In modern complex environments, achieving accurate and efficient target localization is essential in numerous fields. However, existing systems often face limitations in both accuracy and the ability to recognize small targets. In this study, we propose a bionic stabilized localization system based on CA-YOLO, designed to enhance both target localization accuracy and small target recognition capabilities. Acting as the ”brain” of the system, the target detection algorithm emulates the visual focusing mechanism of animals by integrating bionic modules into the YOLO backbone network. These modules include the introduction of a small target detection head and the development of a Characteristic Fusion Attention Mechanism (CFAM). Furthermore, drawing inspiration from the human Vestibulo- Ocular Reflex (VOR), a bionic pan-tilt tracking control strategy is developed, which incorporates central positioning, stability optimization, adaptive control coefficient adjustment, and an intelligent recapture function. The experimental results show that CA-YOLO outperforms the original model on standard datasets (COCO and VisDrone), with average accuracy metrics improved by 3.94%and 4.90%, respectively. Further time-sensitive target localization experiments validate the effectiveness and practicality of this bionic stabilized localization system.

Accurate formation pore pressure prediction is essential for hydrocarbon exploration and development. Extensive exploration practices have demonstrated that subsurface formation pore pressure often exhibits nonlinear abrupt variations. Taking the Xihu Sag in the East China Sea as a case, mutant formation pore pressure (MFPP) is highly prevalent in Paleogene adjacent wells and formations, resulting in frequent drilling incidents within this sag. However, field researchers face challenges in predicting these pressure anomalies using analogy-based methods, which depend on data from nearby wells and formations. Conventional prediction approaches (e.g., Eaton’s and Bowers’ method) are constrained by linear assumptions and regional empirical parameter dependency, leading to significant errors in MFPP prediction. Typically, these methods yield prediction errors of approximately 5% in simple, gradual pressure intervals, while errors can fluctuate between 5% and 20% in complex pressure mutation intervals. To address these limitations and capture the underlying nonlinear relationships, deep learning methods have proven particularly advantageous. This study employed a deep learning approach to construct a model that correlates well-log lithology parameters with pore pressure. A hybrid neural network architecture called CNN-LSTM-ATTENTION was created by merging the benefits of several networks, innovating the model structure. Through 5-fold cross-validation, this architecture demonstrated its suitability for MFPP prediction. With an average accuracy of 96.7% in blind testing, the results suggest that this model performs exceptionally well predicting the MFPP. It greatly improves the MFPP's prediction accuracy compared to traditional pressure prediction methods. Meanwhile, it validates the feasibility of deep learning algorithms in addressing the challenges associated with MFPP.

Aiming at the core issues in complex scenarios, such as feature loss in small object detection, feature overlap and confusion in dense object detection, and the difficulty in balancing accuracy and lightweight efficiency in existing solutions, this paper proposes an improved method inspired by the bee visual mechanism-Bee-YOLO. Built upon the YOLO11 framework, the algorithm enhances detection performance via two core improvements: first, the innovative Bee-eye Dual-branch Enhancement Module (BDEM), which draws on the dual-functional differentiation of bee compound eyes to construct a dual-branch structure, integrating gated fusion and attention mechanisms to alleviate small object feature loss and dense object feature overlap; second, detection head reconstruction, which removes redundant branches and adds a dedicated detection head for shallow-level features. Using high-resolution features and upsampling, this reconstruction mitigates low localization accuracy for small objects while streamlining computation to balance accuracy and efficiency. Experimental results show that on the VisDrone and COCO datasets, the mAP@50 of the proposed method is increased by 3.7% and 2.5% respectively compared with the baseline YOLO11, verifying the effectiveness of the proposed improvements. This study provides a feasible solution for small and dense target detection in complex scenarios and shows potential application value in intelligent surveillance, autonomous driving and other fields.

Abstract To address feature dilution and overlap in small and dense target coexistence scenarios, this paper proposes an efficient improved framework, Enhanced Feature-Aware YOLO (EFA-YOLO). Based on YOLO11, EFA-YOLO integrates targeted feature enhancement, adaptive fusion, and dedicated detection head design, combining frequency-domain enhancement, collaborative attention, and adaptive fusion to preserve fine-grained details and global context.A Multi-Scale Frequency-Aware Attention (MSFA) module adopts multi-directional wavelet decomposition to strengthen high-frequency cues for small targets, while maintaining coherent semantic representations in dense scenes via coordinated channel-spatial attention. To eliminate redundancy from static feature concatenation, an Adaptive Weighted Concatenation Fusion (AWCF) mechanism dynamically balances cross-scale contributions for cleaner fused features.Additionally, the small-target detection head is restructured with shallow high-resolution features to retain micro-scale information, and redundant branches are pruned for efficiency. These designs enable EFA-YOLO to robustly handle small and dense targets without introducing excessive architectural complexity. Experiments on the visdrone2019 and coco datasets show that compared with the baseline model yolo11 mAP50 It has increased by 3.7% and 2.8% respectively, which can effectively solve the core pain points of small target and dense target detection, and provide technical support for actual scenes such as autonomous driving and UAV patrol inspection.

