In aerial images, slender and multi-scale insulator targets are often set against complex backgrounds, which imposes higher demands on the model's ability to retain details and capture long-range dependencies. To address this challenge, we propose a lightweight Dual-branch Residual Interaction Network (RI-Mamba) for insulator segmentation. Unlike typical single-branch architectures, RI-Mamba adopts a parallel dual-branch to prioritize the extraction of detail-semantic features. Specifically, the detail branch enhances the representation of local fine structures through feature expansion, while the semantic branch strengthens the correlation between positional and contextual semantics from a coordinate perspective. During the feature fusion stage, the proposed Residual Mamba Block (R-Mamba) discards traditional convolutional methods for processing local features and instead directly reuses detail-semantic information from the original features via a residual structure, achieving efficient feature reuse while avoiding parameter growth. Furthermore, a Gated Global-local Interaction Module (G2LIM) is introduced, which uses a learnable gated self-attention mechanism to dynamically adjust the fusion weights of global context and local details, thereby reducing redundant computation and enhancing feature adaptability. Experimental results on two public datasets demonstrate that RI-Mamba, with only 4.15M Params and 5.31G FLops, achieves mIoU of 86.28% and 91.59%, respectively, outperforming 17 lightweight insulator segmentation models. These results high ght that RI-Mamba achieves a favorable balance between computational efficiency and segmentation accuracy, making it well-suited for deployment in real-world power inspection scenarios. The code is available at: https://github.com/yaoshuang-yaobo/RI-Mamba.
Difference Visual Question Answering (Diff-VQA) in medical imaging automatically compares patient images across time points to support assessment of lesion progression and treatment efficacy. However, pixel-level matching is unreliable due to non-rigid deformations, view shifts, and acquisition noise, while existing models often rely on synthetic labels and lack effective integration of local and global information. To address these challenges, we propose a multimodal large-model framework that adopts a progressive “local semantic modeling–global difference reasoning” strategy. Key anatomical regions in chest X-rays are localized via object detection and aligned with VinDr-CXR annotations to construct region–disease mappings, transforming misalignment into semantic difference analysis. A dynamic sampling strategy further generates clinically meaningful image pairs with fine-grained difference labels. Finally, a multimodal large model fuses local features with global context to support single-image QA, dual-image disease description, and global difference reasoning. Experiments on the MIMIC-Diff-VQA dataset demonstrate state-of-the-art accuracy in single-image QA and substantial improvements in Diff-VQA tasks over mainstream medical large models. In the single-image QA tasks, our model improves accuracy from 52.5% to 64.1% (22.2% relative improvement), and in the Diff-VQA tasks, the CIDEr score increases from 1.027 to 1.379 (34.3% relative improvement). These results highlight the framework’s potential to enhance diagnostic accuracy and strengthen clinical decision support in radiology practice.
Printed Circuit Boards (PCB) are a fundamental component of electronic devices, serving to connect and support electronic components. They are widely used in various fields, including Consumer Electronics, Industrial Control, and Aerospace. However, modern PCBs are trending toward miniaturization, lightweight design, and high density. Consequently, tiny surface defects appear more frequently. Defects such as missing holes, mouse bites, and open circuits exhibit low contrast, making them challenging to detect. Traditional inspection methods suffer from several limitations. Manual visual inspection and electrical testing are inefficient. They suffer from low accuracy, poor efficiency, and high costs. The automated optical inspection (AOI) still needs people’s assistance. In addition, the early deep learning models like the initial versions of YOLO do not accurately detect small defects. During upsampling, these models lose fine-grained details and fail to adapt to unbalanced data. Consequently, they often fail to detect low-contrast defects. To address these issues, this paper proposes YOLO-iDAT. It is an improved version of YOLOv11. First, this paper changes the C2PSA module into C2PSA_iRMB. This new module combines the iRMB structure, SE attention and residual connections. It makes the model mix the global and local information better. And it significantly enhances the ability to extract the tiny defect feature. Second, this paper uses DySample for up sampling. It is a dynamic method. It prevents the blurring of small defects that often occurs with conventional upsampling methods. It keeps the feature clear. And it does not make the model slow. Third, this paper uses the Adaptive Threshold Focal Loss (ATFL) function. It uses an adaptive threshold and adjusts parameters to adapt to unbalanced data. It significantly reduces the number of missed defects. This paper uses the proposed model to test the public PKU-Market-PCB dataset. The results show that YOLO-iDAT gets a precision of 0.941, the mAP50 of 0.93 and the recall of 0.895. These represent improvements of 0.9%, 1%, and 1.8%, respectively, over the original YOLOv11. And the speed of our model is very fast. It gets FPS of 88.76. This is a sufficient speed for real-time detection in factory settings. These results demonstrate that YOLO-iDAT fits the requirement of industrial PCB defect detection. It provides a more accurate and more efficient solution for PCB defect detection.
