loss functions for each, further enhancing the method’s capabilities in feature encoding, decoding, and fusion. Qualitative and quantitative experiments on multiple datasets demonstrate the superiority of our method compared to state-of-the-art algorithms. Additionally, experiments on object detection tasks validate that our method effectively promotes high-level computer vision tasks.

The performance of existing object detection algorithms significantly degrades when applied to low-resolution infrared (IR) images captured by unmanned aerial vehicles (UAVs), which suffers from slow inference speed, low detection precision, and redundant network parameters. To tackle these issues, this paper proposes a lightweight real-time object detection network for UAV IR images, termed SGA-YOLO, which is designed based on the you only look at once version 8n (YOLOv8n) framework. First of all, the efficient SENetV2-neck enhances the correlation between different channels, which realizes efficient multi-scale feature fusion and improves detection precision. Subsequently, the lightweight S2GM backbone combines ShuffleNetV2-stride2 and C2f_Ghost modules, which significantly reduces the network parameters and increases inference speed. Finally, the adaptive fine-grained channel (AFGC) attention mechanism is coupled to further enhance detection precision and effectively mitigate background interference. Compared with the YOLOv8n, SGA-YOLO achieves a 27% reduction in network parameters, a 4.7% increment in precision, a 2.5% increment in recall rate, a 30.86% reduction in GFLOPs, a 16.3% increment in FPS, a 3.50% increment in [email protected], and a 1.3% increment in [email protected]:0.95. In addition, it supports deployment on resource-constrained embedded system, offering a new perspective on designing lightweight UAV IR object detection networks for real-world applications in intelligent transportation systems. Our codes are available athttps://github.com/gepengqiang2025/SGA-YOLO.

Nowadays, infrared target tracking has been a critical technology in the field of computer vision and has many computational social system-related applications, such as urban security, pedestrian counting, smoke and fire detection, and so forth. Unfortunately, due to the absence of detailed information such as texture or color, it is easy for tracking drift to occur when the tracker encounters infrared targets that vary in shape or size. In order to address this issue, we present a twofold structured features-based Siamese network for infrared target tracking. Above all, a novel feature fusion network is proposed to make full use of both shallow spatial information and deep semantic information in a comprehensive manner, so as to improve the discriminative capacity for infrared targets. Then, a multitemplate update module is designed to effectively deal with interferences from target appearance changes which are prone to cause early tracking failures. Finally, both qualitative and quantitative experiments are implemented on VOT-TIR 2016 and GTOT datasets, which demonstrates that our method achieves the balance of promising tracking performance and real-time tracking speed against other state-of-the-art trackers.

Hyperspectral images (HSIs) with high spatial resolution are challenging to obtain directly due to sensor limitations. Deep learning is able to provide an end-to-end reconstruction solution from low to high spatial resolution. Nevertheless, existing deep learning-based methods have two main drawbacks. First, deep networks with self-attention mechanisms often require a trade-off between internal resolution, model performance and complexity, leading to the loss of fine-grained, high-resolution features. Second, there are visual discrepancies between the reconstructed hyperspectral image (HSI) and the ground truth because they focus on spatial-spectral domain learning. In this paper, a novel super-resolution algorithm for HSIs, named SRDNet, is proposed by using a dual-domain network with hybrid convolution and progressive upsampling to exploit both spatial-spectral and frequency information of the hyperspectral data. In this approach, we design a self-attentive pyramid structure (HSL) to capture interspectral self-similarity in the spatial domain, thereby increasing the receptive range of attention and improving the feature representation of the network. Additionally, we introduce a hyperspectral frequency loss (HFL) with dynamic weighting to optimize the model in the frequency domain and improve the perceptual quality of the HSI. Experimental results on three benchmark datasets show that SRDNet effectively improves the texture information of the HSI and outperforms state-of-the-art methods. The code is available at https://github.com/LTTdouble/SRDNet.