The introduction of noise during image acquisition and transmission is inevitable, leading to a significant reduction in the accuracy of image processing tasks, such as target classification, localization, and recognition. To address this issue, this paper proposes a novel robustness-oriented pooling module called Gaussian adaptive pooling. Drawing on the principles of Gaussian filters, the method introduces a Gaussian weight for feature values in the pooling operation, thus integrating filtering and pooling in a novel manner. This approach is both lightweight and versatile, requiring no additional learnable parameters, and enables seamless integration into neural network architectures with pooling layers. Rigorous mathematical derivations and simulation experiments show that our proposed Gaussian adaptive pooling method surpasses conventional methods (average-pooling and max-pooling) in noise handling. Furthermore, its robustness is comparable to traditional pooling methods in addressing challenges such as rotations, scalings, and translations. Extensive evaluations across multiple computer vision tasks—including image classification (CIFAR-10/100), object detection (MS COCO and RTTS), and semantic segmentation (CamVid)—confirm its effectiveness. Specifically, under varying levels of noise and degraded conditions, Gaussian adaptive pooling achieves significant improvements in standard performance metrics compared to conventional pooling methods. For instance, it delivers notable quantitative gains across different tasks including up to a 12.67% increase in mean intersection over union on the CamVid dataset for semantic segmentation and a 1.1% mAP50 enhancement on the real-world RTTS dataset for object detection.

Infrared small target detection (IRSTD) is inherently challenging due to extremely low signal-to-noise ratios (SNR) in complex backgrounds. Effective detection necessitates capturing local context and suppressing noise to enhance target discrimination. To this end, we propose MSSLNet, which significantly improves detection performance by minimizing irrelevant noise interactions and capturing multi-scale local context. Specifically, a Sparse Large-kernel Convolution Block (SLCB) is introduced, utilizing a sparse-dense dual-branch design to bolster target response while preserving critical information flow. Based on SLCB, a Multi-Scale Sparse Large-kernel Block (MSSLB) is developed to model contextual information for effective target-background separation, further enhanced by a Cross-Gated Module (CGM). Additionally, a Bi-directional Guided Fusion (BGF) module is proposed to optimize cross-layer detail-semantic fusion within skip connections. With only 0.293 M parameters and 2.99 GFLOPs, MSSLNet is highly efficient. Extensive experiments on the NUAA-SIRST and IRSTD-1K datasets demonstrate that our method achieves state-of-the-art performance in both accuracy and efficiency. The code is available at https://github.com/xiaodacheng01/msslnet-main.

A synthetic aperture radar (SAR) raw signal simulator is useful for supporting algorithm innovation, system scheme verification, etc. Trajectory deviation is a realistic factor that should be considered in a SAR raw signal simulator and is very important for applications such as motion composition and image formation for a SAR with nonlinear trajectory. However, existing efficient simulators become deteriorated and even invalid when the magnitude of trajectory deviation increases. Therefore, we designed an efficient SAR raw signal simulator that accounts for large trajectory deviation. Based on spatial spectrum analysis of the SAR raw signal, it is disclosed and verified that the 2D spatial frequency spectrum of the SAR raw signal is an arc of a circle at a fixed transmitted signal frequency. Based on this finding, the proposed method calculates the SAR raw signal by curvilinear integral in the 2D frequency domain. Compared with existing methods, it can precisely simulate the SAR raw signal in the case that the deviation radius is much larger. Moreover, taking advantage of the fast Fourier transform (FFT), the computational complexity of this method is much less than the time-domain ones. Furthermore, this method is applicable for multiple SAR acquisition modes and diverse waveforms and compatible with radar antenna beam width, squint angle, radar signal bandwidth, and trajectory fluctuation. Experimental results show its outstanding performance for simulating the raw signal of SAR with large trajectory deviation.