As a core interaction in the human-vehicle-road system, automated and refined detection of driving behavior has emerged as a crucial research direction in intelligent transportation systems and advanced driver assistance systems. Traditional post-event monitoring models that rely on manual or sensor-based methods are no longer able to meet the requirements of real-time and accurate risk identification. Therefore, this study proposes the You Only Look Once - Lightweight - BiFPN - ECA (YOLO-LBE) detection method. By integrating ghost convolution and GhostC2f modules to diminish computational complexity, the study employs a weighted bidirectional feature pyramid network, and further embed an ECA module to significantly enhance the precision and stability of driver behavior detection. Experimental findings demonstrate that the improved YOLOv8 model improves mAP@0.5 by 5.3%, FPS by 32.4%, Params by 34.4%, and FLOPs by 33.3% compared to YOLOv5s. This research method outperforms existing mainstream models in terms of accuracy, efficiency, and interference tolerance, providing reliable technical support for real-time driving behavior monitoring.
The continuous expansion of the scale of the IoT has led to frequent occurrence of network attack events and posed severe challenges to device operation and data security. To accurately track the behavior of IoT devices in the network and detect any malicious activities that may occur, this study proposes a spatiotemporal feature-based IoT security detection method that combines TPA mechanism, BiLSTM, DenseNet 1D, and CNN. The experiment was validated using two publicly available datasets, UNSW-NB15 and Bot-IoT. The research method could effectively identify various types of attacks (such as user to administrator privilege escalation attacks, remote to local attacks, etc.), alleviate the impact of imbalanced dataset categories, and maintain low online detection latency. In the UNSW-NB15, the precision, recall, and F1 value of the research method reached 97.6%, 96.8%, and 97.2%, while in the Bot-IoT, they reached 95.7%, 94.1%, and 94.9%. When the sample size increased to 1×106, the running time and energy consumption of the research method were only 13.2 s and 0.9 kWh. The research method effectively enhances the ability to identify multiple types of attack traffic, providing a technical solution that balances accuracy and robustness for malicious behavior detection in IoT environments.
The deployment of Deep Learning models on heterogeneous hardware necessitates optimized tensor programs to utilize computational resources. While auto-tuning frameworks like Ansor employ evolutionary search strategies, they suffer from limitations, including slow convergence, memory consumption, and a tendency to stagnate in local optima within high-dimensional, non-convex search spaces. To address these challenges, this paper proposes CoDSA (Coordinate Descent with Dynamic Simulated Annealing), a hybrid auto-tuning framework that integrates an enhanced Droplet Search algorithm into the Ansor ecosystem. We introduce three improvements to the droplet coordinate descent strategy: 1) a Dynamic Simulated Annealing mechanism to escape local optima; 2) an Adaptive Step-Size Control modulated by an exploration factor to balance global traversal and local refinement; and 3) a Diversity Control mechanism utilizing cosine similarity to enforce orthogonal exploration. Experiments on NVIDIA GPUs using Vision Transformer, DeiT, and Swin Transformer models demonstrate that CoDSA outperforms the Ansor framework. Specifically, for the ViTBase model, our method reduces search time by 37.2% and memory usage by 53.1%, while improving inference latency by up to 26.5% on complex architectures like Swin-Small. This work bridges the gap between the global exploration capability of evolutionary methods and the rapid exploitation efficiency of coordinate descent.