The quality of underwater imaging is significantly degraded by light scattering and absorption due to water body and suspended particles. To address the issues of color distortion and contrast degradation, we propose a novel underwater image restoration method based on constrained color compensation and background light color space-based haze-line model. Our method begins by applying the constrained color compensation approach to provide targeted intensity adjustments for attenuated color channels. This process corrects color distortion while simultaneously mitigating the risks of overcompensation and insufficient color saturation. Subsequently, the haze-line model is employed in the transformed background light color space to restore the underwater image. Specifically, the color corrected underwater image is transformed to a novel color space, where the background light intensity serves as the origin. In this color space, all haze lines are identified by grouping pixels with similar color characteristics through the superpixel clustering method. Then, the transmission distribution can be estimated based on the haze-line model. Finally, the scattered light components are removed by applying the underwater descattering model with the estimated transmission distribution to the luminance channel of the color corrected underwater image. Comparative experiments implemented on the UIEB and UCCS underwater image datasets demonstrate the superiority of the proposed method in terms of color correction and contrast enhancement when compared with state-of-the-art underwater image restoration techniques. Our codes are available athttps://github.com/MinjieWan/C3HLM.

The field of remote sensing change detection (RSCD) has seen significant advancements recently, focusing on the precise identification and analysis of temporal changes in remote sensing images. Existing deep learning-based RSCD methods primarily rely on concatenation or subtraction to integrate features of bi-temporal images and reconstruct change features through a feature pyramid network (FPN) decoding architecture. However, these methods face challenges related to inadequate spatio-temporal change representation and insufficient aggregation of multilevel semantic information, resulting in pseudo-changes and poor completeness of detected change objects. In this article, we propose an innovative RSCD framework via spatio-temporal feature fusion and guide aggregation (STFF-GA) to address the aforementioned challenges. The architecture of this network comprises two key components: the STFF module and the GA module. The STFF module is designed as a low-parameter and low-computation structure, effectively enhancing the representation of spatio-temporal change information through split, interaction, and fusion strategies. The GA module uses deep feature guidance (DFG) mapping as prior information to guide the aggregation of multilevel semantic information, thereby correcting the positional information of change objects and filtering out pseudo-changes and other noise interference. In addition, it utilizes convolution kernels of various scales to extract fine-grained features, facilitating the complete reconstruction of change objects. Extensive experiments conducted on three benchmark change detection datasets demonstrate that the proposed STFF-GA consistently outperforms other state-of-the-art (SOTA) detectors. The code is available athttps://github.com/NjustHGWei/STFF-GA.

Change detection (CD) in remote sensing (RS) images is a technique used to analyze and characterize surface changes from remotely sensed data at different time periods. However, current deep-learning-based methods sometimes struggle with the diversity of targets in complex RS scenarios, leading to issues, such as false detections and loss of detail. To address these challenges, we propose a method called difference learning-based convolution and network (DLB-CNet) for building CD (BCD). In DLB-CNet, we use difference learning module (DLM), accomplishing the extraction of building change features by enhancing the feature differences between the two images and enhancing model robustness. Additionally, an innovative attention module called integration attention (IA) is introduced to efficiently process semantic information by jointly focusing on global representation subspaces. Our model achieves impressive results on the LEVIR-CD dataset, WHU-CD dataset, and CDD dataset, with${F}1$-scores of 90.56%, 92.28%, and 94.98%, respectively, demonstrating its superiority over the state-of-the-art methods.

Maritime infrared (IR) small target detection in dense sunglint environments has always been challenging. It is difficult for existing methods to distinguish the small target from dense and strong sea clutter because the sea clutter caused by the sunglints has similar spatial-temporal characteristics to small targets. This paper proposes a method based on appearance stability and depth-normalized motion saliency (AS-DNMS) for small target detection in IR videos or sequences. First, on each frame of the IR video, the isotropic salience measure (ISM) based on principal curvature filtering is proposed for target enhancement, and an adaptive threshold is used to extract a stable number of candidate targets. Then, we use improved pipeline filtering to form trajectory chains of the candidate targets extracted from consecutive frames. To improve the matching accuracy of pipeline filtering, we propose an inverse optical flow method to predict the local images of the candidate targets. Third, the motion vectors of the candidate targets are extracted and depth-normalized. According to the principle of projection perspective, a depth normalization model based on the position of the sea-sky line is established, providing a simple and low-cost solution for acquiring depth information of maritime targets under sea-sky backgrounds. Finally, the appearance stability measure (ASM) and the depth-normalized motion saliency measure (DNMSM) are calculated to construct the AS-DNMS joint feature. A double asymptote decision function is employed to determine the real target. The experimental results show that our method has better detection performance in dense sunglint environments than the baseline methods.