In the field of modern computer vision, pedestrian detection technology holds significant importance in applications such as intelligent surveillance, autonomous driving, and robot navigation. However, single-modal images struggle to achieve high-precision detection in complex environments. To address this, this study proposes a GLNet-YOLO framework based on cross-modal deep feature fusion, aiming to improve pedestrian detection performance in complex environments by fusing feature information from visible light and infrared images. By extending the YOLOv11 architecture, the framework adopts a dual-branch network structure to process visible light and infrared modal inputs, respectively, and introduces the FM module to realize global feature fusion and enhancement, as well as the DMR module to accomplish local feature separation and interaction. Experimental results show that on the LLVIP dataset, compared to the single-modal YOLOv11 baseline, our fused model improves the mAP@50 by 9.2% over the visible-light-only model and 0.7% over the infrared-only model. This significantly improves the detection accuracy under low-light and complex background conditions and enhances the robustness of the algorithm, and its effectiveness is further verified on the KAIST dataset.

Pedestrian detection holds great significance in various computer vision applications, including intelligent surveillance, autonomous driving, and robot navigation. However, relying solely on single-modal images usually fails to achieve highly accurate detection in complex environments. To overcome this drawback, this study presents GLNet-YOLO, a deep feature fusion framework. The aim is to improve pedestrian detection performance by integrating the features of visible and infrared images. The GLNet-YOLO framework modifies YOLOv11 by adopting a two-branch network architecture. One branch processes visible images, and the other deals with infrared images. To better handle feature fusion, it incorporates the Feature Mixer (FM) module, which is responsible for global feature fusion and enhancement. Additionally, the Dual-Modality Refiner (DMR) module is introduced to separate and interact with local features, thus optimizing the feature fusion procedure. Experimental results in the LLVIP dataset demonstrate that GLNet-YOLO improves mAP@50 by 9.2 % for visible images and 0.7 % for infrared images compared to the original single-modal YOLOv11. The proposed framework significantly improves the detection accuracy under low light and complex background conditions, improving the robustness of the algorithm.

Autonomous underwater vehicles (AUVs) have obtained extensive application in the exploitation of marine resources. Terrain-aided navigation (TAN), as an accurate and reliable autonomous navigation method, is commonly used for AUV navigation. However, its accuracy degrades significantly in self-similar terrain features or measurement uncertainties. To overcome these challenges, we propose a novel terrain-aided navigation framework integrating an Improved Marine Predators Algorithm with Depth-First Search optimization (DFS-IMPA-TAN). This framework maintains positioning precision in partially self-similar terrains through two synergistic mechanisms: (1) IMPA-driven optimization based on the hunger-inspired adaptive exploitation to determine optimal trajectory transformations, cascaded with Kalman filtering for navigation state correction; (2) a Robust Tree (RT) hypothesis manager that maintains potential trajectory candidates in graph-structured memory, employing Depth-First Search for ambiguity resolution in feature matching. Experimental validation through simulations and in-vehicle testing demonstrates the framework’s distinctive advantages: (1) consistent terrain association in partially self-similar topographies; (2) inherent error resilience against ambiguous feature measurements; and (3) long-term navigation stability. In all experimental groups, the root mean squared error of the framework remained around 60 m. Under adverse conditions, its navigation accuracy improved by over 30% compared to other traditional batch processing TAN methods. Comparative analysis confirms superior performance over conventional methods under challenging conditions, establishing DFS-IMPA-TAN as a robust navigation solution for AUVs in complex underwater environments.

As one of the key techniques in the field of machine learning, the technology of object detection has been widely applied in many aspects, particularly in remote sensing small object recognition. However, due to the fact that small objects in remote sensing images have high rotation, dense distribution, complex background, etc., the lack of small object feature extraction capability has become an urgent problem in the field of object detection. Therefore, this paper proposes a remote sensing small object detection algorithm YOLOv8-PNAD (YOLOv8-ParNet And Dynamic snake) based on Dynamic snake convolution and attention mechanism. Dynamic snake convolution and ParNet attention are incorporated in the backbone and neck networks to enhance the extraction capability of features at different scales. Multi-Group experiments are performed on the RSOD dataset, and the experimental outcomes reveal that the improved YOLOv8-PNAD network increases the mAP50 by 2.9% and the mAP50-95 by 1.8% compared to the original YOLOv8 network. By comparing with other mainstream models based on deep learning, it is found that the proposed algorithm improves 6.7% over the RA-BIFPN algorithm, which is also higher than several other existing state-of-the-art object detection algorithms, and demonstrates better recognition capabilities for small objects.