Movable antenna (MA) systems constitute a novel and promising approach in the field of wireless communication design, utilizing positional flexibility within a defined spatial domain to enhance signal quality and network performance. Through the dynamic reconfiguration of antenna locations, MA frameworks achieve substantial improvements in key performance indicators, including signal-to-noise ratio (SNR) and coverage robustness, particularly in challenging propagation scenarios. We introduce an alternating optimization framework to jointly address beamforming, user selection, and antenna positioning. The method iteratively updates: (i) ZFBF beamforming weights via water-filling power allocation, (ii) near-orthogonal user subsets via SUS, and (iii) antenna positions via SQP under mechanical constraints. Overall, the simulation outcomes indicate that this approach demonstrates notable superiority in terms of beamforming gain and interference mitigation, significantly outperforming conventional systems reliant on fixed-position antennas (FPAs).
Aiming at the non-stationary, nonlinear and noise-sensitive characteristics of rolling bearing vibration signals, as well as the low recognition accuracy of traditional deep learning in small-sample scenarios, this paper proposes a rolling bearing fault diagnosis method combining Continuous Wavelet Transform with Ridge Tracking (CWT-RT) and Multi-Scale Wavelet Scattering Network. First, Variational Mode Decomposition integrated with Cramer Von Misses statistic (VMD-CVM) is adopted to denoise the original signal and improve the signal-to-noise ratio. Then, CWT-RT is used to transform the denoised signal into time-frequency spectrograms for intuitive time-frequency feature representation. Multi-Scale Wavelet Scattering Network is further applied to extract multi-level structural features, which are fed into Multi-Layer Perceptron (MLP) to realize bearing fault identification. To eliminate data leakage, all original raw vibration files are split into training and test sets at a 7:3 ratio before sliding window sampling. Validation experiments on bearing datasets from South Ural State University, CWRU, and XJTU-SY show that the diagnostic accuracies on two small-sample conditions reach 98.78% and 98.09%, respectively. the 95% confidence intervals for the two accuracy values are [98.21%, 99.15%] and [97.43%, 98.67%], respectively. Across 10 repeated experiments, p-values < 0.001 confirm the statistical significance of the results. The model maintains high accuracy under different loads and noise levels (0/5/10/15 dB). Comparative and ablation experiments verify that the method has high diagnostic accuracy, strong noise robustness and superior small-sample learning ability, with each module effective, and the full-pipeline latency meets the real-time requirements of IoT edge deployment, providing support for industrial engineering applications.
Speech signals-based identity recognition is an active research area, which has a wide range of important applications such as human-computer interaction, forensic investigation, sound surveillance, transaction authentication, and health monitoring. As a special speech event, cough plays a crucial supplementary role in identity recognition scenarios, especially when traditional biometric features are unavailable. However, compared to traditional speaker recognition, very few research efforts have been made by researchers to explore identity recognition using cough acoustic signal analysis. In this work, we propose an EGOOSE-LightGBM model for cougher recognition, in which the hyperparameters of LightGBM are optimized by EGOOSE algorithm. Chaotic mapping, Gaussian mutation, and crisscross optimization are employed in the proposed metaheuristic-type EGOOSE algorithm, which demonstrates superior performance in terms of convergence speed and accuracy compared to GOOSE algorithm. A static-dynamic feature fusion (SDFF) technique is used to fuse cough sound characteristics including Mel frequency cepstral coefficient (MFCC), fundamental frequency, formant, triangle MFCC, and jitter of formant to improve recognition effectiveness and noise robustness. Using the self-re-corded cough sounds, we establish a dataset, the size of which is enlarged by data augmentation techniques. Performance improvement of 3.167% is obtained due to the exploitation of cough data augmentation. Exper-imental results show that EGOOSE-LightGBM outperforms other existing machine learning models such as SVM, XGBoost, RF-Adaboost, LSTM, and LightGBM, achieving remarkable recognition accuracy of 99.500%.