Realizing robust infrared small target detection in complex backgrounds is of great essence for infrared search and tracking (IRST) applications. However, the high-intensity structures in background regions, such as the sharp edges, make it a challenging task, especially when the target is with low signal-to-clutter ratio (SCR). To address this issue, we propose an infrared small target detection method using local contrast-weighted multidirectional derivative (LCWMD). It is a robust detector that comprehensively considers the target property, background information, and the relation between them. First, we consider the approximate isotropy of the infrared small target and present a new multidirectional derivative with penalty factors based on the Facet model to develop the target salience in the local region. Second, a dual local contrast fusion model with the trilayer design is introduced to amplify the difference between the target and the background, so as to further suppress the high-intensity structural clutters. Finally, the LCWMD map is obtained by weighting the above two filtered maps, after which an adaptive segmentation operation is applied to accomplish the target detection. The results of comparative experiments implemented on real infrared images demonstrate that our method outperforms other state-of-the-art detectors by several times in terms of SCR gain (SCRG) and background suppression factor (BSF).

Infrared (IR) small target detection is one of the most fundamental techniques in the infrared search and track (IRST) system. Due to the interferences caused by background clutter and image noise, conventional IR small target detection algorithms always suffer from a high false alarm rate and are unable to achieve robust performance in complex scenes. To accurately distinguish IR small target from the background, we propose a total variation (TV)-based interframe infrared patch-image model that regards the long-distance IR small target detection task as an optimization problem. First, the input IR image is converted to a patch-image that consists of a sparse target matrix and a low-rank background matrix. Then, the interframe similarity of target appearance is utilized to impose a temporal consistency constraint on the target matrix. Next, a TV regularization term is proposed to further alleviate the false alarms generated by noise. Finally, an alternating optimization algorithm using singular value decomposition (SVD) and accelerated proximal gradient (APG) is designed to mathematically solve the proposed model. Both qualitative and quantitative experiments implemented on real IR sequences demonstrate that our model outperforms other traditional IR small target methods in terms of the signal-to-clutter ratio gain (SCRG) and the background suppression factor (BSF).

The precision of infrared (IR) small target tracking is seriously limited due to lack of texture information and interference of background clutter. The key issue of robust tracking is to exploit generic feature representations of IR small targets under different types of background. In this letter, we present a new IR small target tracking method via compressive convolution feature (CCF) extraction. First, a Gaussian curvature-based feature map is calculated to suppress clutters so that the contrast between target and background can be obviously improved. Then, a three-layer compressive convolutional network, which consists of a simple layer, a compressive layer, and a complex layer, is designed to represent each candidate target by a CCF vector. Based on the proposed mechanism of feature extraction, a support vector machine (SVM) classifier with continuous probabilistic output is trained to compute the likelihood probability of each candidate. Finally, the long-term tracking for IR small target is implemented under the framework of the inverse sparse representation-based particle filter. Both qualitative and quantitative experiments based on real IR sequences verify that our method can achieve more satisfactory performances in terms of precision and robustness compared with other typical visual trackers.

Target tracking based on unmanned aerial vehicle (UAV) video is a significant technique in intelligent urban surveillance systems for smart city applications, such as smart transportation, road traffic monitoring, inspection of stolen vehicle, etc. In this paper, a computer vision-based target tracking algorithm aiming at locating UAV-captured targets, like pedestrian and vehicle, is proposed using sparse representation theory. First of all, each target candidate is sparsely represented in the subspace spanned by a joint dictionary. Then, the sparse representation coefficient is further constrained by an L2regularization based on temporal consistency. To cope with the partial occlusion appearing in UAV videos, a Markov random field (MRF)-based binary support vector with contiguous occlusion constraint is introduced to our sparse representation model. For long-term tracking, the particle filter framework along with a dynamic template update scheme is designed. Both qualitative and quantitative experiments implemented on visible (Vis) and infrared (IR) UAV videos prove that the presented tracker can achieve better performances in terms of precision rate and success rate when compared with other state-of-the-art trackers.