The fundamental trade-off between spatial resolution and imaging distance poses a significant challenge for current imaging techniques,such as those used in modern biomedical diagnosis and remote sensing.Here,we introduce a new conceptual method for imaging dynamic amplitude-phase-mixed objects,termed relay-projection microscopic telescopy(rPMT),which fundamentally challenges conventional light collection techniques by employing non-line-of-sight light collection through square-law relay-projection mechanisms.We successfully resolved tiny features measuring 2.76 μm,22.10 μm,and 35.08 μm for objects positioned at distances of 1019.0 mm,26.4 m,and 96.0 m,respectively,from single-shot spatial power spectrum images captured on the relay screen;these results demonstrate that the resolution capabilities of rPMT significantly surpass the Abbe diffraction limit of the 25 mm-aperture camera lens at the respective distances,achieving resolution improvement factors of 7.9,25.4,and 58.2.The rPMT exhibits long-distance,wide-range,high-resolution imaging capabilities that exceed the diffraction limit of the camera lens and the focusing range limit,even when the objects are obscured by a scattering medium.The rPMT enables telescopic imaging from centimeters to beyond hundreds of meters with micrometer-scale resolution using simple devices,including a laser diode,a portable camera,and a diffusely reflecting whiteboard.Unlike contemporary high-resolution imaging techniques,our method does not require labeling reagents,wavefront modulation,synthetic receive aperture,or ptychography scanning,which significantly reduce the complexity of the imaging system and enhance the application practicality.This method holds particular promise for in-vivo label-free dynamic biomedical microscopic imaging diagnosis and remote surveillance of small objects.

Detecting infrared small targets quickly and accurately in complex backgrounds has always been a challenging task. Data-driven methods have achieved good results because of their powerful feature extraction capabilities. Many algorithms use ResNet or VGG as their backbone, but because of the small size and inconspicuous features, pooling layers in their networks could lead to the loss of targets in deep layers. Even though dense network structure is proposed to alleviate this issue, its excessive dense connections makes it difficult to achieve real-time detection. To meet the requirements of both accurate performance and real-time detection, we propose moderately dense adaptive feature fusion network (MDAFNet). We design a moderately dense adaptive feature fusion (MDAF) module that contains only three feature layers as the backbone of the network. This module connects all the internal features with each other and uses a weighted sum of different layers as the output, promoting feature reuse and maintaining infrared small target features in the deep layers of the network. We also design a coarse-to-fine detection head (CFHead) and introduce auxiliary loss to enable the network to predict target contours with greater precision. Moreover, we propose a new data augmentation method that effectively enhances the generalization performance of network. Experimental results demonstrate that our network achieves excellent performance in detection accuracy and meets the requirements for real-time detection on RTX3080 GPU.

In this paper, we propose a learning-based image fragment pair-searching and -matching approach to solve the challenging restoration problem. Existing works use rule-based methods to match similar contour shapes or textures, which are always difficult to tune hyperparameters for extensive data and computationally time-consuming. Therefore, we propose a neural network that can effectively utilize neighbor textures with contour shape information to fundamentally improve performance. First, we employ a graph-based network to extract the local contour and texture features of fragments. Then, for the pair-searching task, we adopt a linear transformer-based module to integrate these local features and use contrastive loss to encode the global features of each fragment. For the pair-matching task, we design a weighted fusion module to dynamically fuse extracted local contour and texture features, and formulate a similarity matrix for each pair of fragments to calculate the matching score and infer the adjacent segment of contours. To faithfully evaluate our proposed network, we collect a real dataset and generate a simulated image fragment dataset through an algorithm we designed that tears complete images into irregular fragments. The experimental results show that our proposed network achieves excellent pair-searching accuracy, reduces matching errors, and significantly reduces computational time. Source codes and data are available at here .

Change captioning aims to describe the semantic change between a pair of images with natural language while remaining immune to viewpoint change. Based on the encoder-decoder architecture, most existing methods primarily focus on encoding effective change representations for transmission to the decoder. However, they suffer from an insufficient understanding of visual semantics, inadequate single-pass feature comparison, and a confounding bias caused by imbalanced viewpoint change data. These impair change representations and hinder unbiased caption generation. In this paper, we analyze and identify the confounding bias from a causality perspective and propose a Relation-aware Multi-pass Comparison Deconfounded (RMCD) network for change captioning, which elevates the encoding of change representations and mitigates the bias. Specifically, in the encoding stage, to sufficiently understand visual semantics, a position-guided context aggregating module is presented to capture the positional and contextual relations among objects in the image. Then, to achieve comprehensive change representations, we present a multi-pass feature comparison module to recognize semantic differences at various feature levels and progressively integrate them. In the decoding stage, to generate de-biased captions, the causal intervention is employed to remove the confounding bias which introduces spurious correlations between encoded change representations and captions. The newly achieved state-of-the-art performance on four publicly available benchmark datasets and further visual analysis demonstrate the superiority of our method.