Investigating the intricate semantics of diverse wetland landscapes is vital for the development of Intelligent computing systems within remote sensing applications. This study introduces a cutting-edge approach that utilizes memristor-based architectures to integrate multi-channel perceptual visual features for classifying wetland remote sensing images, characterized by complex spatial and ecological structures. Our method leverages a deep hierarchical model designed to emulate human gaze dynamics through a memristor-enabled processing unit, employing the BING objectness metric to accurately detect key ecological features and details across multiple scales within wetland scenes. To enhance the human-like visual attention mechanism, we propose a Memristor-Enhanced Robust Deep Active Learning (MRDAL) strategy, which systematically generates gaze shifting paths (GSPs) and extracts their deep representations using memristor-based networks. A distinctive aspect of MRDAL is its resilience against label noise, achieved through a sparse penalty mechanism embedded within the memristor architecture, effectively filtering out irrelevant GSP features. We subsequently apply a manifold-regularized feature selector (MRFS) integrated with memristive components to extract high-quality deep GSP features, which are then utilized to train a linear Support Vector Machine (SVM) for the classification of wetland scenes. Empirical evaluations reveal the method's superi-or performance over conventional models, demonstrating its exceptional capability in discerning complex patterns within a comprehensive dataset of large-scale wetland remote sensing images. This advancement highlights the potential of memristor-based intelligent computing technologies for ecological monitoring and environmental analysis.
Accurate time-series forecasting is challenging when multiple seasonalities interact with non-linear effects. We present MP-Transformer, a hybrid" decompose-then-refine" framework that couples an interpretable Multi-Period ARIMA baseline with a dynamically gated attention residual learner. The ARIMA component extracts dominant linear trends and multi-scale seasonality via seasonal phase templates with synchronous differencing, yielding an approximately stationary residual series and an interpretable baseline. A Transformer then models the remaining non-linear dynamics using a gated fusion of global attention (for long-range periodic dependencies) and content-driven Top-k local attention (for abrupt short-term variations). Period contributions are learned through non-negative, normalized gating weights. Across multiple real-world datasets, MP-Transformer consistently improves multi-horizon accuracy over statistical, deep, and hybrid baselines. The results demonstrate that combining explicit linear decomposition with implicit residual learning yields robust, data-efficient forecasting and enhanced interpretability.
Classifying fetal cardiotocography data is essential in the efficient prenatal risk assessment due to its potential for identifying errors or abnormalities during pregnancy. Traditional fetal heart rate (FHR) analysis frameworks, which unfortunately still rely on manual interpretation of results, subsequently lead to the inefficient use of human resources and sometimes require more time for abnormality detection. With the implementation of Machine Learning (ML) algorithms, automatic analysis and early detection of abnormalities are now possible. The model's performance is directly influenced by the retrieval of features and the optimal management of class imbalance in the dataset. In this regard, we introduce a feature-based innovative strategy for multi-class classification in fetal cardiograph datasets based on feature importance analysis. The proposed model utilizes Random Forest (RF) for feature extraction, which employs two distinct target importance analyses: 1. class imbalance, and 2. class weights. In Phase 1, an artificial neural network and an improved TabNet model were utilised for classifying three classes: Normal, Suspect, and Pathology (NSP), with SMOTE balancing. In Phase 2, we identify the features of classes that contribute to NSP classification, and we consider nine additional features based on class weight for various cardiotography features, such as baseline, ASTV, ALTV, etc. In Phase 2, NSP classification is performed by including class 1-9 features (A, B, C, D, E, AD, DE...) and assigning class weights. Using our proposed ensemble deep learning model, the accuracy of prediction is improved. The RF model retrieves primary features from the fetal cardiograph, and complex relationships among these features enhance the representation of information. The next step is the classification stage, which applies an attention-based deep learning model, TabNet. Due to the nature of the TABNET model in handling tabular data, it can selectively focus on relevant features while ensuring explainability. The proposed model is evaluated using different performance metrics for two novel feature importance analyses. The RF+TabNet+LSTM achieves a maximum accuracy of 97% with SMOTE in NSP target classification (phase 1), while including Class weight in class1-9 features, the model achieves classification accuracy of 92% (phase II) and proves the importance of features contributing to prediction and classification. All code and the curated dataset for Multiclass Fetal Abnormality Detection are available at https://github.com/rrramyaresea/Multiclass-Fetal-Abnormality-Detection, enabling the reproducibility of our findings.
