10 December 2025 in IEEE Sensors Journal
Accurate surface defect detection is crucial for maintaining the structural integrity and performance of wind turbine blades (WTB), thereby optimizing their operational reliability. Due to limitations in drone flight and the small size of certain defects, existing algorithms struggle to detect minor target defects within the high-resolution images associated with WTB. This paper proposes KGP-YOLO, an innovative detection method for WTB surface defects in high-resolution images, which for the first time introduces keypoint localization and Bi-Level Routing Cross-Attention (BRCA). The network comprises a filter and a detector. We propose the Keypoint-aware Geometric-guided Preprocessing Network (KGP-Net) as the filter, select the advanced object detection model YOLOv11 as the detector, and integrate the BRCA module into YOLOv11.Initially, KGP-Net is employed to identify blade regions within high-resolution images and extract cropped images containing only the blade regions. This ensures the model focuses exclusively on the target blade areas, minimizing interference from irrelevant background regions. Subsequently, integrating the BRCA module into YOLOv11 introduces a dynamic sparse self-attention mechanism, significantly enhancing the network’s ability to capture deep-level features essential for effectively identifying complex WTB surface defects. The proposed method achieves a mean Average Precision (mAP) of 87.3% on the self-built SY-PLUS dataset, 92.4% on the public DTU dataset, and 88.5% on the public Blade30 dataset, while maintaining acceptable inference latency. Experimental results confirm that this method excels in detecting small object defects, even in high-resolution scenarios (e.g., 1080P), meeting the demands of practical applications.
23 April 2026 in IEEE Sensors Journal
In traffic and water-surface environments, the fusion of RGB imagery and LiDAR point-cloud data can effectively improve detection performance. However, a key challenge arises when detecting target regions that occupy only a small portion of the image and point cloud, yet these regions are often complex. RGB provides dense appearance features but lacks reliability under reflective surfaces, low-texture environments, and varying lighting conditions; while LiDAR-based depth information, though offering more stable geometric features, suffers from data sparsity. To address these issues, this paper proposes the DAMNet (Deformable Attention Multi-modal Network) model, which utilizes a dual-stream backbone for extracting and fusing RGB and point cloud-derived depth maps. Through a progressive fusion strategy, DAMNet integrates features from both modalities across multiple spatial scales, improving object detection accuracy and robustness in marine environments. The three key modules are: LDMSSM (Learnable Direction Multi-Scale State Space Model) is designed as a direction-aware multi-scale state space module to enhance efficiency and alleviate computational overhead while maintaining accuracy; MASCAM (Multi-scale Asymmetric Spatial Channel Attention Module), which enhances cross-modal feature expression through multi-scale convolutions and asymmetric spatial pyramid structures; and DAMFM (Deformable Adaptive Masked Fusion Module), which ensures precise alignment, fusion, and refinement of features through deformable attention and hybrid dynamic masking strategies. Experimental results on the KITTI dataset demonstrate that DAMNet achieves state-of-the-art performance with mAP50 of 93.0% and mAP50-95 of 70.9%, using 30.65M parameters and 36.98 GFLOPs. The superior performance on water surface object detection tasks with mAP50-95 of 93.5% validates the effectiveness of the proposed multi-modal fusion approach in complex marine environments.
25 October 2024 in IEEE Transactions on Industrial Informatics
Scene perception for autonomous vehicles and vessels is crucial for autonomous navigation. Current mainstream transformer methods typically split the feature map into windows, such as local, dilated, and horizontal/vertical bar windows. However, their token interaction is confined to fixed windows, posing challenges for image-based semantic segmentation. This article proposes a novel model, LBFormer, which enables flexible token interaction across different windows. Specifically, a window-level affinity graph is constructed from coarse-grained features using self-attention clustering and evolves during training, retaining top-k windows with high semantic relevance for each window. Self-attention purification is then employed to compress and filter fine-grained features with low semantic relevance within the top-k windows, ensuring effective token interaction for each feature point. To enhance context modeling within windows and build a more effective window-level affinity graph, a dual branch method extracts multidimensional features from each window, which are then interacted with and fused via the feature aggregation module. Extensive experiments at an image resolution of 224×224 were conducted on our private YZ-DATA water surface scene dataset and the public CamVid urban scene dataset. The results show that LBFormer achieves an MIoU of 89.80% on YZ-DATA and 61.31% on CamVid, surpassing mainstream transformer methods.