Agricultural irrigation is a key component of agricultural production; however, inappropriate water management may lead to water pollution, affecting crop growth and ecological environment. The aim of this study is to use the advanced technique of deep learning convolutional neural network (CNN) in combination with near-infrared (NIR) spectroscopy to achieve intelligent monitoring and analysis of water quality in agricultural irrigation resources. We collected water samples from different agricultural fields and measured the NIR spectral data of these samples using a NIR spectrometer. These data were used to train and validate our designed CNN model. Our CNN model has multiple convolutional and pooling layers to efficiently capture feature information in the spectra. This study provides a new and efficient approach for water resource management and pollution monitoring in agricultural production, which can help to achieve the goals of sustainable agricultural development and ecosystem protection. Future research could further optimise the model performance and consider its application to water quality monitoring systems in real agricultural fields.

Vertical federated learning is a natural and elegant approach to integrate multi-view data vertically partitioned across devices (clients) while preserving their privacies. Apart from the model training, existing methods requires the collaboration of all clients in the model inference. However, the model inference is probably maintained for service in a long time, while the collaboration, especially when the clients belong to different organizations, is unpredictable in real-world scenarios, such as concellation of contract, network unavailablity, etc., resulting in the failure of them. To address this issue, we, at the first attempt, propose a flexible Active-Passive Federated learning (APFed) framework. Specifically, the active client is the initiator of a learning task and responsible to build the complete model, while the passive clients only serve as assistants. Once the model built, the active client can make inference independently. In addition, we instance the APFed framework into two classification methods with employing the reconstruction loss and the contrastive loss on passive clients, respectively. Meanwhile, the two methods are tested in a set of experiments and achieves desired results, validating their effectiveness.

Aiming at the problem of image quality degradation caused by the scattering of particles in the atmosphere under foggy conditions, a polarization dehazing algorithm combining the target polarization degree and atmospheric transmission model is proposed in this paper. Firstly, a bilateral filtering method combined with image gradient information is proposed to solve four target light intensity images from different angles, and the filtered images are used to solve the polarization degree of the target. This method can preserve the edge texture information of the target in the filtering process and effectively improve the quality of the reconstructed image. Secondly, when estimating the atmospheric light intensity at infinite distance, an alpha filtering method combined with bright channels is proposed to avoid the interference of over-bright noise points in the image. The method can effectively suppress the atmospheric light intensity at infinite distance. By analyzing the experimental data, the average gradient and gray variance of the fog-free reconstructed image are significantly improved compared with the original image. Experimental results show that the proposed algorithm has strong defogging ability, and can effectively improve the image quality of the optical imaging system in foggy scenes, and realize the restoration of

Inshore ship detection is an important task in several fields, for example, maritime transportation, maritime supervision, and port management. However, due to the diversified categories and locations of different ships and interference of complex surroundings, capturing discriminative characteristics of multi-scale inshore ships for accurate detection is still challenging. Here, an anchor-guided attention refinement network (AARN) is proposed to alleviate the problems by prominently designing an attention feature filter module (AFFM) and an anchor-guided alignment detection module (AADM). In AFFM, of which the attention supervision generated from high-level semantic features, is used to highlight informative target cues and suppress background interference when establishing a four-layer feature pyramid. In AADM, the anchor-aligned features are adopted to eventually identify potential inshore ships, which both alleviates misalignment between refined anchors and pyramidal features and improves the performance further. Extensive experiments conducted on the public Seaships7000 dataset verify the contributions of the proposed modules and the effectiveness of our method for detecting multi-scale inshore ships in comparison to both domain-specific and general CNN-based methods.

Strictly uniform time interval between adjacent channels is a crucial requirement for the multi-wavelength interleaved (MWI) pulse sources, which difficult alignment can be easily solved by what we believe to be our novel method based on frequency-resolved optical gating (FROG). By utilizing highly precise measurements from FROG, which provide fully two-dimensional information of the pulses in time and frequency domain, we can intuitively identify the time mismatches between different channels in the MWI pulse sources. This enables us to directly align the timing of each channel with sub-picosecond resolution at the first time. MWI pulse sources with total repetition rate of 20.8 GHz (four wavelengths) and 41.6 GHz (eight wavelengths) are precisely aligned by the proposed method, this achievement will pave the way for advancements in photonic analog-digital converters (PADC), high-speed optical communications and so on.