This paper proposes a distributed unsupervised dimensionality reduction network framework for visual data, referred to as DIST-C-SEGAN. It aims to achieve low-bandwidth, high-fidelity, scalable, and privacy-friendly unsupervised dimensionality reduction and compression oflarge-scale image data, laying a solid foundation for downstream tasks in resource-constrained distributed or edge computing environments. This method captures local spatial information through a lightweight convolutional network and explores deep feature information using a graph neural propagation structure that combines compression and activation techniques with multi-head self-attention mechanisms. At the same time, it introduces the Proximal Alignment Algorithm to generate low-dimensional embedding features, achieving end-to-end "near-lossless" preservation of local neighborhood structure and global manifold continuity. Experimental results showthat DIST-C-SEGAN achieves an original variance retention rate of over 98% in all data tasks, demonstrating high reliability and strong continuity. Its structural evaluation metrics are significantly superior to those of classic dimensionality reduction techniques, such as Principal Component Analysis (PCA), Kernel Principal Component Analysis (KPCA), Autoencoder (AE), Variational Autoencoder (VAE), UMAP, and t-SNE. Sensitivity analysis determines the optimal range of key hyperparameters and confirms their robustness to random parameter initialization. Ablation experiments verify the necessity of multi-head self-attention mechanisms and geometric alignment, while downstream tasks further confirm their effectiveness. Moreover, DIST-C-SEGAN theoretically reduces communication costs and provides new solutions for high-fidelity and communication-efficient compression and recovery in distributed environments.
Retrieval-augmented generation (RAG) enhances large language models (LLMs) by integrating external knowledge sources. Yet, existing chunk-or graph-based frameworks remain confined to surface-level semantic correlations, lacking the capacity to model mechanistic causality, cross-document dependencies, and counterfactual reasoning. To address these limitations, this study introduces CogniRAG, a causal reasoning-centric RAG framework that systematically transforms retrieval into a structured process of causal inference. CogniRAG encodes four specialized causal hyperedge types-Linear Chain, Feedback Loop, Intervention Point, and System Stability-within a unified CogniGraphDB, enabling multi-entity causal tracing and intervention analysis. A dual retrieval strategy combining entity diffusion and relation expansion constructs causally enriched prompts that enhance inferential depth beyond conventional similarity-based retrieval. Experiments conducted across four knowledge-intensive domains (Medicine, Political Science, Computer Science, Finance) using an identical LLaMA-3.1-70B backbone and 480 systematically generated questions demonstrate consistent and statistically robust improvements. CogniRAG achieves an approximately five-percentage-point Overall gain over HyperRAG across all domains (for example, Medicine 78.67 vs 73.63), with the largest increases observed in Empowerment and Logical Coherence, corroborated by objective reliability metrics (+1.8 pp Faithfulness, +1.2 pp Factual Consistency, -0.7 pp Hallucination). Cross-backbone evaluation with Qwen-Plus and Mistral Large 2 confirms backbone-agnostic consistency (+3-5 pp), while efficiency analysis reveals moderate computational overheads (indexing +30-40%; latency +50%) proportional to added causal reasoning depth. Overall, CogniRAG advances RAG from semantic retrieval toward causal, counterfactual, and interpretable knowledge reasoning, with future work focusing on real-time causal-graph updating, multimodal evidence integration, reproducible benchmarking, and meta-reasoning for confidence calibration to strengthen scalability and external validity.