Zadoff-Chu (ZC) sequences have been used as random access sequences in modern wireless communication systems, replacing the conventional pseudo-random-noise (PN) sequences due to their superior autocorrelation properties. An analytical framework quantifying the ZC sequence's performance and its fundamental limitation as a random access sequencein the presence of frequency offset between the transmitter and the receiver is introduced. We show that a ZC sequence's perfect autocorrelation properties can be severely impaired by the frequency offset thereby limiting the overall performance of the random access signals formed from these sequences. First, we derive the autocorrelation function of these random access sequences as a function of the frequency offset. Next, we introduce the concept of critical frequency offsets and the spectrum associated with a ZC sequence set to characterize the frequency offset properties of the random access signals. Finally, we demonstrate that the frequency offset immunity of a ZC sequence set can be controlled by shaping the spectrum of the ZC sequence set.

This paper proposes a unique discovery signal as an enabler of peer-to-peer (P2P) communication which overlays a cellular network and shares its resources. Applying P2P communication to cellular network has two key issues: 1. Conventional ad hoc P2P connections may be unstable since stringent resource and interference coordination is usually difficult to achieve for ad hoc P2P communications; 2. The large overhead required by P2P communication may offset its gain. We solve these two issues by using a special discovery signal to aid cellular network-supervised resource sharing and interference management between cellular and P2P connections. The discovery signal, which facilitates efficient neighbor discovery in a cellular system, consists of un-modulated tones transmitted on a sequence of OFDM symbols. This discovery signal not only possesses the properties of high power efficiency, high interference tolerance, and freedom from near-far effects, but also has minimal overhead. A practical discovery-signal-based P2P in an OFDMA cellular system is also proposed. Numerical results are presented which show the potential of improving local service and edge device performance in a cellular network.

Accurate channel information at the transmitter is crucial to multi-user MIMO performance. Unfortunately, the bandwidth of the control channel by which the feedback is conveyed is often limited. An important issue is how to improve multi-user MIMO performance with minimal feedback. Conventional feedback techniques focus on improving the quantized codebook performance using various quantization criteria. In this paper, instead of trying to optimize a single codebook, we apply multiple alternating codebooks to effectively reduce the multi-user MIMO quantization error in addition to the reduction provided by the single codebook optimization techniques. That is, we use an existing quantization codebook design methodology to create not one but multiple such similar codebooks. The codebooks are alternated at each feedback instance creating a larger virtual codebook with the same number of feedback bits as the single smaller codebook. Simulation results show that significant performance gain in multi-user MIMO systems is obtained via the alternating codebook scheme.

The spatial fixed-pattern noise (FPN) compromises severely the quality of the acquired imagery, even makes such images inappropriate for some applications. In order to lower the FPN, some critical nonuniformity correction (NUC) algorithms such as NUC based on linear model, scene-based NUC and so on have been developed. But each algorithm has some drawbacks: restricted application in small dynamic range of objects temperature, low performance under the drift of the environment temperature and complex calculations. In these cases, we develop a bivariate and quadratic model of the FPA and the NUC technique based on the model. The proposed method does not need any assumptions and is a good solution for hardware implementation. It overcomes the drawbacks of the critical algorithm mentioned above. The last simulations and experiments show that the proposed algorithm exhibits a superior correction effect in both large objects temperature range and environment temperature range.

A dual-threshold method is proposed in this paper to reduce the spurious photon events in photon counting imaging with a thermoelectric cooling electron multiplying charge coupled device (EMCCD). Using the thresholds in both spatial and time domains, the read noise, dark signal and clock induced charges (CIC) can be filtered simultaneously. The theory and measurements of the dual-threshold technique are presented. The results show that this method works better than 5σ threshold method in restoring photon images with the EMCCD of the working temperature at -20 °C. The sensitivity of thermoelectric cooling EMCCDs can be effectively improved by the dual-threshold technique. Photon images could be still restored under the illumination level of 0.1e-/pixel/frame.