Recent advancements in the field of deep learning have fundamentally transformed the landscape of medical image segmentation, particularly in the critical area of brain tumor diagnostics. This study introduces DAU2-Net, a novel dual attention-enhanced U2-Net architecture that integrates spatial and channel attention mechanisms with MobileViT blocks to prioritize tumor-specific features and model global contextual relationships. The primary objective of this integration is to significantly enhance the segmentation accuracy, thereby providing more reliable and precise diagnostic tools for medical professionals. Evaluated on BraTS 2019, DAU2-Net achieves a state-of-the-art Dice coefficient of 0.92. This remarkable result not only surpasses the performance of the traditional U-Net model, which recorded a Dice coefficient of 0.88 (+4%), but also outperforms the Fully Convolutional Network (FCN) with a Dice coefficient of 0.85 (+7%). And it also achieves 94.2% sensitivity and 97.5% specificity.These outstanding results highlight the significant effectiveness of the dual attention mechanisms utilized in DAU(2)-Net, which excel at capturing multi-scale contextual information. This capability is crucial for achieving precise tumor delineation, ultimately contributing to more accurate diagnoses and improved patient outcomes in the field of brain tumor diagnostics.
Amid rapid advancement in digital governance and exponential growth of public appeal data, traditional manual text classification increasingly fails to meet governmental requirements for efficient, accurate, and timely service delivery. This study focuses on automatic classification of public affairs appeal texts through systematic investigation of deep learning models. To resolve data duplication, class imbalance, and textual noise, we implemented optimization strategies including deduplication and class resampling. Addressing the generalization and stability limitations of individual models - specifically Enhanced TextCNN, BiLSTM with attention, BERT, and ERNIE3.0 - we propose a deep neural network that in-tegrates BERT's contextual semantic embeddings, CNN's local feature extraction, and BiLSTM's temporal dependency modeling. This architecture employs feature concatenation and dropout mechanisms to effectively synthesize global semantics, local phrases, and sequential features. Experimental results demonstrate substantial superiority over conventional models, achieving 99.06% accuracy and 99.03% F1-score on the validation set, confirming exceptional classification performance and robustness. This approach offers an efficient solution for intelligent public appeal processing while advancing digital governance capabilities and governmental modernization. Furthermore, it establishes a valuable ref-erence framework for complex Chinese text classification tasks.
To address issues such as blurred foreground-background separation, low detail fidelity, and hazy background blur in traditional methods for optimizing depth-of-field effects in photographic images, this study proposes an improved unsupervised cyclic generative adversarial networks model. First, the study builds upon unsupervised cyclic generative adversarial networks, enhancing the generator through deep separable convolutions and coordinate attention modules. It replaces the dual-discriminator architecture with a class activation mapping module within the discriminator, while integrating the Retinex algorithm into post-processing to address background haze effects. Algorithm performance test results indicated that in the depth-of-field information processing scene, structural similarity reached 30.4. In the edge information processing scene, structural similarity reached 28.1. All scenes maintained optimal performance. In depth-of-field optimization tests for bow-and-net images, the model's output images achieved an average information entropy of 7.22 (compared to 7.12 for original images). It effectively mitigated typical depth-of-field defects such as background overexposure and motion blur. The model achieves a balance between depth-of-field effects, efficiency, and visual quality, generating high-resolution shallow depth-of-field images. It provides a solution for unsupervised depth-of-field optimization, applicable to post-production photography and industrial image analysis. Future performance enhancements can be achieved through multi-scene dataset expansion and end-to-end optimization.
The pre-trained large language models (LLMs) achieve impressive advancements not only in text-based tasks but also show significant potential in basic visual-language comprehension. However, it remains uncertain whether LLMs pre-trained on text can comprehend the grammar of sign language based on gestural actions after fine-tuning with limited data. Despite the near-human performance of current LLMs in understanding and generating spoken text, their ability to transition from text language to visual language for multimodal tasks is still confusing. In this paper, we propose the SL-LLaMA model, which leverages the robust capabilities of LLMs for sign language translation tasks and investigate the multimodal abilities of LLMs in transferring from textual language to visual language. We use the LLaMA 2 family of models to perceive and understand sign language grammar and to generate corresponding spoken text. To incorporate video information into the LLM, we propose a sign language translation framework that integrates a vision encoder, an MM-Adaptor, and an LLM to understand sign language and generate the spoken language. Additionally, we employed the language alignment-Supervised fine-tuning training strategy to infuse sign language knowledge into the model. Our study evaluates the performance of gloss-free sign language translation on two benchmarks: RWTH-PHOENIX-Weather-2014-T and CSL-Daily. Compared to current state-of-the-art methods, the proposed model achieves competitive results, demonstrating the strong potential of text-pretrained LLMs in understanding visual gram matical knowledge. Ablation experiments explore the impact of each component on sign language translation, as well as the framework's generalization and scalability, providing a foundational basis for future applications of LLMs in more complex multimodal tasks.
Underground coal-mine scenes are often affected by low lighting, dust, and motion blur., and monitoring images are often affected by low lighting, dust, and motion blur, which severely hinders the recognition performance of intelligent detection systems. To address the decline in object detection accuracy caused by poor image quality, this paper proposes an object detection method combining a detection-guided du-al-branch image enhancement module with YOLOv11. The method includes two key innovations: first, a detection-guided image enhancement module is developed, which generates attention maps from low-con-fidence regions in YOLOv11's detection output to guide the enhancement network in focusing on critical image areas, thereby improving detection accuracy. Second, a dual-branch structure combining brightness enhancement and detail reconstruction is designed, integrating the strengths of Zero-DCE and SRGAN to achieve global brightness correction and edge clarity enhancement. Experiments conducted on the public-ly available underground coal mine image dataset DsLMF+ demonstrate that the proposed method outper-forms traditional enhancement-plus-detection pipelines in multiple evaluation metrics. The enhancement module significantly improves image quality indicators, and achieves notable improvements in YOLOv11 detection accuracy, particularly in mean Average Precision (mAP). This study confirms the effectiveness of combining detection feedback and multi-scale enhancement in complex industrial environments, offering a promising solution for underground object detection tasks.
To address power imbalance, voltage fluctuations, and low new energy consumption rate caused by high penetration of distributed new energy in multi-power distribution networks, this study proposes an energy storage power control method based on an improved ant colony algorithm (ACO). First, A multi-objective optimization model was constructed, considering the reduction of network loss, improvement of new energy consumption, and stability of energy storage SOC. The entropy weight method was used to objectively determine the target weights. Second, to solve the poor accuracy of traditional ACO in continuous control, a "continuous domain discretization" strategy was introduced, and pheromone update rules were optimized to enhance algorithm convergence and precision. Finally, a case study of a 10kV industrial park distribution network is introduced. The research shows that compared with traditional ant colony algorithm and particle swarm algorithm, the improved algorithm reduces voltage fluctuations by 42%, lowers the daily network loss rate to 2.13% (44.2% lower than no energy storage), and increases the photovoltaic consumption rate on clear days to 96.2%. The results of the paper validate the effectiveness of the proposed approach in optimizing the operation of the distribution network and promoting the integration of new energy.
